AltaML blog

The Chat Box is the New Command Line

Vinoth Babu — Fri, 21 Aug 2026 16:00:00 GMT

Every few decades, computing undergoes a fundamental shift. Business to personal computing, Desktops vs. handhelds, etc. Similarly, how humans interact with a computer has historically changed as well—think Command line vs. Graphical User Interface (GUI.) During these periods of transition, emerging technologies often arrive wearing the clothes of their predecessors, retaining familiar interfaces even as the underlying technology changes. We are experiencing this transition now. Despite having access to a newer, more powerful form of computer (an LLM-driven computer), we continue to interact with it through a text-based interface, issuing one request at a time. We at AltaML firmly believe this is poised to change, and soon.

Some History

1983: The Whole System

The command line was both complete and unstructured. The machine could perform virtually any operation that could be expressed as a command, but it provided little guidance about what to do. The user manuals were massive and the user had to supply all information to get the machine to do… anything (within its capabilities).

1984: When Computing Last Changed Shape

The command line was a complete interface. Anything the machine could do, you could ask for, provided you knew how to express it. That completeness came at a cost: you had to hold the system's logic in your head (or a very large manual) and begin each task from scratch.

Then the screen changed, and within a decade, the argument had been settled.


DEC VT100 terminal by Jason Scott, CC BY 2.0 The DEC VT100, introduced in 1978, placed green text on a black screen. You typed a command, it returned a response, and the session effectively ended when the interaction did.	Macintosh 128k transparency by Grm wnr, CC BY-SA 3.0. The original Macintosh (1984) had a slower processor and less memory, yet it succeeded because it made your work visible. You could see what you were doing, interact with it directly, and point to the thing you wanted to change.

Three changes made the difference:

Commands you had to memorize became objects you could see, point to, and manipulate.
One task at a time, in sequence became multiple tasks held together by the system.
Instructions for doing the work became the work itself, directly visible and manipulable.

The desktop made capability visible, and that visibility is what made computers accessible to a much broader audience, ushering in an era of innovation leading to other amazing innovations like the internet, iPhone, etc. But this was made possible only because of how this technology, which was initially reserved for very few, became accessible for the masses to use without fear or concern.

2023: A New Form of Computing

Forty years later, and we are right back where we started.

With the widespread adoption of LLMs, and more recently the agentic systems built on top of them, the user is once again required to express detailed intent and provide complete context, even with interfaces as capable as MCP. Up until now, they have failed at greatness because humans still have to work with an agent, instead of having their agent, or computer, do the work for them.

2026: The Interface at a Turning Point

Two forms of AI dominate today, and both are terminals in disguise.

The Sparkle Button

AI increasingly arrives as a feature bolted onto software you already use: a button in the corner of your mail client, ticket tracker, or document editor. It can make individual tasks faster, but it leaves the underlying structure of your work largely unchanged. You still move from Slack to your inbox, then to your calendar, then to the ticket tracker, and are responsible for carrying some critical context between each system yourself. The machine may be faster, but the workflow remains fundamentally the same.

The Chat Box

Or AI arrives as a single text field with everything behind it, including connectors. It offers enormous capability, yet almost no structure. You have to know what to ask, phrase it effectively, interpret the prose it returns, and maintain the state of the work yourself between turns. Read that description again and it starts to sound remarkably like 1983, just with better technology.

From Prompting to Curating

The shift that matters is a change in who holds the thread of the work. It becomes visible in the smallest interactions:

From	To
A prompt you retype (or run as a scheduled job) every morning	An objective that stands until it's met
AI that assists you inside your apps	AI that assembles the surface around the outcome, then dissolves it when the work is done
A transcript you scroll back through	Surfaces you can point at, where the state lives in the room instead of the scrollback
One assistant you supervise turn-by-turn	A fleet of small, scoped agents working in parallel, reporting into one place
AI that answers when asked or executes tasks you have given	AI that's already done the first pass, and shows you exactly where to look or what to do next

Underneath all five is the same economic reality: your attention is now the scarce resource. Today's software spends it lavishly, asking you to babysit several agents and agent turns, read a wall of text and then swap between a dozen apps and agents, burdening you with a context switch at every step. The point of this shift is to orchestrate that integration for you.

What is a GUI moment for Agents?

Here is what we think:

Surfaces — objects with a location. A conversation log is a long scroll of mind-numbing text where everything scrolls away and nothing has a stable address (how many of us have returned to our previous chat session, unsure of what happened here or even if this was the right session). Work, by contrast, needs persistent places you can return to and hand off to someone else.

Trust — every call, with its arguments. An agent that simply summarizes or cites specific sections is asking to be trusted. An agent that shows the work in easily digestible format is offering to be checked. That is what allows users to delegate with confidence instead of being forced to audit every result themselves.

Safety — permissions you write down. Read the calendar, read the inbox, read the issues, read documents, then write by sending a Slack message. That is the core of it, autonomy is defined by a set of explicit permissions. One agent can read across systems but write in exactly one place. Ask it to file a ticket outside that scope, and it declines explicitly. Narrow permissions are what make delegation manageable and trustworthy.

Latency — assembly you can watch. A spinner or the word “deliberating” or “razzmatazing” asks for patience without providing anything in return. Instead, break the work into pieces that arrive independently, with each pane filling as its evidence becomes available, the channel at 30 seconds and the brief at 55. The wait becomes something to read rather than something to endure.

Scale — many small agents, in parallel. One objective, several small agents, each scoped to a single source, none of them requiring direct interaction, running in parallel and reporting to a single surface. Supervising an assistant is a job; reading a consolidated report is a glance.

In Closing: The Future of Computing

A big reason why Claude Code succeeded was because of their terminal capabilities. The UI was perfect for developers and engineers, which then attracted hobbyists and the general public. Sadly, while this chat based interface is great for coding, it does not serve the general-purpose users and the white-collar knowledge workers. Our hypothesis: that something like what we've outlined above is the next UI evolution for agents, and this is what will support use and adoption in the real and very busy, AI-overloaded world.

The AI Transformation Breakthrough Your Org Chart Needs

Cory Janssen — Tue, 21 Jul 2026 16:00:00 GMT

As AI integration deepens, the organization redesigns itself around it.

Most companies think about AI adoption as a rollout. Buy the tool, train the team, measure the productivity lift. That framing misses the bigger shift underway.

AI transformation happens in phases, and each phase changes the shape of the organization itself. What starts as individual tools becomes shared infrastructure. What starts as personal productivity becomes organizational speed. By the final phase, the org chart looks different, not just the workflows inside it.

Here's how we see it play out.

Phase 1: Individual Augmentation

In this phase, everyone gets an AI assistant. The org chart doesn't change. CEO, VP, director, individual contributor: the reporting lines stay exactly where they were.

Every employee has access to AI, whether that's Claude Cowork or a similar assistant. Each person becomes individually more productive.

The gains are real, but they don't compound. One director's faster report writing doesn't make the whole department significantly faster. One IC's quicker research doesn't change how the team works together. The organization has more AI and some individuals can produce more, but teams don't yet work any differently.

Phase 2: Department Workflows

In the second phase, the unit of change shifts from the person to the team.

Departments start sharing data. That shared data lifts everyone in the department at once, not just the people using AI the most. Team workflows get automated: demand forecasting, reporting, routine coordination.

This is where “AI+” starts to mean something. A department with shared context and automated workflows moves faster than the sum of its individually augmented employees.

But the silos remain. The CEO still hears about progress through the VP or department lead, the same way they always have. Information still moves up the chain one layer at a time. What's changed is how fast and how well each layer moves on its own.

Phase 3: Org-Wide AI-Native

The third phase is where the structure itself changes.

The organization flattens. Mid-level management, built originally to coordinate and relay information, becomes less necessary as AI absorbs the routine coordination work. Former individual contributors move into higher-value roles, closer to judgment calls and away from repetitive tasks.

Cross-functional workflows span departments instead of stopping at their edges. Data that used to live in one team's systems becomes part of an org-wide layer, and workflows built on top of it don't care which department they cross.

The clearest signal of this phase is what the CEO sees. For example, instead of quarterly reports built manually in spreadsheets two weeks after the quarter closes, an automated order-to-cash workflow can produce financial visibility on demand. The decision-making isn't just faster. It's structurally different: real-time instead of retrospective.

What This Means at the Department Level

Zoom into any one department in Phase 3, and the shift becomes concrete.

Agentic workflows take over the routine, high-volume work: the transactions, the repetitive steps, the tasks with clear rules. That doesn't eliminate the people who used to do that work. It moves them.

Former individual contributors become human-in-the-loop, or HITL, operators. Their job is no longer to execute the routine work themselves. It's to handle exceptions, apply judgment where the workflow can't, and bring domain expertise that AI doesn't have. That expertise feeds back into the workflows, which keeps improving the Agentic system over time.

Two roles emerge to build and run this system. A workflow engineer maintains the shared platform underneath every department's workflows, the infrastructure layer everyone draws on. A workflow owner designs and owns each specific workflow, whether that's order-to-cash, demand-to-supply, or insight-to-action, making sure it reflects how the business actually needs to run.

The result is a department where the business owner, formerly a VP focused on outcomes and KPIs, oversees a small team of HITL experts instead of a traditional hierarchy of directors and reports. Teams have greater productivity and decisions are made on real-time high quality data.

The Through-line

Across all three phases, one idea holds: AI transformation isn't something that happens to an org chart. It's something that eventually redraws it.

Phase 1 proves individual value. Phase 2 proves team value. Phase 3 proves that value can move across the whole organization, in real time, without waiting for it to climb the ladder one layer at a time.

The companies that get this right aren't the ones with the most AI tools. They're the ones willing to let their structure change as the work changes underneath it.

Build vs. Buy vs. Book a Ride: A Better Way to Think About AI.

Cory Janssen — Fri, 10 Jul 2026 16:00:00 GMT

Every executive weighing an AI investment eventually asks the same question: build it ourselves, or buy something off the shelf? It's the wrong question, or at least an incomplete one. There's a third option most organizations don't consider, and it's usually the one that actually gets them where they're going.

Think of it this way. When you need to get somewhere important, you have three options.

Option one: buy a car

This is building AI in-house. On paper, owning the vehicle sounds appealing: full control, no dependency on anyone else. In practice, it means months of procurement before you've gone anywhere, including hiring data scientists and ML engineers, standing up infrastructure, and negotiating compute contracts. Once you have the car, you need a team to drive it, plus the maintenance, insurance, and upkeep that come with keeping it running as conditions change. And even with all that investment, there's no guarantee it goes where you need. AI initiatives fail for a lot of reasons that have nothing to do with the quality of the car, including an unclear destination, no one in the organization who's actually driven this route before, or a model built for a road that no longer exists by the time it ships.

Option two: rent a car

This is buying software: an AI platform or point solution licensed off the shelf. It's faster to get behind the wheel, and the sticker price looks a lot more manageable than option one. But renting a car doesn't mean you can drive it. You still need someone on staff who knows the roads, understands the controls, and can react when something goes wrong. And if it breaks down, whether the model drifts, the vendor's roadmap doesn't match your business, or the "generic" tool doesn't fit your actual problem, you're on your own. Rented cars aren't built for your route. They're built for the average driver on the average road, which describes fewer businesses than the brochure suggests.

Option three: book a chauffeured ride

You tell us where you need to go. We get you there safely, efficiently, and with a driver who's done it before. That's AltaML.

This is the option most companies don't realize exists, the middle ground between hiring an entire AI team and buying a subscription and hoping for the best. A chauffeured ride means you're not procuring a fleet, and you're not left to figure out the controls yourself. You state the destination, the business problem, the outcome you need, and an experienced partner handles the driving: the platform, the engineering, and the judgment calls that only come from having made the trip before.

Why the driver matters more than the car

The reason build-versus-buy is the wrong framing is that both options quietly assume the hard part is acquiring the vehicle. It isn't. The hard part is knowing the roads: which problems are actually solvable with AI today, where the potholes are, and how to adjust when the destination shifts halfway through the trip. A car, whether you built it or rented it, doesn't know any of that. A driver does.

That's the case for a true AI implementation partner over either extreme. Not a staffing agency that hands you a car and a set of keys, and not a software vendor that hands you a rental agreement and a support line. A partner who takes ownership of the outcome, brings a platform that's already been road-tested, and stays accountable for the whole trip, not just the parts that are convenient to support.

What this looks like in practice

At AltaML, that means we don't just deliver a model and walk away, and we don't just license you a tool and wish you luck. We bring the platform, the team, and the accountability of a true implementation partner. We've done this drive before, across industries and problem types, which means we recognize the difference between a detour and a dead end faster than a team doing it for the first time. And because we're accountable for getting you there, not just for handing over a vehicle, our incentives are aligned with your outcome from day one.

You bring the problem. We bring the ride.

If you're staring down a build-versus-buy decision on AI, it's worth asking a different question first: do you actually want to own and maintain a car, or do you want to arrive? For most organizations, the honest answer is the second one. They don't need another asset to manage. They need to get somewhere important, reliably, and preferably with someone who already knows the way.

That's what a chauffeured ride offers. You bring the problem. We bring the platform, the team, and the accountability to solve it.

What Responsible AI Actually Means in Practice (Not Theory)

AltaML — Fri, 20 Feb 2026 17:00:00 GMT

AI adoption is accelerating faster than governance can keep up. Responsible AI (RAI) now demands real operational safeguards—not just ethical talk. Without practical controls, organizations face rising legal, financial, and reputational risks.

Organizations are deploying AI into core workflows, but visibility into how those systems operate remains limited. The Cyera 2025 State of AI Data Security Report found that 83% of organizations are using AI, but only 13% have visibility into how it operates. As adoption grows, understanding these risks becomes crucial for responsible oversight.

In 2024, Air Canada was found liable for an error made by its AI-powered chatbot. The chatbot told a customer that they could make a bereavement claim after their travel was completed. After purchasing a full-price fare and taking their flight, the customer learned that the claim had to be made at the time of purchase. The customer took Air Canada to court, where the judge ruled in the customer’s favor, stating, “it makes no difference whether the information comes from a static page or a chatbot.”

The message is clear: you can’t pass off responsibility to an algorithm. If your brand appears on the interface, you are responsible for what it produces.

From our experience deploying AI into high-stakes operational environments, we often see deployment move faster than oversight. This gap underscores the urgent need for strong governance and real-world safeguards. Moving from principle to action is essential.

Responsible AI: Great in Theory, Challenging in Practice

This is where responsible AI (RAI) comes in. When done right, RAI speeds up adoption. Most organizations know why it’s important, but few know how to make it work in practice. To bridge the gap, a shift from ideals to operational solutions is required.

Responsible Artificial Intelligence (RAI)

The practice of designing, building, and deploying AI systems that are predictable, defensible, and aligned with the environments they operate in. Rather than a static list of ethics, RAI is a core engineering requirement that transforms abstract principles into operational guardrails to ensure visibility, mitigate liability, and accelerate organizational adoption.

From RAI Principles to Productions: The 7 Operational Pillars

It’s easy to say you value transparency and fairness; it’s much harder to build the systems that make those values real. Most organizations get stuck because they lack the functional guardrails that teams can actually follow.

At AltaML, we treat RAI as a core requirement, not a compliance checkbox at the end of a project. It is a blueprint for building better solutions. To move RAI from the boardroom to workflow, our approach is organized into seven operational pillars, which we outline next

By the end of this post, you’ll understand how to:

Spot the Breakdown: Identify real-world traps where AI systems can fail.
Apply Practical Controls: Implement specific checks and balances needed to keep your system predictable.
Close the Gap: Turn high-level organizational values into consistent, reliable results.

1. Inclusivity: Audit Representation Before You Build

Inclusivity means making sure the people affected by an AI system are represented in its design and data. If you ignore inclusivity, bias can creep in early, often through flawed proxies or gaps in representation.

The problem is rarely malicious intentions. It’s the hidden assumptions built into data well before the system goes live.

The Breakdown

A U.S.-based healthcare risk algorithm allocated care management services based on predicted future healthcare spending as a proxy for medical need. Certain communities had historically lower access to care and, therefore, lower spending. The model equated lower spending with lower need and systematically deprioritized these patients.

The failure wasn’t mathematical. The assumption was that spending accurately reflected illness burden. When researchers replaced cost-based predictions with direct measures of clinical illness—such as chronic condition counts and diagnostic risk indicators—the bias was significantly reduced.

Practical Controls

Stakeholder Mapping: Identify exactly who is impacted by the tool before you start building.
Proxy Variable Audits: Double-check whether your inputs, like cost, location, or engagement, are serving as proxies for unfair bias.
Data Diversity Review: Confirm your training data accurately reflects the real-world environment where the AI will live.

2. Fairness: Test for Real-World Impact

Fairness means that similar people receive similar treatment. While inclusivity checks who is represented in the system’s design, fairness examines the outcomes—linking design to impact. AI doesn’t create bias, but it can spread quickly if not managed correctly.

The Breakdown

Amazon stopped using an internal hiring tool after discovering it penalized resumes containing indicators associated with women’s colleges. The model was trained on historical hiring data from a male-dominated field, so it learned past biases rather than focusing on merit. This outcome was preventable, highlighting the danger of scaling historical data without first testing for disparate impact.

Practical Controls

Outcome Testing: Measure performance across demographic groups, not just overall accuracy. High aggregate accuracy can conceal unequal outcomes.
Stress-Testing (Bias Red-Teaming): Try to break the system by feeding it edge cases to see if it defaults to unfair patterns.

3. Transparency: Make Decision Logic Defensible

If an AI system denies a loan or recommends a medical treatment, saying “the model said so” isn’t a valid answer. Transparency is about knowing where AI is used. Explainability means being able to answer, “Why did this happen?”

The Breakdown

When Apple launched its credit card in 2019, some customers reported significant differences in credit limits between spouses with similar financial profiles. Regulators investigated Goldman Sachs, the issuing bank. While they didn’t find proof of intentional discrimination, it did expose problems with how credit decisions were documented, explained, and communicated. The bank couldn’t clearly explain how decisions were made, which eroded public trust.

Practical Controls

Model Lineage Documentation: Maintain clear records of training data sources, feature selection logic, model versions, and deployment history. When failures occur, traceability enables root cause analysis and continuous improvement.
Explainability Tooling: Use feature attribution or similar techniques to generate decision rationales that can be reviewed internally and communicated externally.
Defensibility in Design: Ensure that decisions affecting individuals can be explained in plain language and reviewed by qualified personnel when challenged.

4. Safety and Security: Set Boundaries for Autonomy

If you give AI power without limits, it’s a recipe for unpredictability. As AI systems gain greater autonomy and access to tools, they must be designed to prevent harm. Errors can propagate faster and at a larger scale. Safety governs what the system is allowed to do. Security governs who can access or influence it.

The Breakdown

Security researchers found a prompt injection flaw called EchoLeak in Microsoft 365 Copilot. This flaw could have enabled remote data exfiltration via a crafted email. The vulnerability was disclosed and patched before public exploitation. This was a perfect example of why AI systems need robust digital safeguards to prevent them from carrying out harmful actions.

Practical Controls

Risk Tiering: Classify AI use cases by potential impact. High-stakes applications require stricter safeguards and human oversight.
Autonomy Constraints: Limit what tools the system can access and what actions it can execute independently.
Behavioral Guardrails: Apply output moderation, rate limits, domain restrictions, and instruction-context protections to reduce susceptibility to prompt injection and unsafe outputs.
Infrastructure Security: Enforce authentication, encryption, access controls, and continuous monitoring to prevent unauthorized manipulation.

5. Accountability: Know Who Owns the Output

Algorithms can’t be held responsible in court—only people can. Accountability means every AI decision has a clearly defined owner. It establishes who is responsible for oversight, escalation, and remediation before a system goes live. As the Air Canada chatbot case illustrated, if your brand is on the interface, you own the output.

The Breakdown

In the Netherlands, an automated fraud-detection system used to administer childcare benefits falsely accused thousands of families of fraud. Rigid enforcement rules and limited avenues for appeal meant affected individuals had little recourse. The resulting scandal led to investigations and the resignation of the Dutch government in 2021.

The problem started with poor automation but got worse because there was no clear owner, no effective oversight, and no way to escalate issues.

Practical Controls

Clear Ownership: Designate an accountable person for every AI system across technical, legal, and operational domains.
Escalation Paths: Establish formal processes for reviewing contested decisions and correcting errors.
Human Override Authority (The Kill Switch): Ensure a human always has the authority to intervene, suspend, or modify system behavior when necessary.
Governance Reviews: Treat failures as organizational events, not isolated technical bugs. Capture lessons and update controls accordingly.

6. Privacy: Build Protection from Day One

AI runs on data that is often sensitive, personal, or proprietary. Privacy and data protection ensure this information is handled in compliance with the law and in ways that respect individuals’ rights.

Beyond preventing breaches, privacy is about ensuring data is used legally, fairly, and only when needed, right from the beginning. When privacy is treated as a compliance afterthought, regulatory and reputational risk escalate quickly.

The Breakdown

Clearview AI built a facial recognition system using billions of images scraped without consent. Regulators worldwide, including some in Canada, ordered the company to delete the data because the system’s very foundation violated privacy laws.

Practical Controls

Lawful Basis Assessment: Establish a clear legal basis for collecting and processing personal data before development begins.
Data Minimization: Collect only necessary data. Avoid defaulting to full datasets.
De-Identification Practices: Anonymize or pseudonymize data used for model training and inference, where possible.
Retention and Deletion Policies: Define clear timelines for storing AI-related data and enforce deletion protocols.

7. Awareness: Make AI Use Visible and Questionable

RAI shouldn’t be hidden. Transparency lets people examine and explain systems, while awareness helps users know when AI is affecting results and lets them ask questions. If we treat AI outputs as always correct, we create bigger risks.

The Breakdown

Zillow provides AI-based “Zestimates” of users’ home values. Though clearly labeled as an estimate, many users treated it as authoritative. Zillow even relied on these estimates to drive a multi-million dollar home-buying strategy. Overreliance on those forecasts, without human skepticism, led to a $300 million loss and the closure of its Zillow Offers division.

Practical Controls

Clear Disclosure: Inform users when they are interacting with an AI system and explain the role it plays in decision-making.
Visible Limitations: Surface uncertainty ranges, assumptions, and known constraints in plain language.
User Recourse: Provide accessible pathways for users to request human review or challenge automated outcomes.
Internal Escalation Culture: Empower employees to question AI outputs or flag potential misuse without fear of reprisal.

Responsible AI Is an Accelerator

Many organizations see governance as something that slows them down. When in fact careful oversight actually helps them grow and scale.

When you can see how your models work, trust your data, and have clear ways to handle problems, you remove the uncertainty that slows down deployment. In the end, the real advantage won’t go to the fastest AI adopters, but to those who use AI responsibly and at scale.

When you turn broad principles into real safeguards, AI shifts from a risk to a source of lasting competitive advantage.

Agentic AI: What It Is, Isn’t, and Why It Matters | AltaML

Cory Janssen — Wed, 04 Feb 2026 17:00:00 GMT

People talk about agentic AI in all kinds of ways, but the idea is pretty consistent. It’s a system that can help to expedite how work moves forward by reasoning through tasks and taking autonomous action. It does this alongside human involvement. This article offers a grounded, practical definition of agentic AI for business leaders.

In 1996, in a small office park outside of Houston, Texas, a group of executives at Compaq Computer were planning the future of the internet business. They called this future “cloud computing,” and that phrase went on to revolutionize the technology sector, despite confusing the vast majority of the population. Well, the tech industry is back at it with bad naming, this time with “AI agents,” or “agentic AI.”

What exactly is an AI agent? Depending on who you ask, the answer changes. For some executives, it’s a label for lightweight workflow automations. For others, it points to something far more ambitious.

Before reaching any kind of conclusion, it’s worth slowing down and getting clear on why this ambiguity exists, and more importantly, how the leading voices are using it.

Before We Begin: Here’s Why Definitions Vary

The phrase “agentic AI” sits on a spectrum with no shared boundaries, which helps explain why definitions differ so dramatically. A major driver of this variation is the disconnect between how academics use the term and how it’s marketed commercially.

At the academic end of the spectrum, “agentic” refers to systems that can understand a goal, break it into steps, make decisions along the way, and adjust their plan as new information comes in.

At the commercial end of the spectrum, however, agentic AI is often used to label far simpler tools. This includes basic chatbots and prompt-driven workflows that perform a predefined action but don’t demonstrate meaningful planning or showcase autonomy.

Both ends of the spectrum get grouped under the same term, and it’s one reason why Gartner placed agentic AI at the “peak of inflated expectations” in its 2025 Hype Cycle for Emerging Technologies. The term is everywhere, but the capabilities it refers to are far from consistent.

AI Agents vs. Agentic Workflows: What’s the Difference?

Further complicating matters, there’s an important distinction to be made between an AI agent and a larger agentic workflow.

An AI agent is usually a single system built to handle one specific job. It might rely on a large language model or call in a few tools (like an API or a browser). It might also react to real-time inputs and work through a step-by-step sequence to complete a task.

Agentic workflows take things a step further. Instead of relying on one agent to do everything, they coordinate multiple agents. Each agent has a specific role, and they all work toward the same goal. One agent might plan while another gathers information and a third takes action. Agentic workflows can also adjust course as things change, adapting to the nature of work rather than rigidly adhering to a defined sequence of events.

With this distinction in mind, let’s take a look at how leading voices in AI are using the terms.

What Do the Leading Voices Mean When They Say “Agentic”?

Major research labs, leading firms, and market analysts describe agentic AI in a variety of ways, but most cluster around a few core ideas:

OpenAI emphasizes that agentic AI is a system “that can pursue complex goals with limited direct supervision.” This means the technology can take actions, use tools, and complete tasks with minimal human intervention.
Google DeepMind defines agentic AI as “an advanced form of artificial intelligence focused on autonomous decision-making and action.” Or, to put it in the words of Google’s CEO, Sundar Pichai: “Anywhere you can describe a task in natural language, [agentic AI] can act on your behalf to accomplish that.”
Andrew Ng, a British-American computer scientist and one of the leading voices in agentic AI, defines agentic workflows as something that can “iterate over a document many times.” He says that agentic workflows can plan outlines, decide what web searches are needed to gather more information, write a first draft, and even revise the draft taking into account any weaknesses.
Dario Amodei, CEO of Anthrophic, summarizes his definition of agentic AI as a “country of geniuses in a datacenter.” What he means is this: AI systems can be assigned complex, long-running goals and then independently plan, coordinate, and take action across digital and physical systems.
IBM focuses on the workflow dimension. They describe agentic AI as software that “exhibits autonomy, goal-driven behavior, and adaptability” and that can execute tasks across systems and data sources, acting as a digital worker inside an enterprise.
McKinsey and Gartner take a business-first approach. McKinsey describes agentic systems as “a software component that has the agency to act on behalf of a user or a system to perform tasks,” while Gartner leans on agentic AI’s “potential to automate interactions and processes, both for service teams and for the customers making requests.”

The common throughline between all of these definitions is this:

Agentic AI

Refers to workflows powered by large language models that can reason, use tools, and take action alongside humans to achieve a specific outcome—especially in situations that require judgment rather than rigid automation.

With this in mind, let’s unpack the common threads that sit at the core of most definitions.

The Common Threads Between These Definitions

Different organizations may use different words, but they’re describing the same behaviors. These shared capabilities offer the clearest picture of what agentic AI actually is, and they begin with one fundamental shift:

1. Agents act, not just respond.

Across definitions, one of the most consistent themes is that agentic AI takes initiative. At the risk of stating the obvious, AI agents have agency. Unlike traditional generative AI models that wait for the next prompt from a person, agentic systems are expected to take action once a task has been set.

This shift from passive response to active behavior marks a fundamental change in how AI is set to contribute to work inside an organization.

2. Agents pursue a goal

Another shared element is goal orientation.

Agentic AI isn’t designed to provide isolated outputs based on a single input. Instead, it’s designed to autonomously move toward an outcome.

Whether the goal is to generate a report, reconcile data, or triage a request, the system is expected to make progress without any step-by-step handholding. This pursuit of an end state is central to how the leading voices are defining agentic capabilities.

Agentic AI isn’t designed to provide isolated outputs based on a single input. Instead, it’s designed to autonomously move toward an outcome.

3. Agents execute multi-step plans

Nearly every definition of agentic AI points to the ability to move through work in multiple steps. What differs is how those steps are determined.

In some systems, the plan is probabilistic. For example, the agent figures out what to do next as it goes and adapts based on context. This is the model often highlighted in research discussions, where agents dynamically reason and explore paths toward an outcome.

In many practical enterprise workflows, however, the plan is more deterministic. The steps are predefined or strongly guided by humans, policies, or business rules, and the agent’s role is to execute and adapt within the margins of what’s acceptable.

4. Agents can interact with systems on behalf of humans

Finally, agentic AI is expected to operate within real workflows. That means using tools, accessing data, triggering processes, and completing tasks that previously required human intervention.

Of course, a human stays in the loop to guide decisions and own the outcome, but the ability to interact and cooperate with various systems is what distinguishes agentic workflows from other AI models that only generate text or provide basic recommendations.

The Simple Definition

Agentic AI is a collaborative system that helps turn a goal into a finished result. You tell it what you’re trying to achieve, and it works with you to get there.

It does this by “reasoning through the work”, so to speak. That is to say: it deploys the right tools at the right time, navigates obstacles as they come up, and loops a human in where needed to provide judgment, direction, and accountability.

Agentic AI: Turning a Definition into Forward Direction

Almost 30 years ago, the phrase “cloud computing” was little more than a vision. But a shared language gave organizations the clarity needed to experiment with it, invest in it, and eventually transform entire industries.

Agentic AI is at a similar moment in time. And by examining how leading voices define agentic AI, a clear picture emerges: Agentic AI acts rather than responds, pursues outcomes rather than prompts, executes multi-step responses, and interacts with tools and systems on a human’s behalf.

When leaders understand what this language actually means, they can better evaluate opportunities, distinguish marketing hyperbole from real capabilities, and start to make informed decisions about where agentic systems can fit within their existing workflows.

Are you ready to learn how a grounded, real-world approach to agentic AI can help your teams work smarter and deliver stronger outcomes?

Building Canadian Communities with Homegrown AI | AltaML

AltaML — Tue, 02 Sep 2025 16:00:00 GMT

AMO 2025 Conference Recap

The Association of Municipalities of Ontario (AMO) 2025 Conference brought together public sector leaders from across Ontario to discuss how municipalities can build smarter, faster, and stronger for the future. For AltaML, it was an opportunity to listen, learn, and share how AI can help municipalities address their biggest challenges—whether that’s improving housing affordability, enhancing citizen services, or navigating economic uncertainty.

“The AltaML team was proud to sponsor AMO 2025. It was inspiring to see this record-setting crowd come together to collaborate on how best to solve the challenges facing municipalities in Ontario. We were delighted to have the chance to be involved in these important conversations.” – Chantal Ritcey, Public Sector Lead and Responsible AI Lead at AltaML.

A Roadmap for AI Adoption: From Potholes to Progress

At the conference, we aimed to educate Ontario municipalities on how they can use AI to streamline operations, enhance services, and improve decision-making by sharing real-world examples from our work with public sector clients across the nation. Our presentation, From Potholes to Progress: AI’s Role in Municipal Innovation, drew a standing-room-only crowd and sparked conversations that continued long after the session ended.

The presentation highlighted real-world applications of AI we’ve developed with our public sector clients, already in use today, including:

Automating Permit Application Screening: Speeding up approvals and freeing staff for higher-value work
Detecting Sanitary Pipe Defects: Improving infrastructure maintenance and reducing repair costs
Enhancing Citizen Services: Using AI assistants to reduce call wait times and improve access to information

We also introduced a roadmap for AI adoption to help municipalities get started with confidence:

Make AI an Organization-Wide Imperative: Ensure leadership buy-in and a shared vision
Build a Multi-Disciplinary Team: Combine policy, operations, and technical expertise
Start Small, Build Momentum: Pilot projects that deliver quick wins and demonstrate value
Focus on the Business Problem First: Let community needs, not technology, drive decisions

The core message was clear: municipalities recognize the potential of AI, but many require a structured approach to get started.

“Ontario municipalities have demonstrated their drive to innovate to serve their citizens better. The conversations and the demand for our presentation are clear indicators that municipal governments are ready to tackle the challenges facing their communities. It was energizing to see how excited the audience was to learn about our experiences helping municipalities innovate with AI.” – Chantal Ritcey, AltaML.

Key Insights from Ontario’s Municipal Leaders

Across sessions and conversations, municipal leaders raised common themes that highlight both the urgency and the opportunity for AI adoption.

Trends That Are Shaping the Future

Building smarter with the right partnerships and innovation strategies
Speeding up affordable housing development and approvals
Expanding access to healthcare, particularly in underserved regions
Finding new ways to collaborate across municipalities during uncertain times

Hurdles on the Path to Progress

Balancing innovation with tight budgets
Struggling with where to start on AI adoption
Managing affordability and the pace of new housing starts
Navigating supply chain pressures and tariffs
Keeping pace with the speed of technological change

The Opportunity for Homegrown Innovation

Stronger collaboration between federal, provincial, and local governments on digital transformation
A renewed focus on homegrown AI innovation to drive productivity and resilience
Municipalities are eager to share lessons learned collectively rather than work in silos

Where AI Can Help Municipalities Now

Based on what we heard, there are clear areas where AI can deliver immediate, citizen-focused impact. Municipal leaders should be exploring:

Permitting and Approvals: Automating repetitive checks to reduce bottlenecks
Infrastructure Maintenance: Predictive monitoring for roads, water, and utilities
Housing Development: Data-driven forecasting to accelerate planning and approvals
Citizen Engagement: Digital assistants to improve service delivery and responsiveness
Budget Optimization: AI models to support evidence-based decision-making during fiscal constraints
Public Safety: Intelligent resource allocation to improve emergency response times

Each of these areas offers an opportunity to start small, demonstrate value, and build momentum toward broader adoption.

“The true opportunity for AI lies in the everyday. From automating permit checks to optimizing public services, AI can address the most pressing, tangible problems municipalities face right now. The key is to start small and focus on areas where we can deliver immediate, citizen-focused impact to build trust and momentum.” – Ankur Pandit, Public Sector Account Executive at AltaML.

Conversations That Stood Out

“The power of local governments in Ontario is incredible. It was amazing to see so many different cities and towns united by a common purpose—tackling shared, complex problems, from designing public washrooms to the broader use of mass timber. My key takeaway is that there are so many important issues on each municipality’s plate, and that’s where AI can be a real game-changer. By handling the ‘tedious chores,’ AI enables them to focus on what matters most to their citizens. The excitement from local leaders about this potential was really encouraging.” – Paul Longo, Executive Business Partner, Public Sector at AltaML.

Reflections on the Experience

Beyond municipalities, our team also connected with federal leaders in Ottawa. Under the Carney administration, there’s a renewed emphasis on using AI to drive productivity within the public service—an encouraging sign of alignment across levels of government.

“The advanced nature of conversations around AI this year was striking. Municipal leaders are asking about governance, policies, and serious strategies for AI adoption. There’s a real shift toward putting AI to work in meaningful ways.” – Ankur Pandit, Public Sector Account Executive at AltaML.

Looking Ahead in Ontario

The AMO 2025 Conference gave us a valuable pulse check on the priorities of Ontario municipalities and reinforced how determined leaders are to innovate. With offices in Toronto and Waterloo, deep ties to Ontario’s post-secondary institutions, and proven expertise in AI for the public sector, AltaML is well-positioned to help municipalities take the next steps on their AI journeys.

“Innovation is critical to building better communities. As we strive to build better and faster, we should also prioritize building Canadian. Innovation and the resulting intellectual property serve us best when developed and maintained locally. AltaML has a proven track record in developing AI solutions for municipal innovation, and we welcome the opportunity to leverage that expertise to benefit communities across Ontario.” – Paul Longo, Executive Business Partner, Public Sector at AltaML.

From permits to potholes, the opportunity to use AI to serve citizens and governments better is immense. If you’re ready to unlock the potential of AI in your community, let’s talk about how we can help you transform public sector opportunities into tangible innovation.

Uncovering Responsible AI’s Biggest Challenge: Privacy and Fairness

AltaML — Mon, 13 Jan 2025 17:00:00 GMT

Technical breakthroughs in recent years have exposed ways to train machine learning models that combat the influence of human bias and give fair predictions for all. The primary theme in all aspects of fairness and bias work is the assumption that individual demographic data exists and can be accessed during the model training phase.

The assumptions necessary for effective debiasing often conflict with privacy and data protection principles. This creates a critical challenge in responsible AI (RAI) development, posing risks and presenting a concern for every AI developer.

Why Bias Analysis Matters

AI systems have become a common part of daily life. From AI-powered assistants integrated into common software platforms (e.g., ChatGPT-Siri integration) to machine learning optimizing e-commerce recommendations (e.g., Amazon recommendation system), AI is a key element of many digital experiences. Organizations worldwide are investigating ways to use their data to improve operations, reduce costs, and increase revenue. AI offers new avenues for data use, enabling automated trend analysis and future predictions.

As with most emerging technologies, potential benefits come with potential risks. Companies developing AI systems need policies and processes to avoid risks and ensure AI has the intended positive impact on society. AltaML has developed seven key principles to guide the ethical development of AI, including a fairness principle ensuring AI systems do not discriminate based on protected characteristics.

AI systems have the risk of exhibiting discriminatory behavior. Without intervention, AI training algorithms may learn and perpetuate biases in data, leading to discriminatory predictions. For example, the criminal recidivism system COMPAS misclassified Black defendants as re-offenders at twice the rate of white defendants. Similarly, Amazon’s attempt to create a recruiting tool using historical data inadvertently discriminated against female applicants, reflecting the historical gender imbalance in the tech industry and the ability of machine learning algorithms to exploit correlations between variables.

Fortunately, toolkits are available to allow developers to create AI systems while mitigating the risk of harmful discrimination. Techniques include statistical tests and metrics for identifying bias, visualizations for explaining the effect of bias, and machine learning algorithms to augment training to produce fair results. These techniques are considered easy to implement and complement approaches that machine learning developers are already familiar with.

Privacy Risks in Bias Analysis

On paper, derisking a use case for fairness seems simple. In practice, programming with fairness in mind is still quite difficult. Every step taken to effectively debias models requires access to variables that describe protected characteristics. For example, to assess and mitigate the risk of racial discrimination, developers must know the race of individuals in the datasets.

In order to understand where the challenges come from, the principle of “Privacy and Data Protection” must be acknowledged. This principle calls for the protection of individual privacy throughout all stages of AI development and application. Use cases that make predictions regarding individuals and individual characteristics ultimately use data about the individual and, therefore, can present privacy risks.

There are various regulatory reasons accessing demographics for debiasing is not always possible. The General Data Protection Regulation (GDPR) in the European Union and the California Consumer Privacy Act (CCPA) in the U.S. protect consumer privacy rights regarding digital products and services, setting the groundwork for global best practices. These frameworks mandate that consumers provide data only for intended purposes, with clear consent and the ability to delete or remove sensitive information. Specific industries that collect sensitive data may have their own regulations and best practices. For example, in the U.S., the Health Insurance Portability and Accountability Act (HIPAA) restricts the movement and usage of health care data, including demographics.

Government institutions, organizations, and corporations that develop AI systems with data collected internally or by a third party will find that accessing demographic information to debias an AI system is at odds with privacy regulations and best practices meant to protect individual privacy. Companies will find that there is no established mechanism for accessing demographic information in a way that respects the principle of privacy and allows AI system development to protect against risks regarding fair service outcomes. How does this apparent contradiction shape responsible AI (RAI) as an emerging field, and how can we ethically advance it?

Call for Legal Solutions

The friction between privacy and fairness leaves practitioners of RAI in a challenging position. When faced with a choice between upholding privacy and managing fairness risks, companies likely fall into one of two types:

Companies which put RAI at the top of their operating mandate will likely simply avoid working on use cases that contain significant fairness and bias risks.
Since privacy regulations are common and come with high fines, a company without an RAI policy is likely to work on and deploy AI systems without mitigating fairness and bias risks.

Between both scenarios, individuals are left unprotected. In order to protect and avoid stifling innovation in AI, the world needs legal structures that enable the evaluation of fairness in AI systems:

Laws banning discrimination of individuals from automated systems need to exist, similar to how discrimination over protected characteristics between individuals is illegal in many countries.
Systems need to exist that allow the evaluation of an AI system on the basis of fairness and discrimination.

The enablement of fairness assessments on AI systems requires an ecosystem of regulations, digital systems, citizen awareness, auditors, and capable practitioners. An ideal solution would have the following properties:

Marginalized groups control their own data.
Demographics are collected solely for the purpose of ensuring fairness in AI systems.
Evaluation results are publicly available, ensuring transparency for any consumer of an AI system, whether a business or an individual.

Interim Techniques for Machine Learning Practitioners

Despite challenges, there is still a path ahead for AI practitioners to advance RAI forward. Although not fully established, a promising angle to explore involves utilizing location data and spatial aggregation. In fairness and bias analysis, variables can be proxies for demographics. Commonly, this is seen as a risk, but potentially there could be a way to use proxy variables as stand-ins for demographic data to suggest model fairness. One powerful proxy variable is location.

High level approach:

Determine a spatial granularity level in which location data is not personally identifying.
Overlay model output error on a map.
Observe non-uniformities in spatial distribution.
Compare non-uniformities in model distribution to known non-uniformities in demographic distributions (e.g., census).

Conclusion

Privacy and fairness are two important principles in computer ethics that have more relevance than ever with the proliferation of AI, each with unique risks. Without prioritizing one principle over another, developers of AI systems will find it difficult to proceed ethically with high-risk use cases. The rising issues affect both individuals and companies, presenting an increasingly critical challenge for RAI as a whole to overcome. As a society, we need to figure out a regulatory and commercial ecosystem that can support the requirements of RAI at scale, while enabling and enforcing best practices regarding privacy and fairness.

Where Ethics and Development Converge: Building Responsible AI | AltaML

AltaML — Mon, 21 Oct 2024 16:00:00 GMT

As artificial intelligence (AI) plays a bigger role in shaping data—not just how it’s analyzed but even how it’s produced or accessed through generative AI (GenAI), a type of AI that can create new content such as text, images, audio, or video by learning patterns from existing data—we have to take steps to be sure it will be used responsibly. The challenge is defining what that responsibility entails and coming up with a workable solution for responsible artificial intelligence (RAI). AltaML proposes a sustainable approach to addressing algorithmic bias, upholding ethical standards, and delivering better outcomes for all.

AI Business Aspirations

With AI becoming an integral part of operations across organizations, it raises a host of ethical questions about fairness, safety, privacy, and transparency of its applications. These concerns are now amplified by the growing use of GenAI among businesses.

Harvard Business Review reports, two-thirds of senior IT leaders reported plans to introduce GenAI to their operations within the year. Even while embracing the technology, they express concerns over its safety (79%) and its associations with biased outcomes (73%).

This duality relates to a point made by the UK’s Financial Conduct Authority (FCA), Nikhil Rathi, in warning about the need for “an open conversation about the risks and trade-offs” associated with AI.

“We want safe and responsible use of AI to drive beneficial innovation,” Rathi said.

Finding the way to use AI safely and responsibly begins with admitting the problems associated with their use. People cannot just be subject to AI determinations that can lead to unfair and inequitable outcomes. Now is the time to plan on finding the solutions, and it starts with identifying the problems.

Algorithmic Pitfalls in Hiring and Facial Recognition

“Even algorithms have parents, and those parents are computer programmers, with their values and assumptions,“ wrote Alberto Ibargüen, President and CEO of the John S. and James L. Knight Foundation.

The assumptions that were built into data models can have long-reaching effects. We’ve seen this happen with respect to the biases that have emerged from algorithmic reviews of job applicants and facial recognition.

Several years ago, Amazon launched an experiment using an automated rating system for job candidates. As the training data showed far more men in tech roles than women, the algorithm learned to select male candidates as the better choice for jobs.

The algorithm learned to apply this gender bias in its selection and to associate any indicators of female identity on a resume as a reason to filter out a candidate. The results were so skewed that Amazon was compelled to shut down its AI recruiting tool.

Amazon also drew fire for the biases that emerged from its facial recognition software called Rekognition that had been adopted by some law enforcement agencies before a great deal of pressure from researchers and civil liberties groups brought it to a stop in 2020.

Facial Recognition Is Accurate, if You’re a White Guy was the headline that encapsulated the problem. The article reported on Joy Buolamwini’s research that found that Rekognition misidentified women with darker skin tones as men 31% of the time.

In addition to the problem of racial bias and inaccuracy, Buolamwini pointed out the privacy violations that can result from adopting facial recognition. That problem did not go away in later iterations of facial recognition technology, as evidenced by Clearview AI. It builds its facial database by scraping people’s photos from the web without their consent. The company has been fined multiple times by Europe agencies for violating the General Data Protection Regulation (GDPR) privacy rules. While the company claims it’s exempt from GDPR because it’s based in the U.S. and doesn’t sell to European agencies, privacy regulations still apply. However, its latest fine brings the total it could owe to over $110 million, a substantial debt for a business.

What Must Be Done

The fact that an American company can still be forced to act in compliance with GDPR rules is highly relevant to how AI will evolve now that the EU AI Act is set to take effect in August 2026. Violators of the new regulations may face significant penalties.

There are three compelling reasons for taking immediate action on RAI:

Legal requirements
The moral imperative
The consequences of inaction

While there is no international law in effect, the EU law will have far-reaching consequences, likely operating like the GDPR in setting a precedent for privacy regulations put into effect in various jurisdictions across the globe. That means that even businesses in the U.S. or Canada that don’t have European interests would be prudent to keep EU rules in mind as they plan for future AI integrations that may come to fruition as new regulations emerge in different countries.

Being proactive about RAI is also essential to staying on the right track with technology. That means ensuring that AI will not be used to reinforce unfair treatment of individuals stemming from biases that are veiled by the algorithmic operation. Failing to take action to avert such outcomes ends up costing businesses forced to catch up later.

The costs include the technical debt and stagnation that result from failing to plan ahead for AI implementations that meet regulatory standards. An ounce of prevention is far more cost-effective than a pound of cure, especially when considering the delays that can derail projects if the technology requires adjustments or updates.

Ignoring RAI can also translate into losing the trust of customers and employees who will look elsewhere for a company that is not misusing AI. That would increase the difficulty of recruiting and retaining talent.

The 7 Principles of RAI

To guide businesses in balancing AI development with responsibility, AltaML has defined seven fundamental principles of RAI:

Inclusivity

Given that AI has an effect on all kinds of people, their perspectives should be represented in AI solutions. Different types of people have to be recruited to participate in its design and development to avert the racial, ethnic, and gender biases that can get baked in by homogeneous teams.

Fairness

We have to strive to keep human values at the center of AI development. That means not losing sight of the Human Rights Act that recognizes the need to protect people from being discriminated against for traits like age, gender, race, marital status, etc. For that reason, AI use cases that are likely to result in a discriminatory effect should be avoided. This—and the lesson from Amazon—may be why it’s not currently being used widely in recruitment. As of early 2024 only 14% of companies reported using AI for their talent acquisition, according to a global survey of professionals.

Transparency and Explainability

We have to break through the black box. At the most basic level, users have to be informed when they’re interacting with an AI system rather than a human. Beyond that, we should aim to provide visibility into the applications, development, and operations associated with AI so that both users and affected parties understand the what, why, and how of outcomes and be able to flag mistakes.

Safety and Security

We need to ascertain that AI is being used only for valid purposes and prioritize mitigating safety and security risks. This entails comparing the in-production predictions with the ground truths that emerge later.

Accountability

All AI systems are selected and deployed by humans who have to be held accountable for its operation and output. This entails having a subject matter expert (SME) human-in-the-loop to validate the predictions before they are acted on. There should also be regularly scheduled audits to confirm that all outputs are consistent with human rights and environmental values.

Privacy and Data Protection

When data that includes sensitive personal information is used by the program, anonymization must be maintained to safeguard privacy throughout the AI life cycle by adhering to privacy frameworks. For example, if the data on individuals includes their addresses, only the aggregate of locations should show up when displaying results.

Awareness and Empowerment

Public trust in AI depends on transparent communication that fosters understanding and empowerment. Teams should be encouraged to share any of their concerns about AI usage so that they can be addressed. It all starts with a commitment to RAI that is integrated into the company culture.

As AI takes on an increasingly larger role in daily functions, it’s important to remember that while it brings incredible efficiencies, human insight remains essential for guiding its use. The road to RAI begins with awareness of the problems and a commitment to striving for better outcomes going forward.

Navigating Bias in AI with Open-Source Toolkits | AltaML

AltaML — Mon, 06 May 2024 16:00:00 GMT

In an era of rapid technological advancement, the integration of artificial intelligence (AI) and machine learning (ML) has catalyzed revolutionary changes across industries, offering opportunities for data-driven decision-making. However, these advancements have also brought forth ethical concerns, notably the pressing issue of harmful bias in AI systems.

Harmful bias refers to biases within AI systems that result in unfair or discriminatory treatment toward individuals or groups. Incidents involving harmful biased AI algorithms, such as Amazon’s attempted deployment of a compromised HR tool and Northpointe’s discriminatory COMPAS system, have underscored the ethical dilemmas posed by unchecked biases in AI. In response to these challenges, open-source toolkits have emerged providing promising options for researchers and practitioners to identify, explain, and mitigate bias in AI systems.

Here we look into the potential of leveraging existing open-source toolkits to address bias in AI and offer recommendations to promote ethical AI practices.

AI Bias Amplifies Human Bias

Bias and discrimination are deeply ingrained in the historical fabric of human societies. The incorporation of AI into decision-making processes heightens the risk of biased outcomes, as illustrated in the figure below:

AI biases manifest in various forms, such as allocation harms and quality of service harms (Table 1), which can perpetuate existing inequalities and undermine fairness. Examples like the COMPAS and Amazon’s HR tool cases illustrate how AI biases can impact real-world outcomes, underscoring the urgency of addressing bias in AI systems.

Allocation Harms	This situation arises when AI systems either extend or withhold opportunities, resources, or information selectively for certain individuals or groups.
Quality of Service Harms	This refers to instances where AI systems exhibit disparities in performance across different demographic groups, leading to unequal levels of service quality. An example is when a facial detection model effectively recognizes faces of certain races but struggles with others.

Table 1: Types of Harms

In recent times, there has been a growing emphasis on tackling bias issues, with a particular focus on identifying two primary types: data biases and model biases. Data biases, originating from skewed or incomplete datasets, encapsulate inherent prejudices within the data. These biases can be further amplified by model biases, which stem from the algorithms and decision-making processes utilized.

Tackling Bias with AI Toolkits

To effectively combat harmful biases, open-source toolkits have emerged, providing practitioners with a range of metrics and techniques to assess fairness in AI systems. Notable examples include IBM’s AI Fairness 360, Microsoft’s Fairlearn, Google’s What-If Tool, and Aequitas. These toolkits have greatly facilitated efforts to address bias in AI by offering comprehensive frameworks for identifying and mitigating harmful biases.

Steps to Mitigate Bias:

Identify: Biases in machine learning can be grouped into two types: data biases and model biases. The first step is to identify data biases, which can be challenging due to limited access to protected variables such as gender, race, and age. Statistical tests like the chi-square test or group representation metrics available in open-source toolkits like AIF360 can be used for this purpose.
Explain: Once biases are identified, the next step is to analyze and understand how features present in the data contribute to the bias. Explainable AI tools, including but not limited to visualization techniques like partial dependence plots and Individual Conditional Expectation (ICE) plots, can help interpret model outcomes and explain biases.
Mitigate: Techniques for mitigating bias fall into four categories: pre-processing, in-processing, post-processing, and meta techniques. Pre-processing involves adjusting the dataset before training, such as reweighting samples to address biases. In-processing techniques are integrated into the learning process itself, such as debiasing with adversarial learning. Post-processing involves modifying the outputs of the model after training, such as threshold optimization to ensure fairness. Additionally, meta techniques, such as grid search over hyperparameters with respect to bias metrics, provide a higher-level approach to bias mitigation. These techniques are commonly implemented in open-source toolkits to help promote fairness in AI systems.
Communicate: Effective communication throughout the model-building process is crucial. Tools like Google Model Card toolkits can assist with documentation and communication during the handoff of AI models.

Available Toolkits

Toolkits	FairLearn	What-If Tool	AIF-360	Aequitas
Highlight	Best in Functionality	Can identify the most similar cases to compare predictions	The first comprehensive tool with more than 70 metrics	Most user-friendly with a decision tree provided to assist metrics selection
Language Options	Python	Python	Python and R	Python
Assessment and Mitigation	Both	Both	Both	Assessment only
Open Source Visualization Dashboard	Yes	Yes	No	Yes
Who Maintains It?	Microsoft	Google	IBM	School researchers
Notes	Define fairness in relation to potential harm	Requires data upload, no analysis on-premise which can raise privacy concerns	Focus on debiasing	Tool not as well maintained

Table 2: Open Source Toolkits Comparison and Highlight

Conclusion

The integration of open-source toolkits presents a promising approach to combat bias in AI systems. By leveraging these toolkits and adopting best practices for ethical AI, stakeholders can mitigate biases, enhance transparency, and promote fairness in AI-driven decision-making.

However, challenges remain. These challenges include the limited coverage of multi-classification and regression use cases, inconsistencies in terminology, and a lack of awareness among AI practitioners, particularly regarding access to protected variables crucial for effective bias mitigation. For example, collecting demographic data specifically for this purpose is not a common practice, presenting a significant hurdle in addressing bias effectively.

It is recommended that AI practitioners, users, and corporations prioritize the documentation of data sources, adherence to ethical principles, and transparent communication regarding bias mitigation strategies. At AltaML, we recognize the importance of ethical ML practices and are deeply committed to addressing bias in AI systems. Learn more about AltaML’s practices in responsible AI here.

AltaML’s Responsible AI Principles

AltaML — Wed, 20 Mar 2024 16:00:00 GMT

Ethical AI Implementation

Responsible artificial intelligence (RAI) is non-negotiable. At AltaML, we are committed to implementing robust controls, adhering to established policies, and internal directives to ensure the ethical deployment of artificial intelligence (AI) technologies. Our mission is to elevate human potential with applied AI, and this mission is underpinned by a dedication to education and awareness.

Our guiding principles are as follows:

Inclusivity

Inclusive AI development aims to benefit people and the planet for the betterment of all. This is best achieved through diverse participation in the design and development of AI solutions.

Fairness

AI development must prioritize human-centered values, including rights, equality, fairness, and justice, with safeguards to mitigate bias.

Transparency and Explainability

Transparency in AI involves disclosing its usage and enabling understanding of its development, operation, and outcomes. It’s about providing explainability for users or affected parties.

Safety and Security

AI systems must serve legitimate purposes, prioritizing robustness and reliability to build trust while mitigating safety and security risks.

Accountability

Organizations and individuals must be clear on who is accountable for the outputs of AI systems. AI outcomes should be audited to prevent conflicts with human rights and environmental well-being.

Privacy and Data Protection

Privacy must be safeguarded throughout the AI life cycle. Collaboration is key to establishing data protection measures and adhering to privacy frameworks.

Awareness and Empowerment

Public trust in AI hinges on accessible education and expert guidance. Transparent communication helps foster understanding and empowerment ensures teams are encouraged to bring forward concerns about the usage of AI.

AltaML blog

The Chat Box is the New Command Line

Some History

1983: The Whole System

1984: When Computing Last Changed Shape

2023: A New Form of Computing

2026: The Interface at a Turning Point

The Sparkle Button

The Chat Box

From Prompting to Curating

What is a GUI moment for Agents?

In Closing: The Future of Computing

The AI Transformation Breakthrough Your Org Chart Needs

Phase 1: Individual Augmentation

Phase 2: Department Workflows

Phase 3: Org-Wide AI-Native

What This Means at the Department Level

The Through-line

Build vs. Buy vs. Book a Ride: A Better Way to Think About AI.

Option one: buy a car

Option two: rent a car

Option three: book a chauffeured ride

Why the driver matters more than the car

What this looks like in practice

You bring the problem. We bring the ride.

What Responsible AI Actually Means in Practice (Not Theory)

Responsible AI: Great in Theory, Challenging in Practice

Responsible Artificial Intelligence (RAI)

1. Inclusivity: Audit Representation Before You Build

The Breakdown

Practical Controls

Stakeholder Mapping: Identify exactly who is impacted by the tool before you start building.

2. Fairness: Test for Real-World Impact

The Breakdown

Practical Controls

Outcome Testing: Measure performance across demographic groups, not just overall accuracy. High aggregate accuracy can conceal unequal outcomes.

3. Transparency: Make Decision Logic Defensible

The Breakdown

Practical Controls

Model Lineage Documentation: Maintain clear records of training data sources, feature selection logic, model versions, and deployment history. When failures occur, traceability enables root cause analysis and continuous improvement.

4. Safety and Security: Set Boundaries for Autonomy

The Breakdown

Practical Controls

Risk Tiering: Classify AI use cases by potential impact. High-stakes applications require stricter safeguards and human oversight.

5. Accountability: Know Who Owns the Output

The Breakdown

Practical Controls

Clear Ownership: Designate an accountable person for every AI system across technical, legal, and operational domains.

6. Privacy: Build Protection from Day One

The Breakdown

Practical Controls

Lawful Basis Assessment: Establish a clear legal basis for collecting and processing personal data before development begins.

7. Awareness: Make AI Use Visible and Questionable

The Breakdown

Practical Controls

Clear Disclosure: Inform users when they are interacting with an AI system and explain the role it plays in decision-making.

Responsible AI Is an Accelerator

Agentic AI: What It Is, Isn’t, and Why It Matters | AltaML

Before We Begin: Here’s Why Definitions Vary

AI Agents vs. Agentic Workflows: What’s the Difference?

What Do the Leading Voices Mean When They Say “Agentic”?

Agentic AI

The Common Threads Between These Definitions

1. Agents act, not just respond.

2. Agents pursue a goal

3. Agents execute multi-step plans

4. Agents can interact with systems on behalf of humans

The Simple Definition

Agentic AI: Turning a Definition into Forward Direction

Building Canadian Communities with Homegrown AI | AltaML

AMO 2025 Conference Recap

A Roadmap for AI Adoption: From Potholes to Progress

Key Insights from Ontario’s Municipal Leaders

Trends That Are Shaping the Future

Hurdles on the Path to Progress

The Opportunity for Homegrown Innovation

Where AI Can Help Municipalities Now

Conversations That Stood Out

Reflections on the Experience

Looking Ahead in Ontario