Ali Ansari, Founder and CEO of micro1 – Interview Series

Ali Ansari is the Founder and CEO of micro1, where he focuses on turning human expertise into training data and evaluations for advanced AI systems. His entrepreneurial background includes building a software development agency and founding peerlinc before launching micro1 in 2022. He studied computer science at UC Berkeley and Stanford, with a focus on reinforcement learning, and now works on the systems needed to source, assess, and coordinate expert contributions to AI development.

micro1 is a data lab that helps train frontier AI models and evaluate AI agents using expert human input and realistic work environments. Its platform brings together specialist recruitment, data production, quality review, and model evaluation across fields including law, healthcare, finance, engineering, and science. In this interview, Ansari discusses why demand for human expertise is growing, how operational data supports reinforcement learning, and the career opportunities emerging around AI training and evaluation.

You founded micro1 in 2022 after previously building and selling peerlinc. What problem were you originally trying to solve, and how did micro1 evolve into a company focused on training and evaluating frontier AI models?

Originally, it was a problem I had in my own business. I was running a software development agency, and every new project meant finding and interviewing more engineers. I was spending a huge amount of my time on that. So we built an AI interviewer, Zara, to help, and that eventually became micro1.

We started by selling the recruiting software and building a marketplace of vetted talent. Then a customer came to us needing hundreds of engineers to train AI models. We started digging into why they needed so many people, and realized there was this enormous demand for human expertise behind model development.

We had already spent years building the ability to source and vet people at scale. Human data was a very powerful application of that infrastructure. So we went all in, used our AI recruiter to source and vet expert talent, and built other data platforms and quality systems around it. Today, we help labs turn that expertise into better model capabilities.

You’ve argued that as AI models become more capable, demand for human expertise is actually increasing rather than disappearing. What is driving that relationship?

There are a couple of things happening. First, as models get better, we ask them to do harder things. A model that can answer a question is useful. A model that can take on an entire workflow is much more useful, but getting there requires a lot more expert input.

Second, humans keep creating new work. If a lawyer can use AI for a lot of what takes up their day, they have more time to take on more nuanced problems or develop a new service for clients. Those net-new functions then become things we want to distill into models as well. That’s what we mean by “no last mile.” The work keeps evolving.

Synthetic data makes this even more interesting. You can use it to expand on expert-created tasks and feedback, which makes that original human input more useful. I think that combination is why demand for human expertise keeps growing alongside model capability.

micro1 works with subject-matter experts across fields such as law, healthcare, finance, engineering, and science. What can these professionals contribute to AI systems that cannot simply be learned from existing online data?

A lot of what makes someone a great doctor, lawyer, or engineer is the judgment they’ve built by doing the work. Some of that is documented, but a lot of it shows up in how they handle an unusual case, recognize missing information, or decide which detail actually matters.

That’s extremely valuable for training models. An expert can create a difficult task, explain what a good result should look like, and identify exactly where a model falls short. An answer can sound very convincing and still miss the thing that matters most to someone who does that job.

We need experts to make those distinctions. And we need multiple perspectives and a strong review process, because one person’s judgment won’t capture every case. The goal is to turn that experience into a signal the model can learn from. There’s also an enormous amount of nuance within each field. Legal expertise, for example, isn’t one thing: it varies by practice area, by state or jurisdiction, and by the legal systems of different countries. Each layer requires more specialized knowledge and judgment. That same depth exists across every domain and its many subdomains, which is why we need experts who understand what those distinctions mean in practice.

As the industry shifts more attention toward post-training, reinforcement learning, and evaluation, where do you see the biggest bottlenecks in improving frontier models today?

I think access to realistic data is one of the biggest bottlenecks. Companies have spent years solving problems, making decisions, and handling exceptions, and that experience is captured across their documents, systems, and workflows. It’s incredibly valuable if you want a model to learn how to do real work.

That’s why we built our Company Data Partnerships program, or CDP. We pay companies to license their operational data. That can include internal documentation, project histories, CRM records, and the processes people follow to get work done. For valuable company datasets, we’re paying upwards of $2 million, with the amount depending on the data and the partnership.

We use that material to build realistic training and evaluation environments. Think about resolving a customer issue: you might need to look at the account history, find the relevant policy, and work out what action to take. An environment built around that workflow gives a model a place to practice those steps. Experts define the task and how to assess whether it was completed successfully. That feedback is what makes the environment useful for reinforcement learning.

Companies retain ownership of their underlying data. We agree on the scope and handling requirements, including privacy protections, and prepare the data for the intended use.

The value is in the experience those records capture. CDP gives companies a way to earn revenue from that experience while helping models learn to operate in much more realistic settings.

Which forms of human judgment or expertise do you believe will remain especially difficult for AI systems to replicate, even as models become significantly more capable?

Human taste and creativity are areas I think models will have a particularly hard time with. The reason comes back to how we create training data.

For a lot of tasks, we can bring experts together and get them to agree on what good looks like. That gives us a useful signal to train against. But if you ask people what makes a great image, a great film, or a compelling design, the more they discuss it, the more their views can diverge. There isn’t always one answer to converge on.

AI can still become very useful in those areas. But deciding what you want to create, who it’s for, and whether it actually resonates involves a lot of human judgment. I think that judgment becomes even more important as the tools make it easier to produce things.

You’re seeing professionals increasingly treat AI training and evaluation as serious professional work rather than a side gig. Could this develop into an entirely new category of employment, and what might those careers look like?

Absolutely. This is already a new multi-billion dollar category of work. We have experts applying the knowledge they’ve built over their careers to help train and evaluate AI. For some people, that’s becoming a substantial part of their working life.

The work itself can be very deep. A lawyer might create a difficult scenario and define how to assess the model’s response. An engineer might build a task that takes an agent through an entire software workflow. Then you have people reviewing the data, improving the evaluation criteria, and leading teams of experts. The work can also give people more room to focus on what drew them to their field in the first place. In most jobs, a lot of time goes toward administrative tasks and obligations that sit outside the work people find most meaningful. Training AI in their domain can put their expertise at the center of what they do: working through difficult problems, applying their judgment, and sharing what experience has taught them. It’s an opportunity to spend more time on the things they care about and do best, while helping shape how AI approaches that work.

I think we’ll see real career paths develop around all of that. You can become very good at translating your domain expertise into training data, or at running a data pipeline and understanding where models need help. Those are valuable skills, and we’re still very early in figuring out what all of these roles will look like.

As companies recruit doctors, lawyers, engineers, scientists, and other specialists to improve AI models, how do you verify expertise and ensure that the resulting training data is accurate and consistent?

We’ve built a whole system around this. It starts with Zara, our AI recruiter, which helps us source and vet experts for the skills a project requires. Then our Merit tool helps us track the quality, speed, and reliability of their contributions over time. You need to understand both someone’s expertise and how that translates into the data they produce.

Our flow platform brings that expert work together with the models we’ve built to support data generation and review. Experts define the tasks and what a good result looks like. Then flow gen helps expand that starting point into new examples and variations. flow qc checks the data for issues like missing evidence or a mismatch between a task and its scoring criteria. And flow grader evaluates how the model performs against those criteria.

The important part is what happens when an expert catches a mistake. That correction can improve the training data, but it can also help us improve the generators and quality checks themselves. We can turn recurring mistakes into tests for the next version.

That’s the synthetic data flywheel we’re building. Human expertise gives us the starting point and the judgment. Our tools help scale it, and expert feedback helps improve the next cycle. We then test whether those changes actually produce better model performance. That’s what makes the loop valuable.

AI agents increasingly need to perform complex, multi-step tasks rather than simply generate responses. How does training and evaluating these systems differ from the work required to improve traditional large language models?

The big difference is the length of the task. An agent might need to take many actions across different tools before it gets to a useful result. It can do the first few steps well and still fail the overall job because errors build on each other.

So we need environments where the model can practice the whole workflow, and evaluations that tell us where it breaks. Did it find the right information? Did it use the tool correctly? Did it actually get to the outcome we wanted?

Then there’s the context of the company itself. An agent working inside one business needs to understand that business’s processes and expectations. General capability only gets you so far.

That’s where domain experts become really important. They help us diagnose the failures and create the data to improve them. You evaluate, find the gaps, train against those gaps, and evaluate again. That process continues as the agent takes on more work.

As frontier models become more capable, how is the type of human-generated data they need changing, and which forms of expertise or evaluation do you expect to become significantly more valuable over the next few years?

Two things matter more and more: how long the task is, and how closely it reflects the real world. As models get better, the useful training data moves toward complex workflows with more decisions, more context, and more opportunities to make a mistake.

That makes operational expertise very valuable. Someone who’s spent years doing a job understands the exceptions and dependencies that are easy to miss when you’re designing a task from the outside. They can also tell you whether an agent’s result would actually work inside the business.

I think enterprise evaluations become an enormous market over time, potentially a trillion dollars a year. Every company deploying agents will need to understand where they work, where they fail, and how to improve them. The people who can turn their knowledge into useful evaluations will be central to that. It’s a continuing need because the companies, the models, and the work all keep changing.

Some professionals may effectively be helping train AI systems to perform portions of their own jobs. How do you think about that tension, and what do you believe the relationship between human expertise and AI will look like over the next three to five years?

I think we underestimate how much new work people will create when they have these tools.

Take recruiting. If an agent can handle more of the initial screening, a recruiter can spend more time building relationships with candidates or figuring out how to attract people the company couldn’t reach before. Their role changes, and they can do more with the same amount of time.

I expect something similar across a lot of professions. Over the next three to five years, agents will increasingly become coworkers that people direct, give context to, and evaluate. People will take on new responsibilities as that happens.

That’s why I’m optimistic. We’re already seeing new work emerge around training and evaluating AI, and I think there’s a lot more to come as people discover what they can build and accomplish with it.

Thank you for the great interview, readers who wish to learn more about expert-led training data, AI evaluation, and opportunities to contribute their domain knowledge should visit micro1.