James Milin has a tidy way of describing a messy corporate ritual. A company buys artificial intelligence tools. Employees try them. Pilots multiply. Usage climbs. Then somebody in a conference room asks the question that turns the excitement into an accounting problem: what did any of it produce?
That question is the center of Milin's work at Workhelix, the company he co-founded and runs from San Francisco. Its product helps large organizations find where generative AI might improve work, put it into those places, and measure whether the improvement actually occurred. The sequence is deliberately plain: Decide, Deploy, Assess. It sounds like a checklist because the alternative, in Milin's telling, is often a pile of activity wearing a progress badge.
His timing is good. Enterprise leaders have been asked to develop an AI strategy while the technology, vendors, and vocabulary change underneath them. A broad instruction such as “use AI in finance” may satisfy a board slide, but it says little about which person should use which tool for which task. Workhelix shrinks the problem until it can be inspected. A job becomes a bundle of tasks. Each task can be scored for its suitability for AI. The best opportunities rise to the top. Results get compared with a credible picture of what would have happened without the intervention.
“We don't tell leaders they're wrong - we help them ask the right questions.”James Milin
A career in translation
Milin did not arrive at this work through a laboratory. His public career starts with social-work training and a role at ICF International providing training and technical assistance connected to the U.S. Department of Justice. He later moved into enterprise technology at SHI International, where selling infrastructure meant learning how organizations make expensive, slow, politically layered decisions.
From there came Amazon Web Services. Milin was a founding member of its private-equity sales team, working where cloud architecture met an investment thesis. He then co-founded Reserved.ai, a venture-backed platform that used machine learning and automation to reduce cloud costs. Google followed, with a role in enterprise field sales. Across those stops, he says he helped lead cloud migrations and digital transformations for hundreds of mid-market and enterprise customers.
The resume can look like a hard turn from social services into software sales. Look at the work itself and the continuity appears. Each job placed Milin between a complicated system and the people expected to navigate it. Justice programs have policy, training, incentives, and human consequences. Cloud migrations have architecture, contracts, budgets, and teams. Enterprise AI has all of those at once. Translation is not a soft prelude to implementation. It is implementation.
Putting economists in the product
Workhelix began when Milin brought together three researchers who had spent years studying technology and work: Stanford economist Erik Brynjolfsson, MIT research scientist Andrew McAfee, and Wharton professor Daniel Rock. Their research supplied a way to reason about productivity, tasks, and causal evidence. Milin supplied the operator's knowledge of how a large customer buys, adopts, and judges technology.
The pairing matters. Academic work can reveal a pattern without turning it into a corporate workflow. Enterprise software can package a workflow without proving that the workflow measures anything important. Workhelix tries to join the two. Its platform draws on hundreds of millions of public workplace data points, covering thousands of jobs and hundreds of thousands of activities. Data scientists and strategists work with customers alongside the software.
That human layer makes the business less frictionless than a pure self-serve subscription. It also reflects Milin's view of the problem. An organization cannot learn where AI belongs by switching on a generic dashboard. Jobs vary by company. Workflows carry history. The same tool can help an experienced employee in one context and add review work in another. Measurement design needs judgment before it generates a number.
Milin's preferred unit is the task because whole jobs are too coarse. “If you look at all the jobs in an organization and break them down into bundles of tasks,” he has said, “and then score each task for its suitability to be accelerated by generative AI, now you can come up with a really quantitative rigorous way to adopt it.” The sentence is long because the shortcut is the problem.
Find the work
Break roles into tasks and rank the opportunities where AI could create value.
Change the flow
Put tools into specific work with a clear outcome and an accountable owner.
Prove the result
Compare what happened with a credible counterfactual, then adjust or scale.
The counterfactual in the room
Most adoption dashboards answer a simple question: did people use the thing? Milin wants the next question. Did the tool cause the result? A team might ship more software after introducing an AI coding assistant. It might also have hired senior engineers, narrowed the roadmap, or recovered from an unusually slow quarter. Usage and output can rise together without one fully explaining the other.
This is where Workhelix borrows from the “credibility revolution” in economics. The aim is to construct a reasonable counterfactual, a view of what likely would have occurred without the intervention. It is a more demanding standard than a testimonial or a before-and-after chart. It can also save an executive from scaling the wrong project with great confidence.
“We don't assume AI is delivering value just because it's being used. We measure it - causally, rigorously.”James Milin
The tone of Milin's public comments is notable. He does not pitch measurement as a courtroom designed to embarrass an enthusiastic executive. He frames it as a better set of questions. How do you measure success today? Where do you believe the value is appearing? What evidence would make you confident enough to invest more? The questions let the gap announce itself.
Define the outcome and the counterfactual before the pilot begins. If the team cannot say what success changes, it is not ready to measure success.
The first dozen arrived quietly
Workhelix launched its product publicly in April 2024. Its first dozen enterprise customers arrived without paid advertising. Named customers have included Wayfair, Coursera, and Accenture. That early pull gave the company a useful signal: executives did not need another speech about AI's potential. They needed a map for acting on it.
In February 2025, Workhelix announced a $15 million Series A led by AIX Ventures. The investor list mixed venture firms with people who have built or studied major technology systems, including Reid Hoffman, Mira Murati, Jeff Dean, Yann LeCun, Sebastian Thrun, and Andrew Ng's AI Fund. Accenture invested, brought Workhelix into its Project Spotlight accelerator, and said the company's capabilities would be integrated with LearnVantage, its learning and training platform.
The capital was earmarked for expanding the tasks and performance indicators the software tracks, plus the internal tools used by data scientists who work with customers. That plan reveals the actual product ambition. Workhelix is not merely producing a one-time catalog of AI-friendly tasks. It wants to become the continuing record of where a company placed its AI bets and what those bets returned.
After the novelty
Milin's aspiration extends beyond cleaner budget meetings. He talks about AI as a change in how work gets done, who does it, and what good performance looks like. Those are organizational questions before they are software questions. They determine whether a tool removes drudgery, creates another review queue, helps a newer employee learn, or quietly shifts responsibility to somebody who never appeared in the project plan.
His own route supplies a useful check on the standard technology-founder story. The job does not always begin with writing code. Sometimes it begins by watching institutions struggle to absorb a new capability. Milin learned the customer side of transformation across government programs, resellers, cloud providers, and a cost-management startup. At Workhelix, he turned that accumulated view into the connective tissue between researchers, engineers, data scientists, and executives.
There is an appealing modesty in the company's central move. Break the grand transformation into smaller pieces. Name the task. Choose the tool. Decide what should change. Watch carefully. Keep what works. The method does not make AI less consequential. It makes the consequences visible.
The AI boom has produced plenty of forecasts about the future of work. Milin is building something closer to a ledger. A ledger cannot tell a company what to value. It can show where the value appeared, where it failed to appear, and which story survives contact with the numbers. Once every company can buy roughly the same models, that habit of seeing clearly may be the part that compounds.