Michael Higgins
All work

UnitedHealth Group / 2026

From asking what AI is to building with it: AI transformation for an M&A team

Five months to stand up an AI program for Corporate Development: the plan, the governed platform, and working agents, with the deal team building its own tools by the end.

Working draft, content pending

In June 2026 I moved from leading a product team to leading AI for Corporate Development, the function that runs mergers and acquisitions. There was no program, no approved place to host a prototype, and a team of deal professionals who were busy and reasonably skeptical. I took on three roles at once: program leader, solution architect and hands-on builder.

My role

I owned the program end to end and built much of it myself. That combination is the point of the role. Leading the program told me what mattered; building the tools told me what was possible, what it cost and where it broke.

Start with the work, not the technology

I could not lead AI for a function I did not understand, so I went to the work. I completed buy-side training, shadowed deal teams, and mapped the deal process task by task. What we saw became a prioritized list of automation opportunities. It kept us building for real pain points, not assumed ones.

Choose on evidence

Before buying anything we ran structured pilots with real testers and real use cases, debriefed each tester, and wrote up what we found. Some tools earned a place and some did not. One evaluation ended with a recommendation not to build at all, because the gap was judgment, not technology. Pilot data was closed out properly when each pilot ended.

Governance as a design input

In a regulated enterprise, review boards, data-use reviews and procurement all take time. I learned to treat them as inputs to the design, not a hurdle at the end. We built on approved infrastructure from day one, set plain rules for what data could go where, and stood up a governance board so every new skill is tested against past deals before it ships.

A governed place to build

The blocker for a year had been simple: nowhere to put a prototype. I proved a path from the company's AI studio to an authenticated, hosted app, then launched a hub on it where the team finds, tests and uses what we build. The pattern was picked up as the reference for the wider finance AI effort, and other teams now deploy on it.

Working agents

The tools are real. One reads a long deal document with eight agents working in parallel and returns a cited executive briefing, with every citation checked and a redaction gate in front of the output. A first-pass review that takes a day or more becomes something leadership reads in minutes. Others analyze enrollment data, check financial models, and rewrite a sprawling analyst prompt into one half the length that resolves most of its known issues.

Adoption is voluntary or it is nothing

A new tool means nothing if no one uses it. Rather than push from the top, we ran a champions program: four teams, each sponsored by a leader, each building something useful for its own work. Participation held week after week. We also chose to measure usage, not hours saved, because the goal is better jobs, not fewer of them.

What is not done

Some targets moved. When the agent count and a knowledge-graph demo needed more time, I reset them openly rather than show something that was not ready. Production hosting and engineering capacity are still open questions, and I report on them plainly.

What it taught me

Six months ago the team was asking what AI is. Today they build their own tools. The lesson I keep is that transformation is a product problem: understand the work, validate before you build, earn adoption, and be honest about what is not finished.

Contact

Let's build AI people actually use.

I'm interested in AI deployment, architecture, and product leadership roles where design judgment, enterprise scale, and human outcomes meet.