Agentic AI will reach the deal team before the finance function
Where I see AI agents working in private equity today, and the controls that have to come first.
Since 2024 I have been advising clients on agentic AI systems and building them myself, in TypeScript, Python and Rust. That work covers memory stores, security and agent-to-agent (A2A) frameworks. My view, after a lot of prototyping, is that agents will change how deals are done before they change how books are closed.
Why deals first
Deal work is research under time pressure. A due diligence team spends most of its first fortnight finding things: which contract says what, which number in the management accounts ties to which bank statement, which claim in the information memorandum has no source. Agents are well suited to that kind of work. Three jobs stand out:
- Indexing a data room and tracing every material claim back to a source document.
- Flagging gaps, so the team spends its time on judgement rather than on looking.
- Keeping a living issues list that updates as new documents arrive.
Where I would be careful
The same properties that make agents useful make them risky in a regulated setting:
- Permissions. An agent with write access to anything needs the same segregation of duties as a person.
- Memory. What a system remembers between tasks is a data-protection question, not only a technical one.
- Accountability. Every number in front of an investment committee still needs a named human owner.
I learned the value of designing controls in from the start when I founded a fintech in the FCA Regulatory Sandbox. Retrofitting audit trails and explainability is far more expensive than building them in.
What this means for sponsors
The firms that win will not be the ones with the most AI. They will be the ones whose deal teams use it to move faster with the same discipline. That is the combination I bring: investment judgement from the deal side and hands-on build experience from the technology side.