AI is ready — but are you?

The argument about AI in private markets has quietly changed sides. For a while the question was whether a model could be trusted anywhere near a capital account. That question has largely answered itself. Reading a subscription pack, pulling out the commitment and the fee terms, and flagging the clause that disagrees with the side letter is no longer a demonstration. It is a Tuesday.

What has not changed is the state of the average fund's data. That is now the binding constraint, and it is not a constraint anyone can buy their way past. An agent is a fast worker with no institutional memory and no ability to walk down the corridor and ask. It acts on what it can read, and it can be trusted only as far as somebody can check its work. Both of those are properties of your organisation, not of the model.

What "ready" actually means

Readiness is a specific and unglamorous condition. It has four parts, and none of them involve artificial intelligence.

1. One record rather than several

Most funds have a commitment schedule in more than one place: the accounting system, a spreadsheet the finance team actually uses, a version in the data room, and whatever the investor relations deck last quoted. Humans reconcile these silently, out of experience. An agent has no basis on which to prefer one. Until there is an authoritative record — and everyone knows which one it is — you are not automating a process, you are automating an argument.

2. Fields that mean the same thing everywhere

Ask three people in a fund whether "commitment" includes the GP's own commitment, whether "drawn" is net of recallable distributions, and which date a valuation belongs to. You will often get three defensible answers. A person notices the ambiguity and asks. An agent resolves it silently, and the output looks exactly as confident either way. Written definitions are not bureaucracy here; they are the interface.

3. Permissions that already express who may do what

Every fund has permissions in practice. The analyst does not issue the drawdown notice; the associate does not release a distribution. But those rules usually live in habit and seniority rather than in a system. An agent cannot inherit a habit. If the only way to give software access is to give it everything, then the honest description of your first pilot is that you handed a superuser key to a process nobody can supervise. Roles have to exist as settings, not conventions, before an agent can be given one.

4. A process with a defined approval step

The difference between an agent that proposes and an agent that executes is the whole question. Proposing is safe, reversible and immediately useful. Executing requires that somebody has decided, in advance and in writing, which class of action goes out without a human and which does not. Most funds have never had to make that decision explicitly, because the approval was always implicit in the fact that a person had to do the typing.

Why pilots stall

Pilots rarely fail because the model was disappointing. They stall for reasons that are visible in hindsight and boring in advance.

• The agent was pointed at four spreadsheets that disagree, and it inherited the disagreement. Nobody could say the output was wrong, and nobody could say it was right.
• Checking the answer required redoing the work. A pilot that has to be verified by hand has added a task rather than removed one.
• The output could not be traced to a source. "Where did this number come from?" had no answer, so the number could not be used in anything that leaves the building.
• The workflow chosen was the most impressive one rather than the most tractable one — usually something touching investors or money, which then needed an approval process that did not exist.
• The pilot had no owner whose job improved if it worked, so it survived on goodwill until the quarter got busy.

Notice that every one of those is an organisational fact. Swapping in a better model changes none of them.

The honest objection

There is a fair response to all of this, and it deserves to be stated properly rather than knocked down: if you wait for clean data, you will never start. Fund data is never finished. There is always a legacy vehicle on different terms, a side letter nobody has structured, a portfolio company that reports in prose. Treating readiness as a precondition is how firms spend three years on a data project and arrive with nothing an investor can see.

So do not sequence it that way. Readiness is not a gate you pass; it is a property you build one workflow at a time. Pick a single workflow that already satisfies four tests:

1. The data it needs already lives in one place, or can be made to.
2. The output is checkable in under a minute by someone who would know if it were wrong.
3. A mistake is recoverable — nothing irreversible, no money moved, no investor contacted without review.
4. It happens often enough that a difference is visible within a quarter.

Chasing missing onboarding documents qualifies. Drafting the first version of quarterly commentary from structured portfolio figures qualifies. Proposing matches between incoming wires and expected amounts, for a human to confirm, qualifies. Autonomously issuing a capital call does not, and should not, for a long time. Agents built for this domain are designed around that distinction rather than against it.

A readiness checklist

Answer these about your own firm, honestly, without checking with anyone.

1. Can you name the one system that holds the authoritative commitment and capital account record — and would your colleagues name the same one?
2. If two systems disagree on a NAV, is there a written rule about which one wins?
3. Are the definitions of commitment, drawn, distributed and unfunded written down anywhere other than in a colleague's head?
4. Do roles exist as configuration, so that "who may approve a distribution" is a setting rather than an understanding?
5. Is there a single audit log that records who changed what, and when — covering software as well as people?
6. Are your unstructured documents indexed and retrievable alongside the structured record, or filed by year in a shared drive?
7. For the first workflow you would automate, is there a named person whose job is to approve the output?

Count the noes. That is not a verdict on your firm; it is your project plan, in priority order. Most managers find the first three are the expensive ones and the last four are mostly decisions nobody has got around to making.

The argument, briefly

Model capability is now roughly evenly distributed. Everyone can buy access to something competent, which means capability has stopped being a source of advantage — the same way electricity stopped being one. What is not evenly distributed is the state of a firm's own records, permissions and processes. That is where the difference will sit for the next several years.

The pleasant part of this conclusion is that the preparation is not speculative. One record, agreed definitions, real roles, a defined approval step and an audit trail are worth having if AI never improves again. They are also, as it happens, exactly what an agent needs. You are not betting on a technology. You are tidying your own house, which is defensible in front of any investor.