Why AI projects stall in private-market firms
The demo works. The rollout does not. The reason is almost never model capability — it is that the firm has three versions of the pipeline, two definitions of committed capital and a set of numbers that only reconcile because one person reconciles them manually each month.
An AI layer removes that person's judgment from the loop while inheriting the ambiguity they were quietly absorbing.
Naming systems of record
For each critical entity, write down the single system that holds the authoritative version and the single person accountable for it. Everything else is a copy and is labelled as one.
- Investors and contacts — one CRM, one owner.
- Deals and pipeline — one system with defined stages.
- Underwriting models — one repository, one naming convention, one current version per deal.
- Documents and agreements — one store with an executed-versions rule.
- Financial and fund accounting — one ledger, reconciled on a stated schedule.
- Reporting outputs — one archive of what was actually sent, and when.
Definitions before dashboards
Write a short definitions register covering the terms your decisions depend on: what counts as a qualified opportunity, when a commitment becomes committed, how occupancy or utilisation is calculated, what date a deal is considered dead, how pipeline value is measured.
This is not bureaucracy. It is the difference between a metric that changes behaviour and a metric people argue about in the meeting where it is presented.
Cadence makes data survive
Data quality is a function of use. Fields that feed a decision made on a schedule stay current; fields that feed nothing decay within a quarter.
So tie each source of truth to a recurring decision: a weekly pipeline review, a monthly portfolio review, a quarterly investor report. Then delete the fields that no decision consumes.
Governance rules worth setting first
Decide, before any model touches your data, what may never be sent to an external system: personal identifiers, investor account information, executed agreements containing counterparty confidential terms, and anything covered by a confidentiality obligation you do not control.
Set access by role, log it, and set retention periods. These rules are far easier to establish before a tool is embedded in daily work than after.
What becomes possible afterwards
With records, definitions, cadence and governance in place, the useful applications are unremarkable and immediately valuable: drafting first-pass screening summaries from a defined intake format, assembling reporting packs from an authoritative ledger, surfacing pipeline records that have gone stale against a stated rule, and answering internal questions from documents the firm actually controls.
None of that requires ambition. It requires the groundwork most firms skip.