How Heads of Acquisitions Move AI Pilots to Production

A pilot earns attention because a model reads an offering memorandum, drafts a screen, or abstracts a rent roll in minutes. In a room, that looks like production. On a live deal, the same output reaches an analyst who has to defend it to a principal, and the questions begin: which lease says that, which page, which assumption is the model’s and which is the seller’s. A pilot that cannot answer those questions adds a review step instead of removing one.

The reason most acquisitions pilots never ship is not model quality. It is the absence of an operating wrapper. A demo has no input policy, no output standard, no source trail, no named owner, and no place to run other than a chat window. Production work in acquisitions has all five, because a mispriced assumption is expensive and a committee will ask where every number came from.

The move to production begins by narrowing the mandate. Pick a task that is frequent, bounded, and easy for an experienced acquisitions professional to grade: a first-pass underwriting from an OM, rent roll, and T-12; a cited initial screen against the buy box; or a source trace that ties every figure in an investment-committee memo back to the file. Running a whole transaction is not a first workflow.

A narrow job makes its inputs, outputs, and failure modes visible, and it gives the team a fair test. Give the pilot real but permitted deal material, define the output before it runs, and require that every material conclusion be checked against the original documents. That comparison is what separates a production workflow from prompt theater.

A pilot becomes production when it carries the same discipline as the rest of the acquisitions process. Name the permitted inputs and where data may travel. Fix the output as a template the committee already reads. Keep the arithmetic deterministic, so the model normalizes and drafts while the calculations run in code that returns the same answer every time. Preserve a source trail from each material line back to the document, page, or cell.

Then give the workflow an owner and an escalation rule. Someone maintains the inputs, the instructions, the output template, and the review checklist, and decides what happens when the system is wrong. Written down, that wrapper is a reusable operating procedure. The next deal improves a known system instead of starting from a blank window, which is how one working pilot compounds into a program.

Production does not mean autonomous. The approval gate sits closest to the decisions with capital, legal, or relationship impact: the price, the go or no-go, and the assumptions a lender or investor will rely on. The system prepares the packet and shows its work. A named reviewer signs the judgment. That boundary is what lets a firm defend an AI-assisted underwriting in the same room where it defends its own analysts.

Before widening the mandate, measure the workflow against the process it replaces: cycle time, rework, and how often the second reviewer changes a material number. A pilot that shortens the path to a defensible decision earns the next workflow. One that only shortens the demo does not. Scaling an unmeasured pilot is how a firm accumulates tools no one trusts.