Mener un Pilote IA Pendant la Période de Détention : Guide pour Operating Partners en Private Equity

An AI Pilot Inside a Hold Period Has One Job: Produce a Written Answer Fast

A private equity sponsor running an AI pilot in a portfolio company is not running the same experiment a corporate innovation team runs. The corporate team has a budget line and a strategy horizon. The sponsor has a hold period, an exit multiple, and a portfolio in which the same question is being asked ten times at once. That changes what a pilot is for. It is not there to prove that AI works in general. It is there to produce, per company, a written answer to one question, is there anything here worth building, in a timeframe short enough that the answer can still change what happens before exit.

This article is a playbook for that. It covers why the pilots that circulate in this market stall, what a pilot inside a hold period should produce, how to sequence it across a portfolio, and what goes in the board pack. It is written for operating partners, and for the AI Operating Partners that firms are now hiring, a role that Korn Ferry Institute described in April 2026 as emerging because conventional tech-focused operating leaders were judged to lack the depth to get value from AI.

Why the Standard Pilot Stalls in a Portfolio Company

The standard pilot in this market is a proof of concept: a vendor takes a sample of the portfolio company's data, builds a demo in two to four weeks, and shows it working. Korn Ferry's piece on the AI Operating Partner role records that mid-market firms now expect exactly this. The problem is what a PoC measures. It measures whether the model can do the task on the inputs the vendor was given, which are, invariably, the clean ones. It does not measure whether the portfolio company's actual process, forty email formats, half the payload in PDFs, a routing rule that lives in the founder's head and leaves when the earn-out ends, can carry the system in production.

That is why the pilots stall, and they stall at every scale. BCG's July 2026 survey of 152 CEOs at companies with at least $500M in revenue found two-thirds pursuing AI pilots and only 26% with AI embedded in a broader transformation. Those are companies with the budget to hire their way out and the headcount to build anything. If they are stuck, it is not for lack of resources. It is because nobody wrote down how the company actually works before trying to change it. A portfolio company with a founder-led back office and no IT department has the same gap, with fewer people to paper over it. We wrote about the founder-led case specifically in mapping an acquired company with no IT department.

The hold period turns this from a nuisance into a constraint. A pilot that stalls in month four of a corporate programme is a line in a retrospective. A pilot that stalls in month four of a three-year hold has consumed a tenth of the time in which the thesis could have been proven, and the next portfolio company is still waiting.

What a Pilot Inside a Hold Period Should Actually Produce

Reframe the pilot as an instrument rather than an experiment, and the deliverables change. The instrument has to produce the same output in every portfolio company, so the findings are comparable across the fund even when the answers differ, and it has to produce that output fast enough to matter. In practice that means five documents, in order.

• The Exception Ledger. Every real deviation from the documented process, how often it happens, who absorbs it today. In a founder-led target this is where the routing logic that leaves with the founder gets written down for the first time. It is also the document a diligence team wishes it had had before the deal.
• The AI Boundary Map. Each step classified as deterministic code, model judgement, or human approval. The headline is the determinism ratio, and it is usually lower than the investment memo assumed. In one European manufacturer's order handling that SUPALABS mapped, three of eleven steps genuinely needed a model (SUPALABS engagement data, 2024–2026). The other eight were ordinary software, which is cheaper to run and easier to show to an acquirer.
• The evaluation suite. A golden dataset from the company's own cases, per-step pass rates, and a monthly accuracy report. This is the number that goes in the board pack, and it should be quoted as its own line item so it survives the build running late.
• The decision log. Every automated action, its inputs, its confidence, and who approved it, in a surface an acquirer's diligence team can open without asking an engineer. Exit diligence on an AI-enabled process is a question of whether the decisions can be reconstructed; this is the answer.
• The ERP-additive covenant. A written commitment that the system is built on top of what the portfolio company runs and nothing gets replaced. A replatforming is the one AI initiative guaranteed not to pay back before exit.

The first two are produced by a five-day mapping sprint with the people who run the process, and the sprint ends in either a fixed-price build quote or a written no. The written no is a deliverable. Five days spent discovering that a portfolio company's thesis is a pricing rule and a spreadsheet, not an AI problem, is a hold period saved. These are the same five documents described on the method page; the point for a sponsor is that they are the same in every company.

Sequencing Across a Portfolio

The temptation with ten portfolio companies is to run ten pilots in parallel. Resist it, for a reason that is about depth rather than capacity. The Exception Ledger is produced by five days with the people who run the process, and that depth does not survive being spread across three companies in the same week. What should repeat is the instrument, not the calendar. Run the sprints in sequence, one company at a time, and accept that the fourth company's answer arrives in month two rather than week two. The alternative is ten shallow answers that all say yes, because a shallow look at any process finds something to automate.

Sequence by two criteria. First, where does one workflow visibly hurt and have a document trail already: supplier invoices re-keyed by three people, quote-to-cash with five handoffs and no owner, an order desk that runs on one person's memory. Second, where is there an executive at the portfolio company, not at the sponsor, who owns the outcome and can clear an approval boundary. A sponsor mandate is not the same thing. Without a named owner inside the company, the work stalls at the first approval regardless of who builds it, and that is true of every pilot in every company.

After the first sprint returns a yes, run the build in that company while the second sprint runs in the next. Six weeks from the start of the build to the first workflow in production, with shadow mode before it acts, is the cadence. By the time the third sprint reports, the first company has a monthly accuracy report, and the board pack has a number.

What Goes in the Board Pack

Three things, and none of them is a status deck. The determinism ratio per company, because it tells the board how much of the thesis was AI and therefore what it costs to run. The monthly accuracy report per live workflow, because it tells the board whether the system is still right and whether a task has earned more authority. And the count of written noes, because a sponsor who can show five companies mapped, three built and two declined on evidence is describing a process, not a hope.

The decision log is not for the board. It is for the buyer at exit. A process that has been automated for eighteen months and can produce, on request, every decision it took, with inputs, confidence and approver, is a process an acquirer's diligence team can sign off on in a week. One that cannot is a discount.

Before, or Instead of, Hiring the AI Operating Partner

Korn Ferry's piece names three backgrounds for the AI Operating Partner role, entrepreneurial, technical-product, and executive technology, and lists the risks of creating it: overlap with the existing tech operating partner, added complexity for portfolio leadership, and pressure to adopt AI where the business case is not there. The sprints answer all three. They tell you whether there is a thesis at all, before anyone is on the payroll to be right about it. They tell you what shape the role should be, because what the sprints found across the portfolio is a better guide than what the market is hiring this quarter. And they tell you what the hire will inherit: an operating partner who arrives to five Exception Ledgers, five Boundary Maps and one system in production has a job, while one who arrives to eighteen months of seat licences has a discovery project. The fuller argument is on the private equity page.

When Not to Run This

Three situations where a sprint in a portfolio company is the wrong purchase, and it is cheaper to establish them on a thirty-minute call than in week two. The company is large enough to staff its own delivery practice, in which case hire it; embedded delivery is how you learn what to hire for, not a substitute for it. The thesis is arithmetic rather than judgement, consolidating reporting across three legacy systems or applying a pricing rule consistently, in which case a rules engine wins and costs less to keep. Or nobody inside the company owns the outcome, in which case the work stalls at the first approval boundary, whoever builds it.

Sources & References