Roll out AI across a portfolio one process at a time.

Usually not. A single programme sent from the fund tends to arrive as a tool and a training deck, and the people whose work is meant to change are asked last. Adoption stalls and the tool takes the blame.

Split the work instead. The fund owns what has to be common: the usage rules, the inventory, the measure and a way for companies to compare notes. Each company's management owns the process it picks, the build and the result, because they run the business and will still be running it after the fund has moved on.

Start with an inventory and a short usage policy. The NIST AI Risk Management Framework (AI RMF 1.0), published in January 2023, is voluntary guidance built on four functions: Govern, Map, Measure and Manage. Guides to it describe adoption starting with an inventory of AI systems and a governance structure, then moving to risk mapping, measurement and monitoring.

That order suits a portfolio. A fund-level risk register template can track each system's owner, intended use, risk classification and open mitigation items, and every company fills in its own. The framework's profiles allow stricter testing for high-impact systems and lighter review for low-risk internal tools, so a customer-facing model and an internal drafting aid need not meet one standard.

The policy each company's staff will read should fit on one page: what may go in, what may not, who decides the edge cases. Our usage policy guide sets out that page for a fund, and the same shape works for a portfolio company.

Choose one process with enough volume, where the work is mostly reading, drafting or reconciling, and where the owner is reachable. Leave the broadest ambition in the value creation plan for later. A narrow process can be measured; a theme cannot.

Licence counts tell you little. The number that matters is how many people in each company use the tools well, on their own work, read by team each month. It is usually much lower than licence utilisation, and it is the figure that moves when the work is done.

For a single process, measure time and error rate on the company's own volume, before and after, and write down how it was counted. Set no fixed date for results and promise no saving in advance. A figure counted the same way at every company lets you compare them fairly, and it is the figure the next management team will ask to see.

Our guide to AI adoption in private equity explains why utilisation is the wrong number inside a fund, and the argument holds for a portfolio company.

It helps. An operating partner who has watched one process rebuilt inside the fund knows what discovery, design and measurement take, and portfolio management teams hear that differently from a slide. The fund's own documents make good practice: the quarterly portfolio report, the investor inbox, the roll forward.

That is where our work sits. By The People works inside the fund, with deal, operations, finance and IR teams, and does not run portfolio-company programmes. At one private equity firm, investor-email triage went from days to under five minutes. What transfers to a portfolio company is the method: one process, its own people building it, a person in the loop, measured before and after.