Training Portfolio Company Teams on AI

1. The New Job: AI Into Companies With No Tech Bench

A new line has shown up in the operating partner's job description: get AI into the portfolio. LPs ask about it, the investment committee expects it, and the value-creation plan now has a slide for it.

The trouble is the target. A portfolio company is not a fund. It is often a mid-market business with no technology bench, a management team already stretched thin, and no one whose job is to think about AI. The CFO is doing three roles. The head of operations has no time to evaluate tools. The instinct to hand them a license and a deck produces exactly what it produces at the fund, which is nothing, except now it happens across ten companies at once.

So the operating partner's real task is not to deploy a tool. It is to drive adoption inside a business that has no capacity to drive it itself. That is a harder job than the fund version, and it rewards a different playbook.

2. Why the Fund's Playbook Does Not Transfer

If the fund has done its own AI rollout well, the temptation is to copy it into the portfolio. The discipline transfers. The specifics do not, and assuming they do is how a good fund program fails in a portfolio company.

Four things are different. The people. A fund is full of analysts who pick up tools fast. A portfolio company's staff may have never used anything like this, so the training starts further back. The data. The fund's workflows are CIMs and memos; a manufacturer's or a distributor's are invoices, work orders, and customer records, a completely different set of jobs. The appetite. Fund staff have a career reason to learn AI. A long-tenured plant manager may have none, and may reasonably wonder if this is about replacing them. The stakes. At the fund, a bad output costs a redo. In an operating company, a workflow that breaks can stop shipments or mis-bill a customer.

What carries across is the method, not the content: one owner, one workflow won deeply, trust earned on real work, adoption measured honestly. The underlying discipline is the same one in why AI rollouts fail. The execution has to be rebuilt for each company.

3. Triage the Portfolio by Readiness

The first mistake is treating the portfolio as one thing and pushing the same program everywhere at once. Companies differ enormously in how ready they are, and where you spend your scarce time should follow that.

Triage is what keeps the effort honest. Pushing AI into a company in the middle of a turnaround does not just fail there, it spends credibility you needed for the companies that were ready. Start where it can work, and let the results do the persuading.

4. Start With One Function, One Company

Inside a ready company, the same discipline applies one level down. Do not roll AI out across the whole business. Pick one function and win it completely.

The best first function has the traits any good beachhead has. It costs real hours, so the saving is felt. The output is checkable, so trust is earned by verification, not faith. The data is allowed and reasonably clean. And it has a manager who wants it. In a lot of mid-market companies that is the finance function, drowning in document work, or customer service, or a back office buried in routine correspondence and data entry. Win that one function, in that one company, and you have something no deck can give you: a real result the rest of the portfolio can see and want.

Narrow and deep beats broad and shallow here even more than at the fund, because a portfolio company has less slack to absorb a scattered effort. The wider question of which job to pick first, and why the choice matters so much, is in where AI creates the most value.

5. The Management-Team Buy-In Problem

The single biggest difference from a fund rollout is the people you are asking to change, and how they hear the request. A long-tenured management team did not sign up to be an AI pilot, and an operating partner arriving with a tool can land as the fund second-guessing how they run their business.

So buy-in has to be won before the tool, not assumed after it. That means framing it as removing the work nobody wants, not as a verdict on the team. It means starting with a manager who is curious rather than the one who is loudest in resisting. It means proving it on their terms, with a number they care about, on a workflow they chose. And it means being honest about the fear underneath, which is usually about jobs. The credible message is that AI takes the assembly work off capable people so they can do more of what they were hired for, and the way you make that message believable is to deliver a win that helps the team rather than threatens it.

Force it from above and you get malicious compliance: the tool gets logged into once and quietly abandoned the moment you look away. The behavioral side of this, getting smart and skeptical people to genuinely change how they work, is the whole craft, and it is the same lens in using AI for revenue growth in portfolio companies.

6. What to Centralize vs Leave to the Company

A standing question across the portfolio is how much the fund should run centrally and how much to leave to each company. Get the split wrong in either direction and you waste effort or you lose adoption.

Centralize what travels. Vendor selection and pricing, the data and governance standards, a vetted shortlist of tools, the playbook that worked at the first company, and the hard-won lessons. There is no reason for ten companies to each negotiate the same contract or each learn the same mistake. Leave to the company what is local. The choice of which function to start with, the people who run it day to day, and the fit to their specific workflows. Adoption lives inside the company, with their managers and their work, and it cannot be run from the fund's office. The fund supplies the standards and the playbook. The company supplies the ownership.

The failure mode at both ends is familiar. Centralize too much and you get a fund-mandated tool nobody at the company feels they chose, which means nobody adopts it. Centralize too little and every company reinvents the wheel, slowly and expensively. The split is the art, and it is most of what a portfolio operating model for AI is.

7. Measuring It Where It Lands: EBITDA

At the fund, adoption is measured in hours saved and workflows changed. In a portfolio company, the operating partner has a sharper and more honest yardstick available: the financials.

AI in an operating business should show up where value creation always shows up. Lower cost to serve as a function does more with the same headcount. Faster cycle times that free up working capital. Revenue that grows because the sales team spends less time on admin and more time selling. These are EBITDA effects, and they are the language the investment committee and the eventual buyer already speak. Activity metrics from the fund world, logins and messages, mean little here. The question is whether the company runs better and earns more, and that is measurable on the P&L within a quarter or two of a real win.

Tying the effort to EBITDA from the start also keeps it disciplined. It forces you to start with functions where the saving is real and visible, and it gives you the number that makes the next company say yes. The broader map of where these effects come from across the portfolio is in where AI creates the most value.

8. The Repeatable Model Across the Portfolio

The point of all this discipline is not one company. It is a model you can run again and again, getting faster each time, until pushing AI into a new portfolio company is a known process rather than a fresh experiment.

After the first ready company, you have the assets that make the second one easier: a playbook for the starting function, a vetted tool and a negotiated price, the data and governance standards, and a real before-and-after number to show the next management team. The second company is faster than the first, the third faster than the second. What started as a bespoke project becomes part of the value-creation toolkit, applied at every new platform and bolt-on, the same way operational best practices already are.

The maturity of this across a whole portfolio is worth assessing deliberately rather than by gut, which is what a portfolio AI maturity assessment is for, and the full company-level execution is in the playbook for deploying AI in portfolio companies.

9. The Exit Story It Builds

Done well, this is not only an operating improvement. It is a story you tell at exit, and a credible one, because it is backed by results rather than by a slide.

A buyer pays for a business that is more efficient and harder to compete with. A company where AI is genuinely embedded in how core functions run, with the EBITDA to prove it, is worth more and easier to diligence than one that bought some licenses and hoped. The contrast matters: a real adoption story, with numbers, reads completely differently from a checkbox in the data room. And the model itself becomes part of the fund's own story to LPs, evidence that the firm does not just talk about AI in the portfolio but has a repeatable way of putting it to work and measuring the result.

That is the prize that justifies the unglamorous work. Triage, one function, buy-in, EBITDA: it is slow, and it builds an asset a buyer can see and a hand-wave cannot fake.

10. Where to Start

Triage the portfolio this quarter. Sort the companies into ready, willing but not ready, skeptical, and not yet, and pick the one ready company where a win will be most visible. Resist the urge to start everywhere.

Then, inside that company, pick one function with real hours and a willing manager, win the buy-in before the tool, and measure the result in EBITDA. One real win, on the P&L, is the asset that makes the next company say yes and the whole model repeatable.

If you want a partner who runs this across the portfolio, the company-by-company adoption and the central standards both, that is what an AI Operating Partner does. Where a company needs a workflow built rather than just trained, we build it as a Custom Build, and a Guided Launch trains a company's team one function at a time.