Say one of your portfolio companies gets AI right this year. Not a stack of tool licenses – an honest change in how some expensive piece of work gets done, built with the company’s own team, visible in numbers a board can check. The CEO presents it at the quarterly. Everyone is pleased, the deck moves to the next item, and at most funds that is where the story ends.
I think that ending forfeits the difference between owning a company and owning a portfolio. A standalone company that gets AI right gets one win. A fund that gets one company right owns a playbook it can run again. The difference between those two sentences is not effort or spend; it is structure – board control, common governance, and eight or more companies that can each take an adapted version of a proven play. Private equity assembled that structure decades ago, for other reasons. It happens to be built for exactly this.
I am not the first to say so. BCG’s advice to private equity, in its “AI-first” series, is precisely this motion: develop playbooks that nearly every company in a portfolio can adopt, prove the play at one company, then run adapted versions across the rest.
The BCG piece carries no survey data, so take it as an argument rather than a statistic – but it is the same argument.
And the motion is more tractable at this end of the market, not less: translating a play across eight to fifteen companies is a project; translating it across eighty is a program office.
For a deal partner the same point reads even more simply. AI is a rare lever where one good decision, made once at fund level, can pay across every board you chair – provided someone other than you does the translating.
Which raises the obvious question. If the multiplier is sitting in plain sight, and BCG has put the prescription in print, why does almost nobody collect it?
Why nobody collects the prize
The skeptic’s case deserves a fair hearing, because most of it is right. Your companies have almost nothing in common operationally. One has a modern cloud stack because a CTO cared; one still runs on the systems it grew up with, which have the considerable virtue of working; one is mid-integration on an add-on (two CRMs, two definitions of “customer,” one board expecting one set of numbers). The play you proved at the first company quietly assumed things – clean order data, a sales team of a certain shape, someone in the building who can own a new workflow – that the fourth company does not have. Playbooks assumed to fit everywhere die on contact. Every partner who has sponsored a portfolio-wide rollout has watched it happen, usually after the vendor’s slide about “seamless scaling.”
Which is why the playbook pitches get deleted. The pitch leads with where the play travels and goes silent on where it does not, and a partner knows the portfolio far too well to believe in frictionless anything. But the skeptic’s error is in the conclusion, not the observation. Heterogeneity is a fact about the portfolio. It is not an argument against translation – it is an argument about how translation has to be done.
The map of where the play does not travel
Honest translation starts with a “no-go map” – a written account of where the proven play does not travel, and why. Before anyone schedules anything, someone with portfolio-level judgment places each company against the play. At one company the data can carry it. At another the workflow exists but nobody is positioned to own it – and a play without an owner is a license without a user. At a third, the economics mean it would not matter even executed perfectly. BCG’s Deploy–Reshape–Invent framing is useful vocabulary for the placement – most companies sit at “Deploy,” tool licenses handed out with nothing reorganized around them – but the discipline matters more than the taxonomy. The output that matters is the list of companies where the answer is “not here, not yet.”
I am compressing – in practice the placement is a dozen questions per company about data, workflow ownership, and unit economics – but the compression does not change the shape.
That list is what separates a translation from a rollout. A rollout treats the portfolio as homogeneous and pays for the assumption in dead pilots. A translation treats the no-go map as half the deliverable – arguably the more informative half, because it is the document proving that whoever carries the play has studied the companies and not just the playbook. The fragmentation the skeptic points to is real, which is exactly why the map of it comes first.
The economics follow from there. I have argued separately that most of AI’s value is decided in selection, before anything is built – the front gate matters more than the build – and translation is running proven material through that gate again. The second pass is cheaper than the first, because the expensive uncertainty (does this play produce value at all?) was paid down at company one. Each pass after that is cheaper still. That is the compounding, and there is nothing mystical in it: a proven play is a documented change in how work gets done, plus the evidence that it paid.
From the top of the market, the survey evidence points the same direction. FTI’s 2026 study samples funds a notch above this bracket – its floor is $1B under management – but its finding on the outperforming tier travels: the outperformance “holds regardless of fund size, industry or AI spend... What sets these funds apart is not how much they invest in AI. It is where and how they deploy it.” Where and how is a placement discipline. It is not a budget.
The first company keeps its win
None of this survives contact with a management team unless the ownership rules are stated up front and applied to every company the same way. There are, I believe, four.
First: the first company keeps its win. The pilot company is the proof, not the experiment. Its CEO shows the board what the team built; the team keeps the capability and the named credit; and the fund reusing the pattern later takes nothing away from any of it. A win that gets absorbed into “the fund’s AI program” the moment it works is a win no CEO will produce twice.
Second: every receiving company gets its own version, with the same protections. The adaptation is to its systems and its economics, the ownership sits with its team, and its CEO is the one presenting – not the fund. Nobody inherits a foreign playbook with another company’s name rubbed out.
Third: what travels is the pattern and the discipline – never a company’s data. The first company’s numbers, customers, and internals stay home. What the next company receives is “here is a play that worked, here is what it requires, here is how you will know it is paying,” not a look inside a sister company.
Fourth: no company is graded against another. The no-go map is a statement about fit, not merit; “not here, not yet” protects a company from a play that would waste its year, and that is all it says. The moment translation turns into a league table, every CEO in the portfolio starts managing the table instead of the business – and the wins stop being real.
These rules are not courtesy; they are where working deployments come from. MIT’s 2025 research on enterprise AI – a small early study its authors call preliminary, so hold it loosely – found that the deployments which worked were driven by the people who own the workflow, not handed down from a central program. The win belongs to the team that made it.
I also hold this view for a biographical reason. I have been the person a sponsor introduces into one of its companies: I ran analytics inside a ~$1B PE-owned telecom, brought in through the sponsor’s operating partner, and the work went well for one reason above the others – it was done for the company, at its leadership’s request, on questions they wanted answered and could not get to from inside. The team ended up with a capability it kept. That is the receiving end of this motion, and it is the version of it a management team is glad to see arrive.
(If you run a portfolio company and this article reached you as a forward from your sponsor, the rules above are not decoration. They are written down, addressed to you, at what this looks like from your side.)
Who carries it
So translation needs doing, and it needs rules. Who does it?
The first company’s team cannot. It has the play and the scars, but no standing at a sister company – and, more to the point, a business to run. Asking the team that just produced your proof to spend two quarters evangelizing it across the portfolio is a reliable way to lose the proof.
Company-by-company consultants cannot either. Each engagement is local: scoped to one company, paid by one company, gone at the end. Nothing in the arrangement accumulates portfolio-level memory, and portfolio-level memory is the raw material of the no-go map.
The partners could, in principle. They have the standing and the altitude. What they do not have is the hours: translation is operating work, measured in months of unglamorous adaptation, and a partner’s calendar is the scarcest resource the fund owns. If the play only travels when a partner personally carries it, it will not travel. The calendar is the wrong engine.
A play cannot carry itself.
What the job requires is one carrier with two properties at once: portfolio-level standing – the fund’s mandate, the map in hand, memory that persists across companies – and company-level trust, earned by working for each management team under the rules above. In practice that is a retained senior advisor who works across the portfolio. The judgment is retained at portfolio altitude; the work each company takes is engaged by that company and owned by its team – the first ownership rule, in commercial form – with the translation running on the advisor’s hours and never on yours. (The same small MIT study, same caveats, found external partnerships reaching deployment at roughly twice the rate of internal builds; the correlation may just reflect which organizations seek help in the first place. Direction, not gospel.) I have made the longer case for this shape of help elsewhere – the case for the retained advisor – but the point here is narrower: each of the other candidates lacks one of the two properties, and the job needs both.
One conversation, not eight
The same win, in a standalone company, is a good year. In a portfolio, under the rules above, it is the start of a compounding position – proven and paid for once, adapted wherever the map says it travels. That is why I think the right unit of decision is the fund, not the company. Run AI as eight separate company conversations and the best case is scattered wins that end at each company’s walls. Run it as one fund conversation – first win chosen carefully, ownership rules declared, no-go map built, carrier retained – and one win becomes a playbook. It also becomes a better answer to the AI question LPs now ask in diligence than any list of pilots: one real win, translated with stated discipline, and a written record of where you chose not to spend.
Whether AI deserves a place in your portfolio’s value-creation plans at all is a question the work should answer, not the marketing. At some of your companies it plausibly does. At others the honest answer is “not yet,” and a discipline that cannot say so is not a discipline.
If you want the starting point, it is deliberately small. The Portfolio AI Baseline is a fund-level read of where a first play is worth proving and where it is not. It costs a partner under three hours – one working session and one playback; the data request goes to your analysts, and the rest of the work happens off your calendar. It is free because the baseline is how I choose the work – money is made only if the fund decides to continue past it.