Build vs Partner for Portfolio AI: The 3-Model Staffing Test

TLDR: Build vs partner for portfolio AI is not one decision. It is a decision per workflow, and the answer changes with how many portfolio companies share that workflow, how much proprietary data it touches, and how much hold period is left. The Three Repeats Rule says: partner for the first instance, centralize anything that repeats in three or more portfolio companies, and embed a dedicated team only where the workflow is the product. Central AI functions fail when they are built before anything repeats.

Best For: AI or digital operating partners and value-creation directors at mid-market PE funds whose portfolio companies are enterprise scale (1,000 to 15,000 employees), who have been asked how the fund will staff AI and need an answer the investment committee can underwrite.

Updated: October 2026

Build vs partner for portfolio AI is a staffing decision that determines who does the AI work across a fund's portfolio companies: a central AI function at the fund, operators embedded inside each company, or an external partner brought in per engagement. Most funds treat it as an org chart choice made once. I would treat it as an underwriting choice made per workflow, because each model has a different failure condition and a different cost curve, and a portfolio at mixed maturity will need all three at different points in the hold. The argument of this post is that the central AI function is the most common first move and the most frequent failure, and that the decision only becomes defensible when it is tied to a rule about repetition.

Why Build vs Partner for Portfolio AI Is an Underwriting Question, Not an Org Chart Question

Build vs partner for portfolio AI is an underwriting question because each staffing model carries a different risk of producing no measurable outcome by exit. A central team is a fixed commitment against uncertain demand. Embedded operators are a bet on one company's data and leadership. A partner is a variable cost that leaves nothing behind unless the contract says otherwise. The investment committee is underwriting one of those risk profiles whether or not the memo says so.

The three models, defined

A central AI function is a team employed at the fund (or at a shared services entity) that builds and runs AI capabilities for multiple portfolio companies. Embedded operators are AI or data people employed by one portfolio company, often seeded by the fund, who own that company's AI work. An external partner is a firm engaged for a defined scope: an audit, a build, a managed workflow. "Hire" in the roadmap language usually means the second model; "build" can mean either of the first two; "partner" means the third. The three are not alternatives to each other so much as instruments for different situations.

What the surveys say about who is actually doing the work

Very few funds have settled this, and the 2026 surveys show it. In the Accordion PE AI adoption benchmark, a May 2026 survey of 150 AI, data, and technology operating partners, only 9% of sponsors had a dedicated AI center of excellence actively supporting portfolio companies, while 63% of portfolio companies had no formal AI structure at all, only informal guidance. Just 7% of deployments were implementation partner led. The same survey found 63% of operating partners naming talent and capability as a barrier to scaling, second only to data infrastructure at 71%.

At the fund level the pattern looks different. EY's AI Pulse survey from Q4 2025 found that 84% of PE firms had appointed a chief AI officer, yet 62% of those same firms said they struggle to link specific productivity gains to AI adoption. A title exists; a staffing model that produces evidence mostly does not. L.E.K.'s PE Pulse 2026, surveying 100 US buyout professionals in April 2026, found 68% of firms using outside advisors for value creation work and 46% expanding internal operations teams in the same year, which is the "all three at once, with no rule" state this post is trying to replace.

The outcome data explains why the question matters. FTI Consulting's 2026 Private Equity AI Radar, surveying 200 fund and operating leaders in May 2026, found 36% of portfolio companies using AI across multiple use cases but only 7% at enterprise scale, and 35% of respondents named talent as the primary constraint on scaling. Bain's 2026 GP Outlook found 39% of GPs expecting no material financial impact from AI on portfolio companies in 2026 at all.

How thinking on this has moved since 2023

In 2023 and 2024 the default answer was a fund-level head of AI, modeled on the digital value creation leads funds hired a decade earlier. By 2025 the limits of that model were visible: BCG's work on AI-first companies in private equity warned sponsors directly, "do not assume fund-wide AI expertise can sit with a single head of AI," and argued for AI capability by function rather than by title. McKinsey's June 2026 study of 471 PE-backed companies framed the current state as a "rent versus buy" choice, with renting meaning centralized platforms and experts redeployed from the fund, buying meaning dedicated teams inside a portfolio company where proprietary data or product change justifies it, and most leading firms running a hybrid. By 2026 the open question is which rule decides, not which model wins.

Central AI Function vs Embedded Operators vs External Partner: Where Each One Fails

A central AI function, embedded operators, and an external partner each fail in a predictable place. The central function fails on demand: it is built before any workflow repeats, so it spends its first year selling itself to portfolio CEOs. Embedded operators fail on isolation: one company's learnings never leave. The external partner fails on residue: the engagement ends and nothing stays. Choosing a model is choosing which of these failures you can afford for a given workflow.

Where the central AI function fails

The central function is the most intuitive answer and the one that most often turns into overhead. McKinsey's State of AI 2025 found that only 39% of organizations report any enterprise-level EBIT impact from AI even though 88% use it somewhere, and that high performers are nearly three times as likely as others to have fundamentally redesigned a workflow rather than layered AI on top. A central team that does not own a workflow cannot redesign it; it can only advise. The strongest predictor McKinsey found was leadership ownership: 48% of high performers strongly agreed that senior leaders show strong ownership of AI initiatives, against 16% elsewhere. A central function at the fund does not supply that ownership inside a portfolio company. The portco CEO does, or nobody does.

Where embedded operators fail

Embedding is right when the workflow is the product, and hard to justify everywhere else. ManpowerGroup's 2026 Talent Shortage survey found 72% of employers unable to find the skills they need, with AI model and application development the single hardest capability to fill worldwide for the first time. An enterprise portfolio company in distribution or manufacturing is competing for that talent against companies that can offer equity upside and a technical culture. When the hire does land, the result is often good and trapped: a working agent in one company's accounts payable queue, built on that company's ERP quirks, documented in one person's head. If the hold has under three years left, the fund is paying for capability it will not get to reuse.

Where the external partner fails

The partner model has the best short-run odds and the worst long-run residue. The MIT NANDA report on the GenAI divide, published in 2025, found that purchased or partnered AI deployments succeeded about 67% of the time while internal builds succeeded only a third as often, and that 95% of enterprise pilots overall delivered no measurable P&L impact. The partnered deployments that worked were the ones with a defined workflow, a named internal owner, and a measurement standard set before the engagement started. The ones that failed looked identical on the day the partner left. Six months later nobody could run the thing. The failure is not in the partner. It is in the contract that did not specify what stays.

How to Decide Build vs Partner for Portfolio AI: The Three Repeats Rule and the 4-Input Staffing Grid

Deciding build vs partner for portfolio AI works best as a rule applied per workflow rather than a model applied to the whole fund. The Three Repeats Rule is the version that has held up across mixed-maturity portfolios: partner for the first instance of any workflow, centralize once the same workflow appears live in three or more portfolio companies, and embed a dedicated team only where the workflow is proprietary to one company's product or data. Everything else in the staffing decision is a refinement of that rule.

The Three Repeats Rule

The rule exists because repetition is the only thing that justifies a fixed cost. A workflow that appears once is an experiment, and experiments should be variable cost. A workflow that appears three times is a pattern, and a pattern is what a central function is for: the third instance of an invoice intake agent or a contract review queue should cost a fraction of the first, and that fraction is the central team's reason to exist. The contrarian position here is that a fund should not hire a central AI team until it can name the three repeats. Most funds hire first and then look for repeats, which is why BCG's 2025 survey of more than 1,800 executives found only about a quarter of companies creating significant value from AI, with leaders concentrating on an average of 3.5 use cases against 6.1 for everyone else. A central team without repeats produces the 6.1.

The 4-input staffing grid

For each candidate workflow, score four inputs. The inputs are named so an operating partner can defend the score to a partner who asks what it is based on.

Read the scores this way. Repetition at 2 with proprietary dependency at 0 or 1 points to the central function. Proprietary dependency at 2 with hold period at 2 points to embedded operators. Anything with repetition at 0, or owner posture at 0, points to an external partner with a handover clause, or to not doing the work yet. Owner posture at 0 is a stop sign for all three models, which matches the pilot autopsy pattern described in why portfolio company AI pilots fail: the operating model was never tested, so the staffing model never mattered.

The sequencing rule, and the one case that needs a dedicated program

Sequence the three models rather than choosing one. Quarter one: partner on the first instance of a workflow in one portfolio company, with the baseline measured before the engagement starts using the per-workflow ROI formula. Quarter two and three: run the same workflow in a second and third company, with the partner's documented playbook and a portco owner in each. Only after the third live instance does the fund hire the central function, and its first job is the fourth instance, not a strategy.

For a portfolio where several enterprise-scale companies share the same high-volume back-office workflow, where two or more are already spending on AI with different vendors, and where the next partner meeting expects a staffing answer with evidence behind it, a generic approach is not enough; the answer is a dedicated portfolio program that does three things: it baselines the shared workflow in every company where it exists before any model is chosen, it runs the first instance to a measured outcome with a handover standard written into the contract, and it converts that outcome into the repeat criteria and the staffing scores the investment committee signs off on. A single head of AI hire does not do those three things, and a vendor roster does not either.

What the Central Function Should Own, and What It Should Never Own

A central AI function should own the things that are identical across portfolio companies and nothing that depends on one company's leadership. That means the measurement standard, the vendor and partner roster, the reusable workflow playbooks, and the review cadence. It should never own a portfolio company's outcome, its baseline, or its decision to proceed. The moment the central team owns the outcome, the portco executive stops owning it, and the fund is back to the 16% leadership ownership rate McKinsey measured among non-high performers.

Four things that belong at the fund level

The measurement standard comes first, because without it the fund cannot compare one company's result to another's, and Grant Thornton's 2026 PE survey found 45% of PE leaders still piloting AI, 11 points above the cross-industry rate, with only 9% very confident they could pass an AI governance audit in 90 days. The partner roster is second: a fund that has run the same workflow with three different firms in three companies has paid for three discoveries. The playbook library is third, and it is the central function's actual product. The review cadence is fourth, and it is where the portfolio spend register gets read.

Three things that belong inside the portfolio company

The baseline belongs to the company because the company's data team has to defend it. The owner belongs to the company because the fund cannot attend the weekly operations meeting. The go or no-go decision belongs to the company because, as PwC's 2025 survey of 564 PE-backed CEOs found, only 14% of PE-backed CEOs say AI has both raised revenue and lowered costs, and more than half report no AI upside at all. A CEO who was handed an AI program by the fund is not in the 14%. The staffing comparison in how to staff an AI center of excellence applies inside a single enterprise; across a portfolio, the center of excellence has to be far thinner than a company would build for itself.

What Skeptics Get Wrong About Staffing AI Across a Portfolio

The three objections an operating partner will hear in the partner meeting are predictable, and each has a direct answer.

"Just hire a head of AI and a small team, that is what we did for digital." Digital value creation had a repeatable playbook (pricing, procurement, sales effectiveness) before the hire. AI across a mixed portfolio does not yet have one, which is why 41% of sponsors in the Accordion benchmark describe themselves as scaling AI across multiple companies with no operational playbook. Hiring before the playbook exists produces a team whose first year is spent writing one from secondhand examples.

"Partners leave nothing behind." That is true of partner contracts with no handover clause, and it is the fund's fault, not the model's. A contract that specifies the documented playbook, the trained portco owner, the measurement standard, and the exit criterion leaves exactly what a central team would have built, at variable cost, with a live result attached. The 4-question decision grid for portco AI proposals is the same test applied to the partner's proposal.

"This is all RPA with a new label." Some of it is, and Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, partly because of "agent washing" of ordinary automation. The staffing grid is indifferent to the label. If a workflow scores 2 on repetition and 0 on proprietary dependency, it is a central-function candidate whether the tool is an agent or a script, and the baseline will show which.

Making the Staffing Choice Underwritable for the Investment Committee

An underwritable staffing decision for portfolio AI is one where the investment committee can see the inputs, the rule that converts them into a model, the cost shape of that model, and the condition under which the fund would reverse it. Deloitte's 2025 survey of 1,000 senior investors found 81% of PE firms expecting measurable return from generative AI within one to three years; the staffing memo is where that expectation either gets a mechanism or stays a hope.

The one-page staffing memo

The memo has five lines per workflow: the workflow name and the companies where it exists at volume; the four scores; the model the scores point to; the cost shape in words (fixed at fund, fixed at company, variable per engagement); and the exit criterion, meaning the result that would move the workflow to a different model or stop it. McKinsey's PE study found companies with AI in their operating model carrying revenue per employee roughly 20% higher than pilot-stage peers; the memo is how the fund shows which staffing choice gets a company from pilot to operating model and which one just adds headcount.

The reversal condition

Every model needs a stated reversal. A central function is reversed if, two quarters after hire, fewer than three workflows are live in two or more companies. Embedded operators are reversed if the company's baseline shows no movement after one full cycle of the workflow. A partner is reversed if the handover deliverables are not accepted by the portco owner by the contract end. Stating the reversal in the memo is what separates a staffing decision from a staffing preference, and it is the thing partners most often say was missing when an AI program is wound down. AI due diligence at acquisition is where the same four inputs get scored for a target before the fund owns it, so the staffing decision can be made on day one rather than discovered in year two.

This analysis was developed using methodologies and operating experience from Assembly.

Frequently Asked Questions

What does build vs partner for portfolio AI mean?

Build vs partner for portfolio AI is the decision about who does the AI work across a fund's portfolio companies: a central AI function at the fund, operators embedded inside each company, or an external partner engaged per workflow. The decision is best made per workflow using repetition, proprietary dependency, hold period, and owner posture as the inputs.

What is the Three Repeats Rule?

The Three Repeats Rule says a fund should partner for the first instance of any workflow, centralize the workflow once it is live in three or more portfolio companies, and embed a dedicated team only where the workflow is proprietary to one company's product or data. Repetition is the only input that justifies a fixed central cost.

What is the difference between a central AI function and embedded operators?

A central AI function is employed at the fund and serves several portfolio companies with shared playbooks and measurement standards. Embedded operators are employed by one portfolio company and own that company's AI work. The central model suits repeated generic workflows; the embedded model suits workflows tied to one company's proprietary data or product.

When does a central AI team at a PE fund fail?

A central AI team fails when it is built before any workflow repeats across portfolio companies, so its first year is spent manufacturing demand. The Accordion 2026 benchmark found 41% of sponsors scaling AI across companies with no operational playbook, which is the condition a premature central team usually produces.

When do embedded AI operators inside a portfolio company fail?

Embedded operators fail on isolation and hold period. Results stay trapped in one company's systems and one person's head, and if fewer than two or three years remain before exit, the fund pays for capability it never reuses. ManpowerGroup's 2026 survey found AI development the hardest skill to fill worldwide, which makes the hire slow and expensive.

When does an external AI partner fail a portfolio company?

An external partner fails when the contract has no handover clause, no named portco owner, and no measurement standard set before work starts. The MIT NANDA 2025 report found partnered deployments succeeding about 67% of the time, roughly three times the rate of internal builds, but only when the workflow and owner were defined.

How should an operating partner decide the staffing model for a given workflow?

An operating partner should score four inputs per workflow: repetition across portfolio companies, proprietary dependency, hold period remaining, and owner posture, each from 0 to 2. High repetition with low proprietary dependency points to the central function. High proprietary dependency with a long hold points to embedded operators. No repeats or no owner points to a partner, or to waiting.

How does the hold period change the build vs partner choice?

Hold period decides whether a fixed cost can pay back. With under two years remaining, embedded hires rarely return their ramp time and a partner with a handover clause is usually the better model. With over four years, a proprietary workflow can justify a dedicated in-company team. Central functions depend on repetition more than on any single company's hold.

What should a central AI function at a fund own?

A central AI function should own what is identical across portfolio companies: the measurement standard, the partner and vendor roster, the reusable workflow playbooks, and the review cadence. It should never own a portfolio company's outcome, baseline, or go decision, because those belong to the company executive who has to defend them.

Why is a head of AI hire not enough for a PE portfolio?

A single head of AI cannot supply leadership ownership inside each portfolio company. McKinsey's State of AI 2025 found 48% of high performers reporting strong senior leader ownership of AI against 16% elsewhere; that ownership has to sit with the portco CEO. BCG's 2025 PE guidance warned against assuming fund-wide expertise can sit with one title.

What is the right sequence for staffing AI across a portfolio?

The right sequence is partner, repeat, then centralize. Partner on the first instance in one company with a measured baseline, run the second and third instances using the documented playbook with a portco owner in each, and hire the central function only after the third live instance. The central team's first job is the fourth instance, not a strategy document.

How do you make an AI staffing decision underwritable for the investment committee?

An underwritable staffing decision shows the inputs, the rule, the cost shape, and the reversal condition. Scores, the model they point to, whether the cost is fixed at the fund, fixed at the company, or variable per engagement, and the result that would reverse the choice. Without the reversal condition it is a preference, not a decision.

What is a reversal condition for an AI staffing model?

A reversal condition is the stated result that moves a workflow to a different staffing model or stops it. A central function is reversed if fewer than three workflows are live in two or more companies two quarters after hire. Embedded operators are reversed if the company baseline shows no movement after one full workflow cycle.

Is staffing AI across a portfolio just RPA with a new label?

Some portfolio AI work is ordinary automation relabeled, and the staffing grid does not care. Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by 2027, partly because of agent washing. A workflow with high repetition and generic logic is a central-function candidate whatever the tool is called.

How common are dedicated AI teams at PE funds in 2026?

Dedicated AI teams at PE funds are still rare in 2026. The Accordion benchmark found 9% of sponsors with an AI center of excellence actively supporting portfolio companies and 63% of portfolio companies with no formal AI structure, while EY's Q4 2025 AI Pulse found 84% of PE firms had appointed a chief AI officer.

What role should an outside partner play once a central AI function exists?

An outside partner should handle first instances and specialist builds after the central function exists, not the repeated workflows the central team now owns. Every new workflow still starts as a single instance somewhere, and the partner model remains the fastest way to prove it, provided the contract hands the playbook to the central function at the end.