Why Does Portfolio AI Spend Outrun Proof? How to Govern AI Spend Across a Portfolio in 3 Columns

TLDR: Portfolio AI spend grows faster than proof because each portfolio company buys its own tools, several of them overlap, and none were measured against a baseline before the first invoice. Usage dashboards then rise while efficiency quietly falls. Knowing how to govern AI spend across a portfolio starts with a 3-Column Spend Register (tools, overlaps, baselines) and a rule that no new spend is approved without a baseline.

Best For: AI Operating Partners and value-creation directors at mid-market PE funds whose portfolio companies are enterprise-scale (1,000 to 15,000 employees), especially anyone recently in the role who has two or three portcos already spending on AI with nothing the investment committee would accept as proof.

Updated: September 2026

AI spend governance across a portfolio is a fund-level discipline that decides which AI spending at portfolio companies is allowed to start, what it must be measured against, and when it gets stopped. It is different from measuring AI ROI, which happens after the fact, and from AI due diligence, which happens before ownership. The reason it matters now is a pattern that has repeated across PE portfolios since 2024: two or three portcos buy tools independently, several of those tools do the same job, usage climbs because the tools are easy to use, and nobody set a baseline, so nobody can say what changed. This post explains the pattern and the register that fixes it.

Why is portfolio AI spend growing faster than anyone can prove it paid off?

Portfolio AI spend outruns proof because spending decisions are made portco by portco while proof is expected fund-wide. Each company buys what its team wants, usually without a baseline, and reports usage rather than outcomes. Three structural gaps produce this: no shared inventory of what was bought, no rule that a baseline must exist before purchase, and no common definition of what "worked" means, so the fund adds up usage numbers that were never designed to add up.

The pattern in the numbers

The gap between spending and proof is not a portfolio quirk. PwC's 2026 Global CEO Survey of 4,454 CEOs found 56% reporting no meaningful financial benefit from AI, with only 12% seeing both cost and revenue gains. BCG's September 2025 research put 60% of companies in a group getting minimal value from AI despite significant investment, against 5% generating substantial value at scale. IBM's 2025 CEO study found only 25% of AI initiatives had delivered the return expected of them and half of CEOs describing their technology estate as disconnected and piecemeal.

Now put three of those companies in one portfolio. The fund sees three AI line items, three usage dashboards, and three different definitions of success. That is the operating partner's actual situation, and it is why the portfolio number is harder to defend than any single portco's.

Why PE portfolios feel it more than a single enterprise

A single enterprise with one CFO can, in principle, force one baseline. A portfolio cannot, because each portco has its own CFO, its own ERP, and its own idea of what AI is for. EY's Q2 2026 Private Equity Pulse found 76% of PE firms naming AI, automation, and data infrastructure as the top area of increased focus for operational value creation, ahead of margin improvement at 62%. The demand is fund-level. The purchasing is not. Bain's 2026 Global Private Equity Report argues that funds now need 10% to 12% average annual EBITDA growth from operations rather than multiple expansion, which is exactly the kind of number an unproven AI line item cannot contribute to.

How thinking on this has moved

In 2024 the question at most funds was whether portcos should use AI at all, and the answer was to let them experiment. By 2025 the experiments had become line items and the question became whether they returned anything; MIT's 2025 GenAI Divide report put the share of enterprise pilots with no measurable profit impact at 95%. In 2026 the question has narrowed to governance: under what rules a portco is allowed to spend on it at all. The post on how PE operating partners drive EBITDA with AI covers where the value sits; this one covers how to stop spending ahead of it.

Why "usage up" hides "efficiency down"

Usage metrics hide efficiency losses because usage measures activity and efficiency measures output per unit of input, and the two can move in opposite directions. A tool that saves ten minutes on a task but adds a review step, a second tool that overlaps with it, and a subscription nobody canceled will all push usage up. None of them push cost per completed workflow down, and some push it up. Without a baseline, the fund only sees the first curve.

The three ways usage rises while efficiency falls

The first is duplication. Two portcos, or two functions inside one portco, buy tools that do the same thing; usage doubles, output does not. The second is rework. An AI-drafted output that a person has to check and correct records as usage on the tool and as unrecorded time on the person. The third is the subscription tail: seats that were bought for a pilot, kept after it, and counted as adoption. Microsoft and LinkedIn's 2024 Work Trend Index found 78% of AI users bringing their own tools to work; in a portfolio that means the register is missing tools before it is even built.

Gartner's May 2026 note to CFOs put it bluntly: counts of pilots, tools rolled out, or use cases in production show that a function is moving, but they do not prove value. Gartner's March 2026 survey of 204 finance leaders behind that note found 63% experiencing slower-than-expected AI implementation in 2025. In a July 2026 Gartner survey, 45% of CFOs said their AI investments lean toward productivity, which is the category where usage and efficiency are easiest to confuse.

What the agent wave adds to the problem

Agents make this worse, because an agent's cost scales with how much it runs and its value scales with how much of that running replaced something. KPMG's Q2 2026 AI Pulse found 53% of organizations with agents deployed and 18% orchestrating multiple agents across workflows, up from 9% a quarter earlier, while only 26% reported full, real-time visibility into AI operating costs. Half are running agents. A quarter can see what the agents cost to run. Capgemini's July 2025 research found only 2% of organizations had fully scaled agents and fewer than one in five reported high data readiness. A portfolio with three portcos each running two or three agents has a running cost nobody at the fund can see and a value nobody baselined.

How to govern AI spend across a portfolio: the 3-Column Spend Register

How to govern AI spend across a portfolio comes down to one document the fund owns and every portco fills in: the 3-Column Spend Register. Column one lists every AI tool and agent by portco, function, and workflow. Column two marks overlaps, where two entries do the same job. Column three records whether a baseline existed before the spend. The register is small on purpose, so it gets maintained, and it gives partners three numbers: tools, overlaps, and unbaselined spend.

We built the register this way because, across the portfolios we have looked at, the first two questions a partner asks are "what did we buy" and "how much of it is doing the same thing," and the third is "what did it change." Most funds can answer none of the three inside a week. The register makes all three answerable in an afternoon.

Column one: the tool and agent inventory

Column one is an inventory, not a strategy. It lists what exists, including the tools that arrived through an expense report rather than procurement. The rule for inclusion is simple: if a portco pays for it or an employee uses it on company work, it goes in. The post on shadow AI and enterprise AI governance describes how to find the entries that were never reported.

Column two: overlaps

Column two is where the money is. Two portcos with separate contract-review tools, or a portco running a general assistant and a function-specific tool that both draft the same emails, are overlaps. In our experience the overlap share in a portfolio that has never built a register is typically between 15% and 30% of entries, and it is the fastest spend to remove because nothing has to be proven, only consolidated.

Column three: baselines

Column three is the one that governs future spend. A baseline is a measured description of the workflow before the tool arrived: volume, time per unit, error or rework rate, and who did the work. The post on how to establish an AI ROI baseline before deployment covers the method for a single company. The portfolio rule is stricter: no new AI spend at any portco is approved without a column-three entry, and existing unbaselined spend is given one quarter to produce one or it is reviewed for cancellation.

For a portfolio where two or more portcos are already running agents against production systems and a partner meeting is inside a quarter, a generic approach is not enough; the answer is a dedicated program that 1. builds the register across every portco in the first 30 days rather than one at a time, 2. freezes new AI spend until each portco's unbaselined entries have a measured baseline or a cancellation date, and 3. installs a monthly review where each portco reports the three column totals and one workflow-level outcome, so the fund's number is assembled from the same definition everywhere.

Governing AI spend vs. measuring AI ROI vs. AI due diligence

Governing AI spend, measuring AI ROI, and running AI due diligence are three different activities that operating partners tend to collapse into one. Due diligence happens before ownership and asks what exposure and readiness the fund is buying. Spend governance happens during ownership and decides what spending may start and under what measurement. ROI measurement happens after deployment and reports what changed. A fund that only does the third has nothing to measure against.

The earlier post on what AI due diligence is in private equity covers the first column. The post on tracking AI's EBITDA impact in a PE portfolio company covers the third. The middle column is the one most funds skip, and it is the one that makes the third column possible.

What partners and portco CEOs push back on

Operating partners hear the same three objections to spend governance: that a register could be generated in seconds, that this is a purchasing committee in disguise, and that the whole thing is RPA governance with a new name. Each deserves a direct answer, because two of them are half right.

"I could ask an AI to make that register in two seconds"

You could make the template in two seconds. You cannot make the contents, because the contents are the owners, the overlaps, and the baselines, and those live in the portcos. The work is getting three CFOs to fill in column three against the same definition, and the value is the comparison across portcos, which no single company can produce. The template is free. The register is not.

"This is a purchasing committee that will slow the portcos down"

Sometimes it becomes one, and that is the failure mode in the comparison table above. The safeguard is that the register governs measurement, not vendor choice. A portco can buy whatever it wants, as long as a baseline exists and the entry goes in column one. The only spend the fund stops is unbaselined spend and overlaps, and both are decisions the portco CFO would make anyway with the information in front of them. The post on what an AI investment committee is describes the lighter-touch version that fits inside a single company.

"We governed RPA the same way and it did not need this"

RPA had fixed costs per bot and did one thing, so a licence count was a reasonable proxy for usage and usage was a reasonable proxy for value. Agents have running costs that scale with volume, do variable work, and overlap with each other and with the general assistants employees already use. Deloitte's April 2026 survey of 3,235 leaders found only 21% of enterprises with mature governance for agentic AI. The licence count stopped being a proxy for anything when the tool started charging by the run.

How to govern AI spend across a portfolio in one quarter: the 4-step sequence

How to govern AI spend across a portfolio in practice is a four-step sequence over one quarter: build the register, remove the overlaps, baseline or cancel the rest, then review monthly. The order matters because each step produces the information the next one needs. It also matters politically: the first two steps show the partners a number within 30 days, which buys the time the third step needs.

Step one, in the first 30 days, is the register itself, filled in across every portco at once. Doing it one portco at a time produces three registers with three definitions and no portfolio number. Step two, in the same 30 days, is overlap removal, which is the only step that reduces spend without requiring proof; consolidating two tools that do the same job needs a decision, not a measurement. Step three, over the following 60 days, is baselining: every remaining unbaselined entry gets a measured baseline or a cancellation date. Step four is the monthly review, where each portco reports its three column totals and one workflow-level outcome against its baseline.

McKinsey's 2026 State of AI found 37% of organizations attributing at least some EBIT impact to AI, unchanged from the prior year, and only 6% qualifying as high performers with 5% or more of EBIT attributable to AI. The 37% is a self-reported attribution. The register turns it into something a partner can ask a follow-up question about, because every attribution in the monthly review points to a column-three baseline. The review is also where the fund learns which portco should become the reference case for the rest, which is the subject of a later post in this series.

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

Frequently Asked Questions

Why is portfolio AI spend growing faster than anyone can prove it paid off?

Portfolio AI spend outruns proof because spending is decided portco by portco while proof is expected fund-wide. Each company buys its own tools, usually without a baseline, and reports usage rather than outcomes. PwC's 2026 CEO survey found 56% of CEOs seeing no meaningful financial benefit from AI.

What is AI spend governance across a portfolio?

AI spend governance across a portfolio is a fund-level discipline that decides which AI spending at portfolio companies may start, what it must be measured against, and when it stops. It sits between AI due diligence, which happens before ownership, and AI ROI measurement, which happens after deployment. Its core instrument is a register of tools, overlaps, and baselines.

What is the 3-Column Spend Register?

The 3-Column Spend Register is a fund-owned document listing every AI tool and agent by portco (column one), every overlap where two entries do the same job (column two), and whether a baseline existed before the spend started (column three). The red flags are missing owners, overlap above one in five entries, and unbaselined spend above half the total.

Why do usage metrics hide efficiency losses from AI?

Usage metrics hide efficiency losses because usage counts activity while efficiency measures output per unit of input, and the two can move in opposite directions. Duplicate tools, rework on AI drafts, and subscription seats kept after a pilot all raise usage without lowering cost per completed workflow. Without a baseline, only the usage curve is visible.

How is governing AI spend different from measuring AI ROI?

Governing AI spend decides what spending may start and under what measurement, while measuring AI ROI reports what changed after deployment. Governance happens during the hold and belongs to the operating partner; measurement happens per workflow and belongs to the portco CFO. A fund that only measures finds it has no baseline to measure against.

What is the failure pattern of ungoverned AI tools across a portfolio?

The failure pattern of ungoverned AI tools is duplication, rework, and subscription tails: overlapping tools across portcos, AI outputs that people correct off the books, and seats kept after pilots end. Microsoft and LinkedIn's 2024 Work Trend Index found 78% of AI users bringing their own tools to work, so the register starts incomplete.

Should a PE fund centralize AI purchasing across portfolio companies?

A PE fund should centralize AI spend measurement, not AI purchasing. Portcos can buy what they want as long as a baseline exists and the entry goes in the register. The fund stops only unbaselined spend and overlaps. Centralizing vendor choice turns governance into a purchasing committee, which is the failure mode that slows portcos down and loses CEO support.

How do you set a baseline before new AI spend at a portfolio company?

A baseline before new AI spend is a measured description of the workflow as it runs today: volume, time per unit, error or rework rate, and who does the work. The portfolio rule is that no new AI spend is approved without one, and existing unbaselined spend gets one quarter to produce a baseline or a cancellation date.

Why do AI agents make portfolio spend harder to govern?

AI agents make portfolio spend harder to govern because their running cost scales with volume while their value depends on what that running replaced. KPMG's Q2 2026 AI Pulse found 53% of organizations with agents deployed but only 26% with full, real-time visibility into AI operating costs. A licence count no longer proxies for value.

How much of a portfolio's AI spend is typically duplicated?

Between 15% and 30% of AI tool entries in a portfolio that has never built a register typically overlap with another entry, in our experience. Overlap is the fastest spend to remove because consolidation needs a decision rather than a measurement, and it gives partners an early result while baselining is still under way.

How often should portfolio AI spend be reviewed?

Portfolio AI spend should be reviewed monthly, with each portco reporting its three register totals and one workflow-level outcome against its baseline. Quarterly is too slow for agent running costs that change with volume. The monthly review is also where the fund identifies which portco should become the reference case for the rest of the portfolio.

Who should own AI spend governance at a PE fund?

AI spend governance at a PE fund should be owned by the operating partner responsible for value creation, with each portco CFO owning that company's register entries. The operating partner sets the definitions and runs the monthly review. The CFO owns column three, because the baseline is a finance artifact and the cancellation decision is a finance decision.

How long does it take to install AI spend governance across a portfolio?

Installing AI spend governance across a portfolio takes about one quarter: 30 days to build the register and remove overlaps across every portco at once, then 60 days to baseline or cancel the remaining entries. The first 30 days produce a visible spend reduction, which buys the time the baselining step needs before the next partner meeting.

What share of companies actually get value from AI spend?

Roughly 5% of companies generate substantial value from AI at scale, 35% are beginning to, and 60% get minimal value despite significant investment, according to BCG's September 2025 research. For a portfolio, that means most portcos sit in the 60% by default, and the register is how the fund finds out which ones do not.

What is the first practical step for an operating partner facing unproven AI spend?

The first practical step is to build column one of the register across every portco in the same 30 days, using a single definition of what counts as an AI tool or agent. Overlaps appear as soon as the inventory exists, and removing them reduces spend before any baseline is measured, which gives the partners an early number.

When should a fund bring in an external partner for AI spend governance?

A fund should bring in an external partner for AI spend governance when nobody in-house has built baselines across several companies at once, or when portco CFOs will not accept a definition set by the fund alone. The partner's job is the shared definition and the baselines, not the tool choice. If they lead with a platform, keep looking.