Nearly everyone has AI. Here's how the 6% get the EBIT.
A benchmark built from 14 published studies and read against what we see inside PE-backed builds. Where portfolios stand, what the leaders do differently, and what a buyer will check.
Six findings, PE-specific data first, and the five questions a buyer will ask about AI in diligence.
See the findingsSix questions place you on the five levels and name the one gap to close first.
Find your levelpublished by
Past the pilot is where it gets expensive.
The firms getting EBIT from AI rebuild the workflow instead of adding a tool, keep data clean enough for an agent to get the same answer twice, and put someone with authority on the outcome. One finding runs against the market story: no public dataset shows an AI premium at exit yet.
On method. We didn't run a survey for this report. Every number comes from a named, published study, linked in the sources. Our view is labeled as ours; field notes come from Blue Orange engagements with client details removed.
Six findings. Pick one.
Each finding is one number from a named study. Select a finding to see the data behind it, then jump to the chapter.
Most portfolio companies use AI somewhere. Few run it at a scale that shows up in the P&L: 36% use it across use cases, 7% at enterprise scale[10].
Deployed isn't working. Working isn't compounding.
Ask an operating partner how AI is going across the portfolio and you'll hear two things at once. The sponsor is convinced it matters. The portfolio is moving slower than the conviction, one management team at a time.
FTI's 2026 radar puts numbers on the gap: 36% of portfolio companies use AI across use cases, and only 7% run it at enterprise scale[10]. Around 40% of buyout GPs expect no material impact from AI this year[11]. The firms pulling ahead treat AI as an operating program with a P&L target, not a set of pilots.
Most companies are stuck in the expensive middle.
BCG's 2025 study of 1,250 executives[3] sorts companies into three groups. The 5% it calls future-built get value at scale. The 35% in the middle are scaling, but not fast enough. The other 60% report minimal gains despite real spend. The middle is where budget gets spent without a playbook.
"The portfolio company had built several AI tools in house, and each one stalled at the same point. The first 80 percent comes fast. The last 20 percent, making it reliable enough to run in production without someone checking every output, is where most builds stop."
Tools make people faster. Workflows move EBIT.
Eighty percent of McKinsey's 2026 respondents say AI made them more productive. Only 37% see it in EBIT[1]. MIT found the same split: general tools lift individual output while most enterprise pilots never reach the P&L[4].
The 6% McKinsey calls high performers, who attribute at least 5% of EBIT to AI, are about three times as likely to have fundamentally redesigned workflows[1]. A copilot takes minutes off a step. A redesigned workflow removes the step, and that is what shows up as margin.
BCG puts the effort split for AI value at 10% algorithms, 20% technology and 70% people and process[3]. If most of the budget is licenses, you've funded the 30%.
The bottleneck isn't the model, and never was.
It starts at close: a new portfolio company arrives with the prior owner's systems and definitions that differ by entity. PE fund leaders name talent as the top constraint, cited by 35%[10]. It's the same finding: the scarce people are the ones who can make inherited data usable.
Gartner found 63% of organizations don't have, or aren't sure they have, AI-ready data practices, and expects 60% of AI projects without that data to be abandoned[5]. For a sponsor, the cheapest time to fix definitions is the first 100 days, before any agent depends on them.
"If you're building for a human, humans catch those errors innately. An agent won't, so you need catches when it doesn't have the context."
Ownership decides whether AI pays.
High performers are twice as likely to say senior leaders are visibly committed to AI[1]. Most companies don't have that yet: ownership is split across four layers[14]. And handing AI to one function rarely moves the whole company. Finance shows it: adoption sits at 59%, barely up from 58%, and 91% of finance leaders report low or moderate impact[7].
The C-suite as a group owns AI strategy in only 32% of companies; the rest spread it across business leaders, VPs and functional heads[14]. Shared ownership tends to mean nobody owns the result. The model that works names three people: the CEO for the mandate, an operator for each workflow, and the CFO for how impact is counted.
The ownership model that works.
In the builds we see succeed, ownership splits three ways: one person owns the outcome, one owns each workflow, and one decides whether a result counts.
Finance is often a strong first workflow, since AP automation and anomaly detection are already among its top use cases[7]. The program still needs an owner above any one function.
Agents are real. Most aren't governed yet.
KPMG found agent deployment at asset managers and PE firms jumped from 19% to 30% in one quarter[9]. Gartner expects more than 40% of agentic projects to be cancelled by the end of 2027, and counts only about 130 real agentic vendors among thousands[6].
Only 49% of leaders have defined where agents can't decide alone[9]. That is the number to watch. An agent without a written boundary can't be audited, and one that can't be audited won't hold up in a buyer's diligence.
Supervised autonomy is the right boundary.
The agents that hold up are narrow. They watch a rules-based process, surface exceptions with context, and route them to a person. Let the model read, parse and draft. Keep the numbers deterministic, and log every action.
- Agent runs
- Covenant monitoring, variance detection, close orchestration
- Code computes
- Adjusted EBITDA and any figure that has to be right
- Person decides
- Every exception, with the context and a full audit trail
"For a PE firm with 600-plus portfolio companies reporting monthly, adjusted EBITDA runs through a rules engine inside the agentic workflow. The hallucinations disappear by design."
If multiples are the test, AI hasn't been graded.
A review of public research in October 2026 found no dataset showing buyers paying a premium for AI-driven gains, only anecdotes, some cutting the other way[15]. Meanwhile a buyout needs about 12% annual EBITDA growth to match what 5% delivered a decade ago[15]. The case for AI in the hold is growth you can prove.
Diligence is catching up faster than valuation: only 9% of PE firms say they could pass an independent AI governance audit within 90 days, against 22% across all industries[13]. AI that moves a named P&L line, runs without its builders and leaves an audit trail shows up in the EBITDA a buyer already pays for.
Build for diligence, not for a premium.
The five questions we'd expect in a 2027 diligence call. Tick the ones your portfolio company could answer with documents today.
Not diligence-ready yet. Start with the P&L line and the data owners.
One company. Five levels. Two minutes.
Six questions, one per dimension in this report. Answer for a single company as it runs today. You'll land on a level and get the one gap to close first. Nothing you enter leaves this page.
The six dimensions follow the chapters: use, workflow, data, ownership, governance and measurement. The levels run from Level 1, Chat & Copilot, to Level 5, Autonomous Operations. Level 3 is the first level where a workflow, not a person, does the work, and the first where its impact can be counted on a P&L line.
- Level 1Chat & Copilot
- Level 2Connected Intelligence
- Level 3Workflow Agents
- Level 4Orchestrated Systems
- Level 5Autonomous Operations
Level 3 is where cost per seat becomes cost per task.
How we built this benchmark.
This is a synthesis, not a survey. We kept fourteen studies fielded between July 2024 and September 2026 that disclose who they asked and when, covering more than 10,400 respondents. We don't combine numbers across studies, since each asks a different question of a different group.
Five sources are specific to private equity and asset management, marked PE. Field notes are quotes from Blue Orange client sessions in 2026, client details removed. We'll refresh the benchmark each quarter.
Frequently asked questions.
How many PE portfolio companies use AI today?
Most use it somewhere, few at scale. FTI Consulting's 2026 survey of 200 fund and operating leaders found 36% of portfolio companies using AI across use cases and 7% at enterprise scale[10].
What's the biggest barrier to scaling AI?
Data readiness, closely followed by the talent to fix it. Gartner found 63% of organizations aren't sure their data practices are ready for AI[5].
Who should own AI in a portfolio company?
The CEO owns the mandate, a named operator owns each workflow, and the CFO owns how impact is measured. Senior ownership is one of the traits most linked to AI value[1].
Do buyers pay a premium for AI at exit?
Not in any public dataset yet[15]. Buyers price EBITDA they can verify, so AI work that moves a named P&L line shows up in the multiple you're already paid.
What share of companies get EBIT from AI?
About 90% of companies use AI in at least one function, 37% report any positive EBIT impact, and 6% attribute 5% or more of EBIT to AI[1]. McKinsey calls that 6% the high performers.
Are PE firms using AI agents?
Increasingly. KPMG found agent deployment at asset managers and PE firms rose from 19% to 30% between Q2 and Q3 2026, while only 49% of leaders have defined where agents can't decide alone[9].
What separates AI high performers from everyone else?
High performers are about three times as likely to have fundamentally redesigned workflows around AI, and twice as likely to say senior leaders are visibly committed[1]. They rebuild the work rather than adding a tool to the old process.
Sources15 sources, each linked
- [1]McKinsey, The state of AI in 20261,719 respondents, 97 countries · May to June 2026
- [2]McKinsey, The state of AI in 20251,993 respondents · June to July 2025
- [3]BCG, The Widening AI Value Gap1,250 executives · September 2025
- [4]MIT NANDA, The GenAI Divide 2025300+ initiatives, 52 interviews, 153 surveys · H1 2025
- [5]Gartner, Lack of AI-Ready Data Puts AI Projects at RiskData management leaders · Q3 2024 · February 2025
- [6]Gartner, 40% of Agentic AI Projects Canceled by 20273,412 webinar attendees · June 2025
- [7]Gartner, Finance AI Adoption Remains Steady183 CFOs and finance leaders · May to June 2025
- [8]Informatica, CDO Insights 2025600 data leaders
- [9]KPMG, AI Quarterly Pulse Q3 2026PE314 leaders plus 100 in AM and PE · Jul to Sep 2026
- [10]FTI Consulting, 2026 Private Equity AI RadarPE200 fund and operating leaders · May 2026
- [11]Bain and StepStone, 2026 PE GP OutlookPE100+ buyout GPs · Dec 2025 to Jan 2026
- [12]Bain, Global PE Report 2025: gen AI field notesPEInvestors with $3.2T AUM · September 2024
- [13]Grant Thornton, Buyers are asking questions about AIPE2026 AI Impact Survey
- [14]Pearl Meyer, AI ownership survey (Fortune)108 executives and board members · Feb to Mar 2026
- [15]Forbes, J. Sviokla, AI at the back of the deal cycleReview of published evidence · October 2026 · secondary
Let's find where AI pays.
Two hours, read-only, with a senior engineer. You'll leave with a ranked list of what is worth building in one portfolio company.