There is a gap at the center of AI value creation in private equity. By 2026, 98% of sponsors had mandated AI adoption across their portfolios, yet only about half of portfolio companies were actively implementing it — and across enterprises generally, roughly 95% of generative AI pilots deliver no measurable P&L impact. The gap is rarely the model and rarely the use case. In the published research and in our own engagements, it is almost always the same thing: the data underneath isn’t ready, and the value creation plan never budgeted to make it ready.
What the adoption data actually shows
The numbers are striking even after you discount for definitional differences between studies:
| Finding | Source |
|---|---|
| ~95% of enterprise GenAI pilots show no measurable P&L impact; only ~5% achieve rapid value | MIT NANDA, The GenAI Divide: State of AI in Business (2025) |
| Externally partnered AI deployments succeed ~67% of the time — internal builds succeed roughly a third as often | MIT NANDA, same study |
| >80% of AI projects fail — about twice the failure rate of conventional IT projects | RAND Corporation (2024) |
| 42% of companies abandoned most of their AI initiatives, up from 17% a year earlier | S&P Global Market Intelligence (2025) |
| 98% of PE sponsors have mandated AI adoption; only ~half of portfolio companies are actively implementing | Accordion, PE AI Adoption Benchmark — 150 operating partners (2026) |
| 68% of portfolio-company CFOs say they don't know where to begin with AI | Accordion, same benchmark (as reported) |
| Top adoption blockers: data security (67%) and data quality / availability (65%) | Deloitte, GenAI in M&A survey |
| Majorities of GPs rate AI ineffective so far for deal sourcing (64%) and portfolio monitoring (75%) | S&P Global Market Intelligence, 2026 PE Survey |
Two honest caveats, because they matter. These studies measure different things — “no P&L impact” is not the same as “formally cancelled,” and survey populations range from GPs to operating partners to mid-market CFOs — so the percentages shouldn’t be averaged into one number. And the MIT study is a preliminary report with a modest sample. Directionally, though, every serious measurement lands in the same place: mandates are nearly universal, measurable results are rare, and the blockers named most often are data blockers.
There’s also a measurement paradox worth naming. In Grant Thornton’s 2026 survey of fund and operating leaders, 95% report their AI initiatives are meeting or exceeding the original business case — while S&P Global found the share of companies abandoning most of their AI initiatives jumped from 17% to 42% in a single year. Both cannot be representative at once. The likeliest reconciliation is that “success” is being self-reported against soft criteria — a pilot shipped, a demo impressed — rather than measured against the P&L. That gap between declared success and measured impact is itself a data-readiness symptom: without a trustworthy measurement layer, nobody can prove which camp an initiative is in.
Why portfolio companies hit this harder than most
A mid-market company arriving in a portfolio was usually built by acquisition, bootstrapping, or both. In our engagements the on-arrival pattern is remarkably consistent: an ERP or accounting system (sometimes several, if the company itself did tuck-ins), a CRM that half the team uses, operational data living in the tools that run the business, and the connective tissue in spreadsheets. Each system holds its own version of basic entities — customers, jobs, revenue — and no two fully agree. This is the N-systems problem, and it compounds at the fund level: across our PE and VC clients it’s common to see a single firm operating with 10–15 different fund administrators and multiple CRMs accumulated across vehicles and vintages.
None of this data is “bad.” It ran the business. But it is not AI-ready: an LLM or analytics layer pointed at five systems that disagree produces fast, confident, wrong answers. That failure mode, not model quality, is what the 95% statistic looks like up close. It’s also why “is it really an AI company?” has become a standard diligence question: an AI story built on unready data doesn’t survive contact with production.
Anatomy of a stalled pilot
The failure has a repeatable shape. A portfolio company stands up an AI assistant over its operating data — a reporting copilot, a customer-question bot, an analysis tool. The demo, built on a curated extract, dazzles. Then production arrives: someone asks for last quarter’s revenue for a top customer. The CRM says one number, the ERP another, and the assistant — with no way to know which system is authoritative, that two records are the same customer under different names, or that “revenue” means bookings in one system and recognized revenue in the other — confidently blends them. The CFO catches the discrepancy against the board pack in week one. From that moment the tool is “the thing that gets numbers wrong,” usage collapses, and the pilot joins the 95%.
Coverage of the MIT study adds two corroborating details. Researchers describe a “learning gap” — enterprise AI tools that don’t retain context or improve from feedback — as a leading reason pilots stall, and they note that employees often quietly use personal AI accounts for drafting work even while the official initiative flounders. Both point the same direction: the models are capable, the generic use cases work, and the failures concentrate where the tool must be right about this company’s numbers. The pilot was asked to be smarter than the data underneath it. No model is.
What "data ready" concretely means
Data readiness isn’t a platform purchase; it’s a short list of conditions, each checkable:
1. One source of truth. Core data consolidated from the operating systems into a governed layer — typically a data warehouse organized so every number traces back to its source (medallion architecture is the common pattern).
2. Entities resolved. “Acme Services LLC,” “Acme Services,” and “ACME SVCS” recognized as one customer across systems — entity resolution is unglamorous and decisive, because every join and every AI answer inherits it.
3. Metrics defined in code. One definition of revenue, gross margin, churn — applied identically everywhere, so the AI and the board pack can’t disagree.
4. Lineage and quality gates. Anomalies flagged at ingestion, and the ability to show where any number came from — the difference between an answer and an auditable answer.
5. Safe access. The data reachable by the tools and people who need it, with controls that survive diligence. Once these five hold, AI use cases stop being science projects: retrieval has something trustworthy to retrieve.
What each level of readiness unlocks
Readiness isn’t binary, and neither is the AI roadmap. Each level of data maturity makes a class of use cases safe to deploy — and deploying a use case above the company’s level is how pilots end up in the stalled 95%:
| Readiness level | The data reality | AI that's safe to deploy |
|---|---|---|
| 0 — Spreadsheet-run | Systems disagree; joins happen by hand each month | Document drafting, summarization, meeting notes — nothing that answers with a number |
| 1 — Consolidated | Core sources landed in a warehouse; one place to look | Automated reporting, dashboards, anomaly flags on the consolidated data |
| 2 — Reconciled | Entities resolved; KPIs defined once, in code | Portfolio Q&A over trusted metrics, variance alerting, LLM-assisted analysis |
| 3 — Governed | Lineage, quality gates, and access controls in place | Data-room RAG, agent workflows, AI output an auditor or board can rely on |
The ladder doubles as a budgeting tool. Each level unlocks a visible class of capability, so the groundwork can be funded increment by increment against payoffs the operating partner can see — rather than as a single act of faith labeled “data transformation.”
What this means for the value creation plan
The implication is sequencing, not spending more. Plans that treat AI as a line item of use cases tend to fund a series of stalled pilots that each rediscover the same data problems. Plans that work fund the groundwork once — consolidation, entity resolution, metric definitions — and then run use cases on top, each cheaper than the last because the foundation is shared. The groundwork is not overhead on use case one; it is the platform for every use case, and for the reporting that proves any of them worked. We’ve written separately about what that foundation looks like at the fund and portfolio level.
The build-vs-buy evidence is unusually one-sided here. MIT’s finding — partnered deployments succeeding at ~67% versus internal builds at roughly a third of that rate — matches what sponsors discover on cost grounds anyway: replicating a senior data-and-AI team inside every portfolio company is the expensive path, and the scarce skill isn’t writing the pipeline, it’s knowing which groundwork matters for the thesis. The pattern that works is specialist-built and portco-owned: outside engineers stand up the foundation and first use cases, the company owns everything that gets built, and internal operators run it from there.
The sequencing rule: fund the data groundwork before the use cases, buy the build rather than the headcount, own what gets built — and measure the result in the same reconciled numbers that flow to LP reporting, so the gains are evidence, not anecdote.
The same logic is moving upstream into diligence. Sponsors have begun underwriting AI-readiness before close — scoping the data groundwork during the deal so the value creation plan prices it from day one, instead of discovering it as an unbudgeted surprise in quarter three. It mirrors the shift technology diligence went through generally: the question stopped being “does the technology work?” and became “what will it cost to make the thesis true?” Our technology due diligence methodology treats AI readiness as a first-class assessment dimension for exactly this reason: the remediation number belongs in the model, and sometimes in the price.
The five-question readiness test
An operating partner can gauge a portfolio company’s data readiness in one meeting:
1. If I ask two systems for last month’s revenue, do they agree? 2. Is there one place where customer, product, and financial data are joined — or does that join happen in a spreadsheet each month? 3. Does “gross margin” have one written definition, applied in code? 4. When a number looks wrong, can anyone trace where it came from? 5. If the one person who maintains the reporting left tomorrow, would the reports still arrive?
Two or more “no” answers means the honest first line of the AI plan is groundwork — and the encouraging part is that the groundwork is fast when it’s scoped as a first increment rather than a transformation: weeks to a consolidated, reconciled first report, not a year-long program.