The Demo Always Works
I’ve watched this scene play out more than once. A PE firm brings in a vendor — or asks their existing software provider — for an AI demo. The pitch is compelling. The model flags anomalies across data room documents, surfaces cross-portfolio performance comparisons, and pulls sector benchmarks faster than any analyst running a manual query. Leadership is interested. Budget gets approved.
Six months later, the tool is either barely used or producing outputs the deal team has quietly stopped relying on. Not because the AI failed, but because it was only ever working with one slice of the firm’s data. Ask it about pipeline coverage and it answers from the CRM. Ask it about portfolio performance and it answers from the monitoring platform. Ask it something that requires both, and the answer sounds plausible but isn’t complete. The team figures this out within a few weeks and reverts to doing it manually.
The firm’s conclusion is usually that AI isn’t ready for private equity. That’s the wrong conclusion. The right one is harder to sit with: the firm’s data wasn’t ready for AI.
Count the Systems. Then Count the Gaps Between Them.
Here’s a useful exercise. Map every platform that holds data your deal team touches in each quarter. For most mid-market PE firms, the list covers at least seven distinct categories: CRM and pipeline tracking; market intelligence and deal comps; proprietary sourcing of founder-owned businesses; virtual data rooms during diligence; fund accounting and portfolio monitoring; LP communications and capital account reporting; and deal documents and internal research stored across shared drives and intranets. Underneath all of it sits Excel, everywhere for everything, as the default connective tissue between all of the above.
Seven categories, often eight or more distinct platforms. None of them share a common data model. A market intelligence subscription may have an API connection into the CRM for certain fields, but there is no unified schema that lets you ask a question spanning both systems and get a coherent answer. The data exists in each platform on its own terms, structured around what that platform was built to do — not around what your firm needs to ask across all of them. This fragmentation is one of the reasons many private equity firms struggle to operationalize generative AI beyond isolated use cases.
General partners track opportunities in one system while investor relations teams manage LP relationships in another. Portfolio monitoring occurs through separate spreadsheets, creating information silos that force teams into manual workarounds when stakeholders need unified insights.
That’s not a technology failure. It’s the industry’s natural history. Each of those tools was purchased to solve a specific, bounded problem well. They each do their job. The problem surfaces the moment you try to ask a question that requires more than one of them to answer it.
The Multi-System Assembly Problem
Here’s a scenario that plays out routinely at mid-market PE firms.
A partner wants to answer a straightforward question before a quarterly LP meeting: How is our industrials portfolio tracking against public comparables?
To build that answer, an associate works through something like the following. First, open the market intelligence platform, pull the public comp universe — revenue growth, EBITDA margin, EV/EBITDA — and download it into Excel. Second, go to the portfolio monitoring system and pull each portfolio company’s most recent financials, submitted last month in whatever format each portco CFO preferred. Third, locate the original deal entry assumptions, which means either opening a third spreadsheet or digging through a shared drive folder to calculate performance against the acquisition case. Fourth, build the comparison manually, reconciling accounting differences between the portco’s management accounts and the public comp methodology, and produce a summary that’s clean enough to share.
Three separate systems. Two manual exports. A reconciliation step that requires judgment about which EBITDA add-backs are genuinely comparable across public and private companies. Somewhere between two and six hours of analyst time — to answer one question that should be answerable in minutes.
Multiply that by every LP inquiry in a quarter, every IC update, every operating partner request for a cross-portfolio view, and you start to understand why firms report 40–45% time savings in portfolio monitoring once data has been unified — but those gains only materialize when the data is actually connected. When it isn’t, AI doesn’t reduce the analyst burden. It produces an additional output that still needs to be verified by hand.
Why “Ask Your Vendor for an AI Demo” is the Wrong First Move
According to RSM’s 2025 Middle Market AI Survey — which drew responses from more than 1,000 middle-market executives — 91% of middle-market firms have adopted generative AI, but only one in four has successfully embedded it into core operations. That gap is the most revealing statistic in PE technology right now. It isn’t a skills gap or a change management problem. It’s a data infrastructure gap.
When leadership asks existing SaaS vendors for AI capability demos, they get AI that operates entirely within that vendor’s data perimeter. The AI inside the CRM sees CRM data. The AI inside the portfolio monitoring platform sees portfolio monitoring data. The AI inside the LP reporting tool sees LP reporting data.
None of them can see across the full picture, because no single platform has the full picture. And the full picture is exactly what you need to ask the questions that actually drive decisions:
- Which companies in our current sourcing pipeline have relationships that already exist somewhere in our CRM — contacts we haven’t connected to this opportunity yet?
- How are our healthcare portfolio companies performing against the sector benchmarks defined at underwriting, and which ones are diverging?
- Which portfolio companies are showing early stress signals across both their financial KPIs and the commentary buried in their most recent board decks?
Those questions cross system boundaries. Today, answering them requires an analyst to assemble the data by hand. The AI cannot do it because no one built the connection between the systems. That’s a data infrastructure problem, not a model problem.
The Compounding Cost Nobody Calculates
The manual assembly problem doesn’t just cost time — it costs accuracy. And the accuracy cost is harder to quantify but more consequential.
A single data entry error in one spreadsheet cascades through the IRR and TVPI calculation for the entire fund. Catching it requires going back to the source document and tracing each figure manually. The data collection and normalization phase consumes 60–70% of the reporting cycle at most firms that have not systematized it — and then it repeats identically next quarter.
60 to 70 percent of the reporting cycle spent on collection and normalization. Not analysis. Not the work that sharpens investment judgment or surfaces the insight that changes a portfolio decision. Just locating the data, cleaning it, and reconciling it so the actual work can begin.
The firms furthest ahead operationally haven’t solved this by adding headcount. They’ve solved it by reducing the time between data existing and data being usable — so the people they have spend their time on strategic problems that actually require them.
What the Diligence Process Reveals About Your Data Problem
There is a moment in almost every diligence process that makes the underlying data infrastructure problem visible in miniature.
The data room opens. Documents arrive from multiple advisors, prepared in different formats, with inconsistent naming conventions and varying definitions of the same metrics. The financial statements may show one EBITDA figure; the management presentation shows another, adjusted for a different set of add-backs; the CIM references a third, normalized for a one-time item that isn’t clearly disclosed anywhere. The deal team must not only extract key figures — they must reconcile which version of reality each document is presenting, and decide which version is defensible for underwriting purposes.
Today, junior members of the deal team read through those documents manually. They build extraction templates in Excel. They flag issues in a running tracker updated asynchronously across the team. The process works — in the way many manual processes work — because talented people put in the hours to make it work. Not because the system was built to scale without them.
What AI changes in diligence is not the judgment call about which EBITDA figure is the right one. That still requires an analyst who understands the business and can press management on the discrepancy. What AI changes is the time it takes to surface the discrepancy in the first place. A two-week exclusivity window gives a deal team enough time to thoroughly review the financial statements, key contracts, and management presentation. It does not give them enough time to read every document in a 400-file data room with equal attention. AI can surface which documents warrant the most scrutiny — flagging unusual contract terms, inconsistencies between sections of a CIM, or revenue recognition policies that differ from what management described — before a senior person spends time on them.
AI handles volume and retrieval. Humans handle judgment and decisions. The judgment layer is the job. The volume problem is what’s consuming the time that should go toward it.
The Foundation That Changes the Problem
The firms making AI work in private equity are not the ones who purchased the most sophisticated point solution. They’re the ones who built the data infrastructure layer beneath their existing tools before asking those tools to do anything intelligent.
One of the largest PE firms has built a dedicated data science function that works directly alongside portfolio company management teams on demand forecasting, pricing optimization, and operational analytics. Vista Equity Partners, per Bain’s 2025 Global Private Equity Report, now requires each of its portfolio companies to submit goals and quantified benefits from generative AI initiatives as part of annual operational planning, and runs a GenAI CEO Council so that what works at one portco gets deployed across the rest. These are software-focused programs built over years with significant data infrastructure investment. Mid-market firms are beginning to face the same structural question — just without the same budget. The entry point is identical: unified, clean, queryable data.
The approach is integration architecture, not platform replacement. The individual applications stay in place. The CRM remains the deal team’s workspace. The fund accounting system stays the system of record for portfolio financials. The LP reporting platform continues to manage capital account communications.
What changes is that underneath those tools, a unified data foundation now exists — one that AI can query across all of them simultaneously, with consistent definitions, without requiring an analyst to bridge the gaps by hand first. You don’t need system standardization to get there. Modern integration platforms connect directly to different ERPs, CRMs, and accounting systems across multiple entities, handling data normalization automatically — mapping different chart of account structures to a common model.
The question your AI should be able to answer — which companies in our pipeline match our current target profile, and do we have existing relationships with their management teams anywhere in our CRM? — is a data plumbing question before it’s an AI question. Build the plumbing. The AI already knows what to do with clean, connected data.
Building this foundation is not the end of the work. It’s the beginning of a different kind of work — one where the questions your team has been asking manually for years become answerable in seconds, and where the portfolio visibility problem that fills everyone’s calendar every quarter finally has a structural solution. That’s what Part 2 of this series is about.
Something worth sitting with: If you asked your AI right now to compare your current portfolio against the public comp sets defined at each acquisition and show where each company is tracking above or below the original underwriting case — how much of that answer would depend on an analyst assembling the data by hand first? And what does that tell you about where the real work needs to happen?
