Private Equity Has Trust Issues With AI

This article was written by Lalit Lal, Co-founder & CTO of Keye, the first AI platform built exclusively for private equity. Keye understands the data, performs real analyses, and mirrors the workflows of top-performing deal teams. For more information, visit https://www.keye.co/

Private equity is caught in a weird place with regard to AI. Every firm sees the upside; faster due diligence, better insights, scaling value creation across portfolios. But most firms are still just dipping their toes instead of actually using it.

According to Bain research, most portfolio companies are testing AI, but only 20% have actually put it to work in ways that matter. McKinsey found pretty much the same thing. They found that 78% of companies use AI somewhere, but barely 1% of executives would say their deployments are actually mature.

This clearly shows an excess of expensive experiments, and not a lot of evidence of actual results.

That said, the hesitation makes sense given accuracybut sitting on the sidelines is getting risky.

Why Everyone's Stuck

So why isn't this moving faster? Because in high-stakes finance, "pretty good" doesn't cut it. Investment committees need what you might call bulletproof evidence. Stuff that holds up under scrutiny. Numbers that are 100% right. Analysis that explains why something's happening, not just what's happening. And AI models, when you let them run wild, don't always deliver that.

A startup called Patronus AI recently tested regular AI setups on financial questions and found they delivered inaccurate information up 81% of the time. Even the fancy long-context models still failed about 25% of the time. That's nowhere near good enough for an IC presentation without serious safeguards. The reputation risk can be brutal. One made-up covenant term, one wrong revenue number, one bogus guidance figure and your credibility takes a hit that lasts way longer than any single deal.

So, naturally, firms are hesitant. The CFA Institute found that 68% of finance professionals are curious about AI, but 60% are nervous about it, with almost half saying there's pushback at their firms.

How to Fix the Trust Problem

This is the dangerous spot everyone's in right now. Playing it safe feels smart, but it's becoming a competitive problem. If your competitors can run reliable AI scans through data rooms, summarize document changes in minutes instead of hours, and stress-test business models with solid, traceable outputs faster than your team can even get organized, then your "being careful" just became code for "moving slower."

So how do you solve this without breaking compliance rules? Treat AI like you'd treat any other operational risk — manage it properly instead of just avoiding it.

• Don't Add Generative AI Where It Doesn't Need to Be: not every process needs AI. If your existing workflows are already fast and accurate, don't fix what isn't broken. Focus AI on the stuff that's actually painful — sifting through massive data rooms, comparing complex documents, or pulling insights from unstructured data. Smart deployment beats wide deployment every time.
• Set Clear Standards for What You'll Accept: write down exactly what counts as good enough for your IC. Source citations you can check, numbers that tie back to real documents, prompts you can reproduce, confidence scores that make sense. If your team can't verify it, don't use it. NIST's AI risk guidelines give you a decent place to start.
• Keep Your Models on a Tight Leash: only let your AI pull from sources you trust (SEC filings, audited statements, QoE reports) with strict rules about citations and number-checking. Letting AI grab stuff from the open web for investment decisions is like using random internet financials for deal modeling. Just don't.
• Test Everything Before It Goes Live: before any AI touches real due diligence, test it on finance-specific tasks. Number crunching, policy extraction, document comparison. Set standards for how accurate it needs to be. Only deploy what passes your tests, and keep records of everything like you would for any valuation model.
• Keep Humans in the Loop Where It Matters: all IC materials need a human to sign off. Track every edit, every escalation, every time someone overrides the AI. Use that feedback to make your systems better. Turn it into a learning process, not just a one-time setup.
• Build Systems That Can Say "I Don't Know": pick AI that can admit when it's not sure. Systems that say "not enough information" or "sources don't agree" are way better than ones that confidently make stuff up. When the stakes are high, you want conservative systems that tell you when they're confused, not confident ones that hallucinate.
• Watch These Systems Like a Hawk: monitor AI like you would any critical system. Track when it starts drifting, what kinds of errors pop up, how fast it's running. When something goes wrong, do a proper post-mortem and fix it. AI failures should get the same attention as trading errors.
• Treat Vendors Like Any Other Financial Service: require all the usual stuff, like SOC 2 compliance, clear data policies, audit trails, proper handling of confidential information. Map these controls back to your existing risk management and regulatory requirements.
• Start With Safe Wins: begin with applications where you can still double-check everything. Comparing documents across regulatory filings, pulling policies out of credit agreements, summarizing portfolio performance with clear links back to the data. These give you speed wins without giving up the ability to verify the work.
• Measure Real Business Impact: track business results, not just cool tech demos. How much time you're saving on specific deliverables, how many errors show up in final outputs, how often humans have to step in. Then connect those numbers to deal cycle times and portfolio performance to see what AI's actually worth.

The Window Won't Stay Open Forever

This is only gonna get more intense. As more firms deploy reliable AI workflows, LPs are gonna expect the better analysis and faster turnaround that only AI-assisted teams can deliver. The firms that figure this out now will have real advantages in deal sourcing, due diligence speed, and portfolio value creation.

The opportunity won't last forever. Just like electronic trading or algorithmic portfolio management, the first movers will build advantages that keep building on themselves. The question isn't whether AI will change how private equity works; it's whether you'll be leading that change or scrambling to catch up.