Every discussion about AI in private equity lands in the same place. One side imagines fully automated diligence. The other admits most AI pilots never get beyond a proof of concept. Both are true, and neither side is talking about the actual bottleneck.
After working with hundreds of private equity clients at TresVista, we keep seeing the same pattern. A firm rolls out ChatGPT or Claude, analysts start using it for small tasks, productivity gets a temporary lift, then adoption stalls. The stall isn’t where most firms think it is.
That distinction matters because most conversations focus on what AI can do, not what firms are prepared to support. It shows up first in how firms talk themselves out of starting at all.
The cost of waiting
A lot of mid-market firms are still sitting on the sidelines, and there are good reasons for that. Compliance teams worry about confidential information. Legal teams worry about governance. In private equity, a mistake involving sensitive data is expensive. These concerns call for governance, not avoidance, with clear limits on what AI can see and decide. But that caution has been backed up by numbers that make waiting look like the safer bet.
MIT’s Project NANDA, in its 2025 report The GenAI Divide, found that despite $30-40 billion in enterprise AI spending, around 95% of organizations still aren’t seeing measurable financial returns. It’s an easy statistic to point to when the safest decision is to wait. Waiting, though, doesn’t make the workload any lighter.
Deal teams are handling more information than they were a few years ago, hiring has slowed, and LPs increasingly expect firms to explain how their AI investing decisions are paying off. Sourcing and diligence are usually the first places those expectations show up.
Sourcing still depends on judgment
AI is excellent at sorting information. It can scan thousands of companies, organize messy datasets, identify businesses that match an investment thesis, and summarize results in minutes. That’s valuable work, but the decision about whether a company is worth pursuing still belongs to the investment team, for reasons no dataset fully captures.
No model knows which management team will become difficult during diligence, or whether recent growth came from a sustainable customer base versus a single contract that happened to renew at the right time. Those calls depend on judgment built from doing this before, so the strongest sourcing projects combine automation with review.
Here’s what that looked like in one recent TresVista engagement: building a database of compounding pharmacies from state-level records spread across PDFs, spreadsheets, downloadable files, and other inconsistent sources. Scripts extracted the data, one model
structured it, another classified it, and a human review process checked everything against external references. A project that would have taken roughly 500 manual hours got done in about a fifth of that time.
That’s a much more realistic picture of AI in sourcing. It removes repetitive work, so analysts spend more time making decisions. The same pattern carries into diligence, just starting from a messier baseline.
Diligence improves in layers
Due diligence is probably where expectations get ahead of reality. Anyone who’s opened a financial model that’s been passed between investment banks for several years knows how messy those files become. Nested formulas, broken links, overlapping assumptions, hidden tabs, and inconsistent formatting aren’t problems AI can magically clean up. Research, though, is a different story.
Once the underlying structure exists, AI can review filings, industry reports, market research, and company materials at the same time instead of forcing analysts through each source one by one. The gains show up, but they build gradually rather than arrive all at once.
That layered improvement only holds up, though, if the data underneath it is reliable once the deal closes and the file moves from diligence into monitoring.
Portfolio monitoring depends on the data first
Portfolio monitoring probably has the biggest long-term opportunity, but it’s also where weak data becomes impossible to hide. An AI model can’t answer questions that no system has captured. Customer profitability, operating metrics, working capital trends, or product- level performance all depend on structured, reliable data before automation can add much value.
That’s usually the gap we see first when a firm asks about monitoring: not which model to use, but whether portfolio companies are even reporting the same numbers the same way. Building that structure is unglamorous work, but it’s what makes automation possible later. Investor reporting is what that later stage looks like once the foundation is already in place.
Investor reporting follows the same principle
K-1 reconciliation has traditionally taken about 25 to 30 minutes per investor across funds with roughly 200 investors. In one TresVista engagement, Python automation with built-in validation cut that down to about one minute per K-1, roughly a 90% reduction in manual effort. The same shift showed up elsewhere too.
PCAP reviews followed a similar pattern. Manual teams often reviewed only about 20% of statements because reviewing every file wasn’t practical; automation made full coverage possible without adding proportional effort. Same lesson, different task.
Neither improvement started with AI. Both started with usable data, and that ordering is exactly what separates firms whose pilots take off from firms whose pilots stall.
Why pilots stall
After enough projects, the reasons AI fails in private equity usually come down to a few predictable problems. Many firms still don’t have clean data. Others buy software without changing the process around it. Success is often measured by licenses issued or prompts submitted instead of hours saved or turnaround time. The firms that avoid these traps look different from the outset.
The firms seeing meaningful returns aren’t chasing every new model release. They’re investing in the less glamorous work first: data quality, governance, repeatable workflows, and review processes. That’s the layer where we spend most of our time with clients, not picking a model but building the workflow underneath it.
At TresVista, that means pairing our teams’ deal experience with the automation layer — structuring messy data, building the review workflows, and running the reconciliation and reporting pipelines so the model has something reliable to work from. It’s less about deploying AI and more about making your firm ready for it.
If you’re trying to figure out where AI belongs inside your firm, don’t start with the tool. Start with the workflows that already run on clean, structured data. Those are usually the first places where AI earns its place. If you’re mapping where AI belongs in your own workflows, get in touch. We’re glad to walk through what we’ve learned.
