🤖 AI in Private Equity: The Holdout Is Over

Hi all - you may have noticed we have not shared any new deals over the past couple of weeks. Simply put, nothing that has crossed my desk recently has met our bar. Maintaining a high signal-to-noise ratio is core to what we are building at Road To Carry.

Behind the scenes, we are investing heavily in initiatives like the Serial Acquirer Conference, the Roll Up Bootcamp, and several early-stage projects that will bring even more value to the community.

In the meantime, if you are seeing compelling opportunities, please feel free to send them my way. Thank you for your continued support—we just crossed 8k readers! 🎉

In the world of PE, where perfection is expected on the first go, AI adoption runs counter to the industry's core culture. But according to Ed, that mindset is dangerous. Models improve rapidly. The competitive advantage comes from iteration and learning, not one-time deployment. The firms willing to experiment are quietly pulling ahead of the ones waiting for a perfect answer.

Edwin: What's the one misconception about AI that could lead PE firms to make the wrong decisions?

Ed: The single biggest mistake is what I'd call the one-and-done. You try it once — whether that's ChatGPT, Claude, or a vendor solution — you don't get the answer you expected, maybe you get something that looks hallucinated, and you conclude AI doesn't work. You put it on the shelf.

The problem is twofold. First, that's just a bad way to approach any new technology. Second, the models are continuously improving and the vendor landscape is continuously changing. If you write AI off after one bad experience, you're not leveling up the skill set of your team. And that puts you behind competitors who are willing to iterate.

PE professionals have a very low tolerance for inaccuracy and a very high expectation of quality. That's a feature in almost every other context. With AI, it's a liability — because the tool rewards iteration, not perfection on the first pass.

Edwin: You had a great career at KKR, retired, and then came back to build an AI company for PE. What did you see?

Ed: I'll be honest — I didn't see it myself. I was five years into visiting national parks. My oldest son Matt called me up and said he thought generative AI was going to change the way everyone works. My first instinct was to talk about my Utah trip.

But we started brainstorming. And we kept coming back to the same problem: diligence.

In private equity, diligence is not a classic data problem. It's a document problem. The companies being acquired — especially in the lower middle market — have decades of knowledge locked inside PDFs, contracts, financial statements, and management presentations. It's messy. It's unstructured. Nobody's cleaning it up.

What we zeroed in on was document ingestion and document analysis — how you make sense of all of that and merge it with expert network calls, research, regulatory filings, and everything else that goes into a proper diligence process. It turns out that's a genuinely hard problem to solve. Early ChatGPT had about 8,000 tokens. Now we're talking 200,000, a million, 2 million. But even with massive context windows, the challenge in a real data room is tens of millions of tokens, spread across hundreds of documents that don't necessarily talk to each other.

That's what convinced me to come out of retirement.

Edwin: A year ago, AI usage at PE firms was basically associates using ChatGPT to clean up emails. Now, many firms already use tools like ChatGPT Enterprise. Why is that not sufficient and what are you doing differently at ToltIQ?

Ed: When you use ChatGPT Enterprise or Claude Enterprise, you're using a multimodal model in its native form. That model is brilliant across a huge number of tasks. But there's a trade-off — especially for business documents.

Think about a credit agreement with multiple amendments. A CIM with embedded charts. A stacked bar chart with three different shades of blue that you can read as a human but a model might not differentiate. A legal document where the key clause is buried in a footnote on page 147. General models are optimized for speed and breadth. They're not optimized to handle the complexity of how knowledge is actually organized in PE documents.

What we do differently: we ingest the actual documents, deconstruct them into their component parts, and build what's called a vector-based knowledge structure from embeddings. So when you ask a question, we retrieve the precise slice of knowledge that's relevant, and feed that curated subset into the model. Every answer comes with citations — down to the paragraph, the table row, the bullet point.

There's also a security dimension most people gloss over. ChatGPT and Claude are not under NDA with you. When you're loading deal documents, that matters. We sign NDAs with every client. Your documents sit in your own segregated environment in Amazon. That's a fundamentally different posture.

We're also model-agnostic. We run on OpenAI, Anthropic, Google, and Cohere — different parts of our stack use different models based on their capabilities. Because it's an arms race. Who the leader is today may not be the leader in 18 months. We've seen that flip three or four times in the last year and a half alone.

Edwin: Are you seeing faster adoption in smaller funds or larger firms?

Ed: The conventional wisdom would say the large caps are furthest ahead. I'd push back on that.

The mega funds — KKR, Blackstone, Carlyle, the top 15 or 20 firms — are largely trying to build their own AI infrastructure. They have the balance sheets and the engineering teams to attempt that. But it's also produced slower adoption. There's organizational inertia at scale.

The mid-market and lower mid-market are actually moving faster — partly out of necessity. When you've got a 15-person investment team, a 2-3x productivity multiplier on your associates isn't a nice-to-have. It's how you compete with firms that have 150 people. You're handicapped on headcount and you know it.

The biggest near-term bang for the buck is in document-intensive work — everything from sourcing through close. That's where the value is clearest. The structured data problems — carry calculations, IRR, LP reporting — are harder to crack and the tolerance for error is zero. Those will come. But they're not where AI is winning today.

Edwin: What does AI actually mean for the people working in private equity?

Ed: My honest view is that by the end of 2027, we are in a meaningfully more complicated workforce world in PE.

2025 was the experimentation year. 2026 is the year firms are deploying AI on real problems in real workflows — diligence, sourcing, marketing, operations. Inside ToltIQ, we're a 30-person firm and we'll probably never be more than 50, because the AI leverage across sales, marketing, and engineering is giving us a 2-3x multiplier off those resources.

The org chart in PE won't disappear. But it will flatten on the sides. You won't need to keep adding associates proportionally to deal volume. Associates will cover more ground. The nature of the work will shift.

The skill that matters most in this environment isn't just knowing how to prompt a model — it's being genuinely inquisitive. Taking the adversarial position. Knowing when to push back on the output rather than accepting it.

Those are the skills that separate good investment professionals from great ones. AI doesn't make them less important. It makes them more important — because the cost of not having them is higher.

Edwin: A lot of people in this community feel anxious about AI and don't know where to begin. What's your advice?

Ed: Start outside of work.

You're all making enough in private equity to afford a $20 or $200 a month personal license for ChatGPT or Claude. Get one. Download investor presentations for public companies you follow and try to assess what's changed. Pull earnings transcripts. Analyze something you care about. Take a picture of the food on your plate and see how well the model identifies it. Plan a vacation and have a back-and-forth conversation with it.

The anxiety you feel at work comes from the high-stakes environment. Lower the stakes first. Get comfortable with the iterative nature of working with these tools — that you can ask a question five times, refine the prompt, and get progressively better output. That mindset is the direct opposite of how most PE professionals are trained, and it's the thing most in need of changing.

One specific resource: follow Ethan Mollick out of Wharton. He posts consistently thoughtful content on AI, was an early adopter, and wrote a book called CoIntelligence that's a genuinely useful framework for thinking about where this is going. Parts of it are slightly dated now — which is itself the point. The pace of change is real.

Edwin: Is AI becoming part of LP diligence conversations?

Ed: Yes, increasingly so.

I’ve sat in on several annual meetings this past year. LPs are asking how AI is being used in sourcing, diligence, and portfolio support. They are also exploring similar tools themselves as they manage large volumes of GP reporting data.

Over time, AI may become a standard part of how both GPs and LPs evaluate investment opportunities.

Any topics I should cover next? Share thoughts with [email protected]
Were you forwarded this newsletter? Subscribe here.

Make sure you receive Road To Carry every Thursday and Saturday!

Everyone: reply to this email with a “hi”

Gmail mobile: Click the 3 dots (...) at the top right corner, then "Move," then "Primary"

Gmail desktop: Go back to your inbox and move this email to the "Primary" tab

Other users: Follow these instructions