How Private Equity Firms Run on AI

Private equity firms and dealmakers run on AI across the deal lifecycle. Proprietary sourcing engines scan millions of companies to surface targets, generative-AI diligence compresses weeks of document review into hours, and portfolio value-creation programs push AI inside the companies they own. Increasingly this happens through both in-house data-science teams and specialized deal-tech platforms like Hebbia, Ontra, and AlphaSense. The firms getting a real return did not buy a strategy. They put one tool on one stage of the process and kept a partner on the judgment.

If you sit on a deal team, run a fund, or advise the people who do, you have heard two years of noise about AI in private capital. Some of it is real. A lot of it is a vendor describing a future that has not shipped. This piece is narrower. It looks at named firms and platforms, what they use, and what they have reported, with a source for each, and it groups the work the way a deal actually moves: find it, check it, then build value in it.

Where AI sits in the deal lifecycle: use case, what AI does, an example, and the caveat

Use case	What the AI does	Example firm	Caveat
Deal sourcing	Scans millions of companies and signals to surface targets	EQT (Motherbrain)	Surfaces, does not decide; a partner still runs diligence
Due diligence	Reads data rooms and filings, answers questions, drafts memos	Hebbia, AlphaSense, BlackRock (eFront Copilot)	Output needs review; confidential data needs boundaries
Firmwide diligence speed	Assistants let investors assess a company faster	Carlyle Group	Reported speed, not audited; humans still sign the thesis
Portfolio value creation	Embeds data scientists and AI tools inside portfolio companies	Blackstone (BXDS), Hg (Catalyst), Vista	Impact figures are firm-reported estimates
Legal and back office	Automates routine private-markets contracts and entity work	Ontra	Standard agreements only; bespoke terms stay with counsel
Banking support	Firmwide copilot summarizes, analyzes, drafts for bankers	Goldman Sachs	A productivity layer, not a dealmaker

---

Stage 1 · Deal sourcing

EQT: a sourcing engine that found 15 of its own deals

EQT built Motherbrain in 2016 as a dedicated team and platform to bring data and machine learning into the top of the funnel. The system combines external signals, from web and social activity to hiring, with EQT's own historical diligence records, and uses that pool to flag companies a human scout might never reach.

The specific claim is the one that matters. Motherbrain is credited with sourcing 15 investments, among them Peakon, AnyDesk, and CodeSandbox, with Peakon later acquired by Workday. That is not a machine picking winners. It is a machine widening coverage so the investment team spends its attention on a better-filtered set of companies. EQT is direct that the point is technology amplifying human expertise, not replacing the decision. Source: EQT and the AI Expert Network case study on Motherbrain.

The reason this leads the piece is that sourcing is where AI earns the clearest edge in private capital. There are more companies than any team can track, and the cost of missing a good one is high. An engine that reads the long tail and surfaces the few worth a call is a real advantage, as long as a partner still runs the diligence behind each name.

Stage 2 · Due diligence and research

Hebbia: agentic diligence over the whole data room

Hebbia builds Matrix, an agentic research tool that ingests very long files and returns answers in a structured, spreadsheet-style format. In finance terms, it reads SEC filings and other documents to organize and compare a company against its competitors, which is the grinding core of early diligence.

The market signal is loud. Hebbia raised a 130 million dollar Series B at a 700 million dollar valuation, led by Andreessen Horowitz, on 13 million dollars of annual recurring revenue and, unusually for the category, actual profitability. Named customers include the investment bank Centerview Partners and the private equity firm Charlesbank. The tool does not run the deal. It reads the room faster than a team of associates and lays the facts out for review.

The honest read: agentic diligence still needs a professional to weigh what it surfaces, and any tool touching a live data room needs a confidentiality posture set in advance. Hebbia is a strong example of where the work is going, not a reason to trust an unread answer. Source: TechCrunch.

AlphaSense: research across half a billion documents

AlphaSense is a market-intelligence platform that runs generative search over 500 million-plus premium documents, including earnings transcripts, SEC filings, expert-call notes, and broker research, and layers in a firm's own internal content. For a diligence team, it collapses the scattered-source problem into one searchable surface with summaries and alerts.

The penetration numbers stand in for a customer count. AlphaSense reports use by 80 percent of the top asset-management firms and 88 percent of the S&P 100. That reach is the point for a dealmaker: the platform is already the research layer at a large share of the institutions on the other side of the table. It speeds the reading and the monitoring; the analyst still forms the view. Source: AlphaSense.

BlackRock: a diligence copilot built for private markets

BlackRock added generative AI to its core platforms through Aladdin Copilot and eFront Copilot, the latter aimed squarely at private markets. The eFront tool lets users pull quick analytics and visualize risk and performance factors, while the broader Aladdin Copilot surfaces answers across the platform to support faster decisions.

The detail worth copying is the guardrail, not the feature list. eFront Copilot is designed with content filtering and hallucination controls, and it will not give investment advice or answer questions outside the Aladdin platform boundary. That is what a serious diligence tool looks like: it retrieves from an approved surface and refuses to wander. For a category that runs on material, non-public information, the boundary is the product. Source: Microsoft and BlackRock.

Carlyle: diligence measured in hours, not weeks

The Carlyle Group offers the cleanest read on how far this has spread inside a large firm. Chief innovation officer Lucia Soares reports that 90 percent of Carlyle's employees use tools like ChatGPT, Perplexity, and Copilot in their work, which is a very different picture from a small innovation team running pilots.

The outcome she points to is the one deal professionals care about. Carlyle's credit investors can now assess a company in hours rather than weeks, because the AI does the reading and organizing that used to consume the front end of the analysis. That is speed on the least differentiated part of the job, which frees time for the parts that actually separate one bidder from another. As with every case here, the figure is the firm's own, and the investment thesis is still written and defended by people. Source: Forbes.

Stage 3 · Portfolio value creation

Blackstone: a data-science team inside the portfolio

Blackstone runs a dedicated data-science group that works alongside its portfolio companies on AI strategy and integration rather than only advising the fund. The firm frames this as operational, not experimental. Its people sit with management teams and build.

The number Blackstone puts on it is an estimated 200 million dollars of bottom-line impact from that work, with named examples at portfolio companies including Link Logistics and Signature Aviation. Read it as a house view of what AI is worth when it is pushed down into the operating companies, not an audited figure. The lesson for a smaller sponsor is the model, not the headcount: value creation is now a place where AI shows up on the P and L, not just in the deck. Source: Blackstone.

Hg: an in-house AI incubator across 60-plus software companies

Hg, the European software investor, built its AI capability as an internal team rather than a consulting line item. It runs 100-plus AI professionals, and its Catalyst incubator brings together more than 80 engineers, product managers, and designers who embed directly into portfolio businesses out of London and New York.

The activity numbers are the striking part. Hg reports 1,600-plus live AI initiatives across the portfolio, 100-plus agentic products and features shipped, and roughly 260 million dollars of budgeted EBITDA impact from AI. For a software-focused fund this is the logical shape: treat AI as a product-and-engineering program you install in every company, staffed by people who ship, and hold it to a budgeted number. Source: Hg.

Vista Equity Partners: an Agentic AI Factory for the software portfolio

Vista built what it calls an Agentic AI Factory, a platform meant to scale AI agents across its enterprise-software companies. The approach rearchitects workflows around AI, deploys multiple agents per user, and integrates go-to-market through hyperscaler partnerships.

Vista's own framing is ambitious: it projects 5 to 10 AI agents per user and, across the full portfolio, a potential 4 to 8 billion autonomous agents. Those are projections, not a current count, and worth reading as intent. The concrete example is Gainsight, a Vista company deploying agents that handle customer renewals, spot upsell opportunities, answer product questions, and trigger follow-ups. This is the value-creation play at software scale, with the honest caveat that the biggest numbers here describe the plan rather than the present. Source: Vista Equity Partners.

Supporting infrastructure

Ontra: the routine legal work, automated

A large share of a deal professional's week is not analysis. It is legal process. Ontra automates that layer for private capital, handling routine contracts, negotiation on standard agreements, and entity management. The founder started the company after estimating he spent 15 to 20 percent of his day on exactly this kind of repetitive legal work.

The customer list is the proof of adoption. Ontra serves roughly 850 customers, including Blackstone, Warburg Pincus, and Motive Partners, and has processed more than 1.5 million documents. The caveat is built into the product: this is for the standardized, high-volume agreements, and bespoke or contentious terms still belong with counsel. Automating the routine legal grind is one of the least glamorous and most reliable AI wins in the whole lifecycle. Source: Fortune.

Goldman Sachs: a firmwide copilot for the bankers next to the deal

Private equity does not close deals alone, so the investment banks matter here too. Goldman Sachs rolled out its GS AI Assistant to 10,000 employees in early 2025, then expanded access across the full workforce of roughly 46,500. Bankers, research analysts, and developers each get a version tuned to their work.

The assistant summarizes documents, analyzes data, drafts content, and translates materials, the connective tissue around a live process. Goldman reports efficiency gains for engineers on internal tools and faster handling of routine work. It is a productivity layer, not a dealmaker, and that framing is exactly right. The advisor on the other side of your deal is now backed by a copilot, which is one more reason to understand what these tools do and do not do. Source: Fox Business and the firm's internal memo as reported.

The playbook, in the firms' own words

KKR: writing down how to run generative AI across the lifecycle

KKR published its own framework for implementing generative AI across the investment lifecycle, which is useful precisely because it is a large sponsor describing method rather than selling a tool. The through-line is disciplined adoption: apply AI where it creates tangible operational improvement, and resist chasing hype for its own sake.

It is worth being precise about what this is. It is a thought-leadership framework, not a claim of a single audited internal deployment, and KKR is separately one of the largest investors in the physical infrastructure, power and data centers, that AI itself runs on. Both facts point the same way. The firms with the most at stake are treating AI as a discipline to be governed, not a gadget to be bought. That is the posture to copy. Source: KKR.

---

What still does not work, and what every serious desk kept human

Eleven strong examples can read like a straight line, so here is the other half.

Most of these numbers are self-reported. The 300-hour figures, the impact estimates, the projected agent counts, they come from the firms and their vendors, not an independent audit. They are credible and often attributed to named executives, but they are the cases the firms chose to publish. Treat them as what is possible, not what is promised on your process.

Confidentiality is the whole game. Private capital runs on material, non-public information and privileged legal work. The firms doing this well use enterprise platforms with hard data boundaries, like BlackRock's eFront Copilot refusing to answer outside its platform, rather than pasting a data room into a public chatbot. If you take one rule from this page, take that one, and write it down before anyone touches a live deal.

The frontier claims are still claims. Vista's billions of agents and the autonomous end of the market describe a direction, not a delivered state. The safe posture today is AI that reads, drafts, and flags, with a professional who reviews, decides, and signs.

It rewards structure, not gadgets. The desks getting real returns, from EQT's sourcing engine to Hg's embedded teams, built a program and a method around the tool. Firms that buy a subscription and hope get marginal wins at best. The gap between the two is process and training.

How to think about your first step

If these firms point at one move, it is this. Do not start with an AI strategy, start with one stage of your own process. Pick the single most repetitive, highest-volume task on a deal, usually reading a data room, building a first comparison, or drafting a standard memo, and point one reviewed tool at it. Keep yourself on the output and measure the hours over a full cycle.

Then write your confidentiality and information-barrier rule before anyone loads a document, decide what may enter a tool and what may never leave the building, train the team on that one workflow, and only add a second task once the first is boring. That is the unglamorous order behind every firm on this page, whether they said so or not.