AI in PE deal origination. What’s actually working in 2026.

The state of adoption.

The numbers are real. 43% of sales organisations had implemented AI solutions by 2024, and 92% planned to expand AI investments in 2025.^[1][2] Within PE specifically, the adoption curve looks similar. The McKinsey research that gets cited everywhere puts AI investment growth in financial services among the highest of any sector.^[3]

What that adoption is actually doing is much narrower than the marketing suggests. There are four use cases producing measurable lift in 2026, and a much longer list that hasn’t.

What’s working.

One — universe mapping and enrichment.

Building a complete target universe for a thesis used to be a multi-week analyst project: web scrapes, license filings, association directories, manual normalisation, contact enrichment. AI tooling has compressed that to days. The lift isn’t in “finding the perfect target.” It’s in covering the entire viable universe instead of the top 200. For lower-mid-market roll-ups in fragmented sectors, that’s the entire game.

Two — signal detection.

Continuous monitoring across thousands of accounts — leadership changes, capacity filings, equipment financing, regulatory activity, hiring patterns — was operationally impossible for a human team. It’s now a baseline. Signal-anchored outreach has been measured to produce reply rates 3–5x higher than template-based outreach in published benchmarks.^[4] That number is the most important AI-driven shift in deal sourcing today.

Three — outbound personalisation at scale.

The promise that “AI writes personalised outreach” was overstated in 2023. The actual operational use in 2026 is narrower and more reliable: pulling 3–5 specific, public, attributable facts about a target company and incorporating them into a templated structure that a human operator reviews. Not autonomous content generation. A research-and-assemble function that scales human-quality first messages to thousands of targets.

Four — AI dialing.

Voice agents that handle initial outreach calls, identify decision-makers, qualify intent, and either escalate to a human or schedule follow-up have moved from prototype to production over the last 24 months. The compliance landscape around AI voice is real (see TCPA considerations), but operated within the regulatory framework, the throughput improvement is substantial — particularly for the “canvas pass” phase of sourcing where most of the work is identifying who’s in play.

What’s not working.

Fully-autonomous outbound.

Agents that source, write, send and close conversations without human supervision do not produce above-baseline results in PE deal origination. The judgement layer matters. The cost of a low-quality conversation with a target founder — in terms of reputation, deliverability and future relationship damage — is high enough that autonomous outreach is structurally worse than supervised outreach.

Generic “intent data” products.

The intent-data industry sells signals derived from web behaviour aggregated across publishers. For consumer-facing B2B SaaS, this can work. For PE add-on sourcing — where the relevant signals are public filings, leadership changes and capacity events — generic intent data is mostly noise. The signal categories that matter are specific to the sector and require custom monitoring infrastructure.

LLM-generated “personalisation.”

Out-of-the-box LLM personalisation — asking a model to write a custom intro to each prospect — produces text that reads as AI-generated to any sophisticated reader. Founders and CFOs at lower-mid-market businesses are sophisticated readers of outbound. The reply rates on unconstrained LLM personalisation are below template baseline in most measured campaigns.

Predictive deal-likelihood scoring.

Models that predict “this company is X% likely to sell in the next 12 months” have proven hard to operationalise. The base rate is so low (a 5% annual base rate of transition is high for most sectors) that even a well-calibrated model produces too many false positives to drive sourcing decisions. The teams using this kind of scoring tend to treat it as a tie-breaker rather than a primary driver.

The actual pattern.

The AI tooling that produces real lift in deal origination is the boring kind. Better universe coverage. Continuous monitoring of public signals. Mechanical personalisation that incorporates 3–5 attributable facts. Voice automation operated inside compliance frameworks. None of these are particularly impressive in a demo. All of them compound to a sourcing function that operates at 10x the throughput of a human team without sacrificing quality.

The firms doing this well in 2026 don’t have an “AI strategy.” They have a sourcing infrastructure that uses AI in the places it produces measurable output, and human operators in the places it doesn’t. The distinction matters because it determines whether AI investment shows up in deployment results or in the slide deck.^[5]