Nevin Raj, general manager at Grata, now part of Datasite, recently joined ACG Insights host Carolyn Vallejo to discuss where AI is working in deal sourcing and where it quietly fails. If you run a lower-middle-market fund or an M&A advisory shop, the conversation is worth your time.
Listen to the full ACG Insights episode: Addressing AI's Deal Sourcing Blind Spots
Adoption Is Not the Problem Anymore
Vallejo notes that recent estimates put AI adoption above 80% or 90% among middle-market and lower-middle-market private equity firms. Raj's experience supports the broader point: it is increasingly difficult to find deal professionals who are not using ChatGPT or Claude in their day-to-day work.
The question at investment committee is no longer, “Should we use this?” It is, “Why did the associate's target list miss three of the companies our competitor just closed on?”
AI is earning its keep across the business-development workflow: generating theses from market trends, mapping investable companies inside a sector, drafting personalized outreach, transcribing calls, entering CRM notes, and preparing follow-up.
The Three Blind Spots
The model makes things up when the data is not there
Contact information is the classic case. Ask a general-purpose model for the CFO's email at a 40-person industrial-services company and it may invent one. Raj's own example is telling: Grata was acquired, yet the company still receives inquiries about participating in its Series A. The model's picture of the world is stale, and it may not recognize that it is stale.
Models also favor what is most visible. The under-the-radar, founder-owned company with little press and a six-year-old website may be exactly the target you want—and exactly the one a general web search never surfaces.
One-shot prompts produce robotic output
The first outreach draft is often generic. Refining it takes iteration, and probabilistic systems can return different answers to the same prompt. That is not fatal at small volume. It is dangerous when the team scales before the workflow is tuned and reviewed.
The downside is not symmetrical
The light version is a missed deal and a thin pipeline slide at investment committee. The moderate version is reputational erosion in a sector you are trying to roll up. The severe version is your firm's email domain being flagged as spam, affecting ordinary business communication as well as sourcing.
What Actually Fixes It
Connect the model to trusted data
The mechanism is MCP, often surfaced in AI tools as connectors. Instead of asking the model to generate a company list from memory, the workflow routes the request to a curated source such as Grata, PitchBook, or your own CRM and reasons over the returned records. Precision now lives in both the data and the instructions.
Benchmark against a truth set
Take a dataset you trust—funding events, closed transactions, or a manually verified target universe—and compare the AI output against it. Errors often cluster around recent events and sparse private-company data. A truth set shows you where the human check belongs.
Measure conversion, not activity
Track emails sent, calls made, meetings booked, meetings held, and then NDA, IOI, LOI, and close. Early in the process, higher activity and lower conversion may be normal. Scale only as conversion improves through the funnel.
The metric chain that matters
Start with one agent and one human
Do not spin up five agents at once. Pair one agent with one operator, prove that the process converts, and expand only after the controls and benchmarks work.
A Useful Proof Point—with an Important Caveat
The episode highlights Ridgefield Partners, a transportation-focused M&A advisory serving businesses with $1 million to $25 million of EBITDA. According to Grata's account, Ridgefield began with roughly 4,000 CRM contacts supported by ZoomInfo and HubSpot. After connecting Grata data through MCP, the firm's reach expanded to approximately 70,000 companies in under six months, while it moved into adjacent sectors including healthcare.
That demonstrates a material increase in sourcing coverage. It should not be confused with independent proof of conversion. The episode does not provide Ridgefield's meeting, IOI, LOI, or close rates—the downstream measures Raj himself recommends using before scaling.
What This Means for a 10- to 50-Person Firm
Raj closes with the point that matters most: deals are still done by people. AI accelerates relationships and judgment; it does not replace them. The firms getting the most value are treating AI as a plumbing project first and a prompting project second.
That plumbing is where PE Tech Partners works. For funds and advisory firms in this range, the operating model comes down to three priorities:
- Connect the model to the system of record. Integrate the data sources and CRM platforms your team already pays for and trusts so the model stops guessing.
- Build the agentic workflow in phases. Start with one operator and one stage of the business-development funnel, with a conversion benchmark defined before the first scaled campaign.
- Instrument the funnel for outcomes. If the investment committee deck says “AI sent 8,000 emails” but cannot show how AI-sourced meetings converted to IOI, the program is not yet manageable.