A repeatable deal sourcing workflow: criteria, evidence, review

Most acquisition and investment pipelines are still assembled from memory, conference badges and whatever advisers happen to send over. AI agents can widen that funnel quickly, but a longer list is not the goal. The goal is a list you can defend: every candidate is on it for reasons someone can check, and nobody contacts a founder until a person has looked at those reasons. This guide lays out a sourcing workflow that a corporate development team, a private equity or venture fund, or a search fund can run again and again: a thesis turned into criteria, discovery across several kinds of sources, evidence for each candidate, scoring, and a human review before any outreach.

Turn the thesis into criteria an agent can apply

A thesis such as “founder-owned logistics software in Southeast Europe” is clear to the partner who wrote it and vague to everyone else, an agent included. Before any search runs, rewrite it as inclusion and exclusion criteria that two reviewers could apply to the same company and reach the same answer.

Keep two kinds of criteria apart. Hard criteria are gates: geography, ownership, what the company sells and to whom. A candidate that fails one is out, however attractive the rest looks. Soft signals set priority: hiring, a new product line, a founder nearing retirement, the loss of a large client. When the two are blended into one score, a strong signal can hide a failed gate.

Thesis: B2B software for logistics operators in Türkiye and nearby markets
Include only if ALL hold:
  - Sells software to freight forwarders, carriers or warehouse operators
  - Headquartered in TR, GR, BG or RO
  - Independent or founder-owned (majority owner is not a listed group)
  - Evidence of paying customers (pricing page, named clients, tender awards)
Exclude if ANY holds:
  - Pure consultancy or system integrator
  - Already acquired, or in an announced sale process
Unknown is an allowed answer: record it, never guess

The last line matters more than it looks. An agent asked to fill every column will fill every column, with guesses if necessary. Making “unknown” a legitimate value, and showing it to the reviewer, keeps guesses out of the pipeline. Treat these criteria the way you would treat acceptance criteria for any piece of work: written before the run, and saved as a new version when they change.

Where candidates come from, and what each source misses

No single source covers a market. Use several, and record which ones each candidate came from, because the gaps are systematic rather than random.

Choosing and combining these sources is a subject of its own; see choosing data sources for company research. For deal events such as a new funding round or a completed acquisition, tracking funding rounds and M&A covers the feeds and the deduplication problem.

Evidence for every candidate, not a score for every name

The unit of output is not a company name with a number beside it. It is a row in which every criterion has a value, the source it came from, the passage that supports it and the date it was retrieved. With that in front of them, reviewers can check a candidate instead of researching it again from scratch. A hypothetical row might look like this:

Three rules make a table like this worth reviewing:

• Keep “no evidence found” apart from “the evidence says no”. The first is an open question for the reviewer; the second is a rejection.
• Quote, do not paraphrase. A summary sentence written by the agent is not evidence. The passage it relied on is.
• Record coverage. A list of a few dozen matches says nothing about companies the sources never mentioned. Note which sources were searched, so the next run can widen them on purpose.

How investment firms use AI for sourcing and research

The same pattern appears across firm types, with a different emphasis in each.

• Private equity: add-on acquisitions for an existing platform company, regional roll-up maps, and lists of owner-managed businesses where succession may be on the table.
• Venture capital: thesis-driven scanning for young companies in a category, plus signals such as hiring, product releases and early customer announcements.
• Corporate development: adjacency scans, build-or-buy questions, and watching what competitors acquire.
• Search funds and independent sponsors: narrow, deep lists in one sector and region, where the hard part is finding companies that barely exist online.

Beyond the list itself, agents earn their keep in the research around a deal: a first-pass profile of each shortlisted company, a comparables table, a market map, and later a structured diligence pass. Automated due diligence research covers that later stage. Once a company is in the pipeline, ongoing monitoring of targets and portfolio companies keeps the picture current; see continuous monitoring for investment research.

One caution is specific to investing: confidential material. Teasers, information memoranda and anything received under a non-disclosure agreement should stay inside the firm's own boundary. Before an agent touches that material, know where it is processed, who can read the output, and whether anything is retained.

Scoring and review before anyone is contacted

Fix the weights for soft signals before the first run, and keep hard criteria as gates rather than points. Every candidate that passes then goes to a review queue, where an analyst does one of three things: accepts it into the pipeline, rejects it with a reason, or sends it back for more research on a named open question.

Rejection reasons are data. If reviewers keep turning candidates down for the same reason, that reason belongs in the exclusion criteria of the next version.

Outreach is a different kind of action from research. A first message to a founder cannot be unsent, and it can carry stale facts, the wrong name, or worse, a hint about another company you are looking at. The workflow should draft outreach but never send it by itself. A named person reads the draft against the evidence, edits it and decides. A sample acceptance criterion for the whole workflow:

Every candidate in the review queue has a quoted source for each hard criterion, or an explicit unknown. No message leaves the workspace without approval from a named team member.

Keep the workflow repeatable

• Version the criteria. When the thesis changes, write a new version instead of editing the old one, so you can tell which list answered which question.
• Re-run on a cadence that suits the market, quarterly for a slow sector for example, and deduplicate against your CRM before review.
• Show what changed since the last run: new candidates, candidates that now fail a gate, and evidence that has gone stale.
• Watch for trigger events between runs, such as an ownership change or a new investor, rather than waiting for the next full sweep.
• Keep a run log: criteria version, sources used, who reviewed, what was accepted.

Where GoCux fits

GoCux is not a company database and does not read the open web. Web material arrives only through a browsing or export step that someone has authorized, and it comes in carrying the original page reference and its retrieval date. Inside that boundary, GoCux handles the parts of this workflow that need control.

FindAll takes your matching criteria and the columns you want and returns a table in which each row carries its own evidence. Coverage limits are stated rather than smoothed over, and a count of matches is never treated as proof that the list is complete. Extraction pulls values from material you supply, such as an information memorandum or an annual report, each with the passage behind it, and leaves a missing value as unknown. Sending anything out, outreach included, belongs to the action classes that pause for a person: someone on the team approves, edits or rejects, and that decision stays on the work order's record next to the sources. The finance workflows page shows how this supports evidence-led research.

Checklist before the first run

• The thesis is written as hard criteria (gates) and soft signals (priority), with exclusions.
• “Unknown” is an allowed value, and reviewers can see it.
• Each candidate row carries a quoted passage and a retrieval date for every hard criterion.
• Sources are listed with their known blind spots, and a match count is not read as full coverage.
• Weights were fixed before anyone saw results.
• Every rejection carries a reason that feeds the next criteria version.
• Confidential deal material stays inside the firm's boundary.
• No outreach leaves without a named person's approval.