AI Deal Matching That Screens Companies Against Every Investor Mandate

#The starting position

The firm sources acquisition targets for investors. Each investor has a mandate: sectors they want and sectors they will not touch, revenue and size bands, geographies, ownership structures and a long list of softer preferences that usually live in someone's head. The firm's job is to find companies that fit, explain why, and put the best ones in front of the right investor before a competitor does.

That job was being done by hand. Analysts read company websites, compared what they found against several mandates at once, and kept the results in spreadsheets and email. The work did not scale with the number of investors, because every new mandate multiplied the comparisons. It was also hard to audit: when an investor asked why a company had been proposed, or why a similar one had not, the reasoning was scattered across inboxes.

The firm wanted two things at once. It wanted the speed of AI research, and it wanted control: every recommendation explained, every approval recorded, and the whole path from research to deal visible in the CRM.

#What we built

01 An investor and deal model the AI can read

We started with the data, because a model can only match against criteria it can read. Investors, acquisition mandates, contacts, a deal queue and pipelines were modelled in Sales Cloud, with website-to-Salesforce intake for inbound opportunities. Investor likes and dislikes became structured fields, and mandate documents stored in Google Drive were connected to the investor record, so the criteria are available to both analysts and the AI.

02 The AI Matching workspace

On top of that model we built a matching workspace in Lightning Web Components. Candidates are first ranked against each mandate in a fast pass. The strongest go to Deep Scoring, where Prompt Builder prompts, grounded in the mandate and in web research gathered with Firecrawl, produce a score, a written rationale and red flags. The result is a Match Analysis that tells an analyst what fits, what does not, and what to check.

03 Match records, approvals and feedback

Every match is stored as its own record, with its score, rationale and status. Analysts approve or reject matches individually or in bulk, and their feedback is captured on the record. Approved matches become deals with a create and send action, and investor-specific reporting shows what each investor has been offered and why.

04 List Ideas and Target Batches

Not every candidate arrives one at a time. We built a List Ideas interface and Target Batches so analysts can load a set of companies, tag them for several investors at once and push them through screening together.

05 Investor operations

Around the matching engine we delivered the day-to-day pieces a firm runs on: investor onboarding, account hygiene, agenda export for investor meetings, and transaction, pipeline and retainer reporting, so the partners can see what each retainer has produced.

06 Agent infrastructure outside the org

Research and orchestration steps that do not belong inside Salesforce run on a self-hosted stack with n8n workflows and a PostgreSQL store, deployed through Coolify. Salesforce remains the system of record: the outside services do the heavy lifting and write their results back to the match records.

#How the matching works in practice

A typical cycle starts when an analyst loads a set of companies, either a Target Batch from research or a List Idea built around a theme, and tags them for the investors they might suit. The first-pass ranking runs across every tagged mandate and sorts the list, so the analyst sees the most promising pairings first instead of reading the whole set in order.

The shortlist then goes to Deep Scoring. For each company and mandate pair, the system gathers what it can about the company from the web, grounds a Prompt Builder prompt in the mandate and in that research, and writes back a score with its reasoning. Red flags are called out separately, because a strong fit on sector and size can still be ruled out by ownership, geography or a detail buried on a company's own site.

The analyst reads the Match Analysis, approves or rejects it, and adds feedback where the AI got it wrong. Approved matches move into the deal queue, where the create and send action produces the deal record and the investor communication from the same data. Nothing is retyped, and the score, the reasoning and the approval stay attached to the deal for as long as it lives.

#Controls that make it auditable

An AI process in a deal business is only as good as the record it leaves behind. Every match is a record, so the firm can report on what was screened, what was proposed and what was declined for each investor. Every approval carries the name of the person who made it. Every rejection carries a reason when the analyst gives one, which is also the most useful data for improving the ranking over time.

Mandate documents stay in Google Drive, where the firm already keeps them, and are linked to the investor in Salesforce, so the criteria the AI read can be checked against the source the investor signed off.

#How we delivered it, end to end

Data model first. Mandates, investors and the deal queue were agreed before any prompt was written. The quality of every later AI step depends on the structure of the criteria it is matching against.

AI in two tiers. The ranking and scoring tiers were designed separately, so that cost and latency are spent where they change a decision. That split is also what makes large batches feasible.

Human in the loop by design. Approval is a required step in the process, and feedback is stored for every decision. The firm can show an investor the reasoning behind any recommendation.

Experiments kept separate from production. AI idea generation, an alternative ranking proof of concept and large-batch performance testing run as experiments alongside the production process. Their results are being measured before anything is promoted, and target throughput is treated as a goal under test, not a reported result.

Ongoing delivery. The firm remains an active client, with the matching engine and reporting extended in regular releases.

#The decisions that mattered

Building inside Salesforce. A standalone research tool would have been quicker to demonstrate. It would also have left scores and approvals outside the CRM, with someone copying results into deals. Building in Salesforce kept one record, one permission model and one audit trail.

Explaining every score. A score on its own invites an argument. A score with a rationale and red flags invites a decision. Writing the explanation was part of the specification from the start.

Not automating the final call. It was technically possible to send top-scoring matches straight to investors. The firm's reputation rests on judgment, so the final step stays with an analyst.

#What changed

#What we would tell a firm starting this

Write down the mandate before you buy the model. Most of the value in AI matching comes from turning investor preferences into structured, readable criteria, and that work is useful even before the first prompt runs. Then keep the AI in two tiers, keep a person on the final decision, and store every decision. That is what turns an AI experiment into a process a regulated, reputation-driven business can stand behind.