Agents that find, research and qualify the right business before a word is written

Most cold outreach fails before the email is even drafted, because it goes to the wrong company. Spray a list, hope a few care, annoy the rest. We wanted to do the opposite: spend almost all the effort up front, on working out exactly who is worth approaching and why, so that by the time anything gets written there is a real reason to write it.

So we built a multi-agent outreach system, first for ourselves and then for a campaign aimed at UK manufacturers who were likely missing government grants they could claim. The part we are proud of is not the email at the end. It is the research and the qualification that happens before it.

Defining the business worth approaching

Before any of it runs, we write down what a good fit actually looks like, in a form the agents can score against. Not a vibe, a rubric. For each campaign there is an ideal-customer-profile file with weighted points and, just as importantly, hard dealbreakers.

The dealbreakers do the heavy lifting. A sole trader is out, because cold-emailing one is not legal here without consent. A dissolved, dormant or distressed company is out. A competitor is out. Anything outside the target size band is out. One dealbreaker disqualifies a company no matter how well it scores everywhere else, which keeps the system honest about who it is allowed to talk to.

What survives the dealbreakers then gets scored, on industry fit, company size, region and live growth signals, against a profile we tune per campaign. The output is a tier and a number with the reasoning written out next to it, so a human can see exactly why a company qualified or did not.

How the agents do the research

The work is split across a small team of agents, each with one job, run in order. That sounds like overhead. It is the point. A single agent asked to find, judge and pitch a company in one breath will quietly invent the bits it cannot find. Splitting the job lets us check the work between steps.

• Discovery. It pulls candidate companies from the Companies House API by industry code and region, so every lead starts from the official register rather than a bought list.
• Dedup. Every candidate is checked against our own database first. Anyone already contacted, or who has ever asked us to stop, is dropped before any effort is spent on them.
• Research. A researcher agent reads the official filings, the registered officers and people-with-significant-control data, then the company's actual website, to find the real decision-maker and a genuine reason to get in touch. Not a generic sales inbox, the named director.
• Grant and competitor intel. It checks public funding records for the company, and for similar firms nearby who have claimed grants. A neighbour in the same trade who has already won funding is the strongest, most concrete reason for the recipient to read on.

The hard part: getting an AI to not make things up

The genuinely difficult problem was not gathering information. It was stopping the system from confidently stating things that were not true. Early on, a research run described offices that did not exist, products that were not on the site and "no recent news" on a company that had posted news that month. Every one of those would have landed in a real director's inbox, who would have clicked their own site, seen the claim was wrong and never trusted us again.

So we made it a rule of the architecture, not a hope. The researcher has to attach a verbatim quote and a source URL to every factual claim it makes. Then a separate verifier agent re-reads the brief against the live sources and marks each claim pass or fail. Anything without a source is deleted before it can reach the writing stage. Opinions are allowed, but they have to be labelled as opinions, never dressed up as facts.

Only after a lead has cleared all of that does a writer draft a short, personal email that links to its claims, and a reviewer checks it one last time for accuracy, tone and compliance. Nothing sends on its own. A person approves every message before it leaves.

What it means

The result is a system that spends its effort where it matters: on defining the right business, proving the facts about it and only then writing to a named person with something true and relevant to say. It is the unglamorous half of outreach that everyone skips, done properly by machines that are made to show their working.

This is the kind of thing we mean by AI and automation: not a chatbot bolted on, but agents doing real research with the guardrails to keep them honest.