Prospecting has always been research under time pressure: which companies are worth contacting, why now, and what can honestly be said to them. Research tools and agents make the first two questions far cheaper to answer. They also make it far easier to send something confidently wrong to a stranger, at scale, in your company's name. This guide covers how to set up prospect research so an agent can do most of the reading: qualification criteria a researcher can apply, signals worth looking for, an evidence pack for every prospect, a written brief for the agent, and an approval step before any message leaves.
Split the work between research and decision
A useful division of labor: the agent reads, collects and drafts; a person decides who is contacted and what is said. That line is not about distrust of the tooling. Research mistakes are cheap to fix while they sit in a table. Once a message is sent, the mistake is public, attached to a name, and cannot be recalled.
Concretely, the research side produces three things for each company: a qualification verdict with reasons, the signal that makes now a good time, and a draft first message built only from those facts. The decision side approves, edits or rejects that draft. Everything upstream of sending can be automated aggressively. Sending itself should not be.
If you are designing the whole funnel rather than the research step, from sourcing to handoff, see the companion guide on building a lead generation pipeline. This page stays with the research and the first message.
Qualification criteria a researcher could apply
“Good fit” is not a criterion. A criterion is something two researchers would apply the same way to the same evidence. Group your criteria under four questions and state, for each, what evidence counts and what disqualifies.
| Question | Example criterion | Evidence that counts | Disqualifies |
|---|---|---|---|
| Fit | Operates in logistics, with its own fleet | Company site, registry activity code | Pure freight brokerage with no fleet |
| Need | Manages maintenance across several depots | Depot list on the site, job postings for depot roles | A single location |
| Timing | A dated signal from the last few months | Press release, careers page, public tender | No dated signal: the company stays in nurture, it is not dropped |
| Access | A named function owns the problem | Team page or a job posting naming the role | No identifiable owner |
Three rules keep this honest. Each criterion resolves to yes, no or unknown. Unknown is not no: a company with an unconfirmed criterion goes to a review queue rather than being dropped or, worse, quietly passed. And fit criteria should come from your own won and lost deals, not from a generic persona; if you have not written your ideal customer profile down yet, the guide on finding companies that match your ICP starts there.
Signals: what makes now the right time
A signal is a public event that changes the odds a company cares about your problem this quarter. Good ones for most B2B products:
- Hiring for a role your product supports or replaces.
- A new leader in the function that would buy, especially in their first months.
- Expansion: a new site, market or product line that strains existing processes.
- A public tender or request for proposals in your area.
- A regulatory change that lands on their sector.
- A visible technology change, such as a migration away from a tool you integrate with.
Every signal needs three things to be usable: a date, a source address, and one sentence on why it matters for this company. Without the date, a two-year-old announcement reads as news. Without the sentence, the signal is trivia. Signals also expire; set a window after which a signal no longer justifies outreach on its own.
Funding announcements are the signal everyone watches and the one most often misused. A new round tells you a company has money, not that it has your problem. Use it to adjust priority, not as the reason for the message. If you want signals to arrive continuously instead of being researched on demand, the guide to sales trigger monitoring covers that setup.
An evidence pack for every prospect
The output of research should be a short, structured pack, not a paragraph of prose. Prose hides which statements were checked. A pack makes it obvious.
prospect example-logistics.com fit yes 'Our fleet of refrigerated trucks...' /about retrieved YYYY-MM-DD need yes depot list shows several sites /locations retrieved YYYY-MM-DD timing yes posting: Fleet Maintenance Planner /careers/123 retrieved YYYY-MM-DD access unknown no team page; role inferred from the posting angle planning maintenance across depots as the fleet grows open who owns maintenance software today? not found current tools in use (no public mention)
The rule that makes the pack valuable: the outreach draft may only state facts that appear in it. That one constraint removes the most embarrassing kind of AI-written outreach, the invented compliment about an award the company never won or a product it never launched. It also separates two things that look alike in a spreadsheet: a fact supported by a quote, and a question for which nothing was found. Keep both, in different fields. A missing answer is a question for the seller to ask, not a gap to fill with a plausible guess; more on that in missing evidence is an open question.
Briefing the agent
Agents do noticeably better with a written brief than with a one-line request, and the brief doubles as the record of what was asked. A workable template:
goal qualify the attached company list against criteria v3 and build an
evidence pack for each company that passes or is borderline
inputs company list (domain + name); criteria v3; our product summary
sources company websites, official registries, public job postings;
no login-only sites, no purchased contact data
steps 1 confirm the company is active 2 test each criterion with a quote
3 look for a dated signal 4 draft a first message from the pack
output one pack per company; 'unknown' wherever no evidence was found
done when every company has a verdict and every yes has a quote and a source
never send, schedule or export anything; contact no oneWrite the “done when” line before the agent starts. It turns a vague research request into something a reviewer can check, and it is the line most often skipped; specifying a task walks through the same idea. In GoCux this brief maps onto a work order: a goal, the boundaries of what it may touch and the acceptance criteria, and each step is written to the order the moment it happens. A research plan with its source allowlist is approved before it runs, and running it cannot quietly add sources the plan did not name.
Approval before outreach goes out
Sending is the step that cannot be undone, which makes it the right place for a person. A reviewer looking at a draft should be able to answer, without opening a browser:
- Is this the right company, and the right recipient for this role?
- Does every factual statement in the message appear in the evidence pack?
- Is the signal still current, and is the reason for writing stated plainly?
- Is the recipient clear of suppression and opt-out lists, checked at send time rather than when the list was built?
- Would you be comfortable if the recipient forwarded this message to a colleague?
The reviewer's options should be approve, edit or reject, and each outcome should be recorded with who decided. Rejections carry information: a run of rejections for the same reason usually means a criterion or the drafting instruction needs to change.
In GoCux, sending something out is one of the actions that stop and wait for a person on the team, who can approve it, edit it safely or reject it. The approver sees the exact message that was about to go out, quoted as it was requested, and the approval covers that one attempt. It is checked again when the send actually runs, so approving one draft does not open the door for the next fifty. Drafts can also be prepared as Responses, answers assembled for review inside the workspace before anything leaves it. Why an agent should ask before it acts covers the reasoning behind this design.
A short checklist before the first batch
- Fit, need, timing and access criteria are written, and each resolves to yes, no or unknown.
- Every signal has a date, a source and one sentence on why it matters; old signals expire.
- Each prospect has an evidence pack, and the draft may only use facts from it.
- Unknowns go to a review queue, never silently in or out.
- The agent's brief says what done looks like and forbids sending.
- A person approves, edits or rejects each message; decisions and reasons are recorded.
- Contact data has a lawful basis under the rules that apply to you, such as KVKK or the GDPR, and opt-outs are honored in every tool.
The sales solution page shows how account research, evidence and approvals fit together in the workspace.
Bring this workflow into your workspace.
Set the scope. Review the evidence. Approve the steps that need it.