Design an AI Lead-Research Workflow for a Consulting Business

Before adding another prospecting tool, decide what a useful research result would let you do.

For a consulting business, that might be a simple decision: investigate this company further, leave it for later, or rule it out. An AI lead-research workflow should produce enough traceable evidence to support that decision, with the uncertainties still visible.

The workflow below is a proposed starting point for researching a small, supplied list of companies using their publicly available information. It ends with a human-reviewed research brief. It does not collect personal contact data or send outreach.

Check whether research is the constraint

Look at your last few attempts to win work. Where did progress actually stop?

• You couldn't explain who the offer was for
• You had promising companies but kept repeating the same research
• You collected information without deciding what to do next
• You understood the companies but hadn't made time for conversations

This workflow addresses the repeated-research problem. If the offer is unclear, write a narrower description of the problem you solve first. If conversations are waiting on you, producing more briefs may only lengthen that queue.

Choose one decision the research will support. For example: “Which companies on this list show a documented operating change relevant to our onboarding consulting offer?” That is a proposed research question, not a claim that any company needs the service.

Define a bounded research job

Start with a list you already have permission to use. Give each company its verified official domain so similarly named businesses don't become mixed together.

Write a short brief before running the research:

• Offer: the specific business problem you help solve
• Working set: the supplied companies and domains
• Required fit: observable conditions a company must meet
• Relevant signals: company announcements or operating changes worth checking
• Exclusions: companies or situations outside the offer
• Output: one short brief per company, with evidence and unresolved questions
• Limit: the number of companies, source pages, and research time allowed

For a first trial, five companies and a small page limit per company are manageable proposals. Adjust those limits to your review capacity. A time limit should end in a visible “incomplete” status if evidence is missing.

Keep this first workflow read-only. Use a research-capable tool that can open source pages and return their URLs, plus a document or sheet you already use. Avoid adding automated contact discovery, CRM changes, and message sending to the same job.

Collect evidence before writing the summary

Restrict collection to publicly accessible company sources: official product and service pages, company newsrooms, careers pages, and official reports or filings. Search results can help locate a page; open the original page before treating it as evidence.

Capture a record for each useful observation:

• The company and official domain
• The specific observation
• The exact source URL and page title
• The publication or event date, when shown
• The date you observed the page
• Any qualification that changes its meaning

Publication date and observation date answer different questions. Reading an old announcement today does not make its news current. An undated page should remain undated.

If two sources conflict, preserve both and flag the conflict. If a page is inaccessible, record the gap. Don't let the system substitute a plausible fact to finish a paragraph.

Anthropic's guidance on evaluating research agents separates checks for source support, coverage, and source quality. That is a useful basis for deciding what a research brief must demonstrate. Read Anthropic's research-evaluation guidance.

Separate evidence, inference, and unknowns

Use three explicit labels in every brief.

Evidence is what the opened source supports. Attribute company claims to the company when independent verification is unavailable.

Inference is your interpretation of why that observation may matter to your offer.

Unknown is something the available material cannot establish.

Consider this hypothetical example:

• Evidence: An official careers page lists an implementation-manager role whose description includes coordinating customer onboarding
• Inference: Onboarding coordination could be a useful subject for further research
• Unknown: Whether onboarding is currently a problem, whether outside help is wanted, who owns the decision, and whether there is a budget

The job listing does not establish purchase intent. Keep the inference modest enough that a future conversation can genuinely test it.

Likewise, “not found in the reviewed sources” is different from “does not exist.” A small public-source review cannot establish a complete picture of a company or the market.

Use a company research worksheet

The following is a proposed worksheet. Keep it short enough to review alongside its source pages.

Company research brief

Identity - Company name: - Verified official domain: - Research completed on: - Sources reviewed and coverage limit:

Fit assessment - Required condition: - Supporting evidence and source URL: - Assessment: supported / contradicted / unknown

Relevant observation - What the source actually says: - Page title and exact URL: - Publication or event date: - Observed date: - Important qualification:

Interpretation - Why this may relate to the offer: - Alternative explanation: - What remains unknown:

Review decision - Investigate further / defer / exclude: - Reason for the decision: - Next question to resolve: - Human reviewer and review date:

Avoid an overall numerical score in the first version. Visible conditions are easier to challenge than a precise-looking ranking with unclear assumptions. If scoring later becomes useful, write the rules and show the evidence behind every component.

Give the AI clear research instructions

Adapt these proposed instructions to your chosen tool:

Research only the supplied companies and verified official domains. Use publicly available company sources. For each material observation, include the exact page URL, page title, publication or event date when available, and observed date. Separate evidence, inference, and unknowns. Do not infer budget, buying intent, personal contact details, or internal problems from a weak signal. Do not invent missing information. Flag conflicts and inaccessible pages. Stop at the agreed limits and return incomplete results honestly. Produce the company research worksheet for human review. Do not contact anyone or modify business records.

Instructions should match the tool's actual permissions. Where possible, give this workflow only the access needed to research and prepare its output.

Review every brief before it influences outreach

Open the sources behind the decisive claims. Check identity, dates, fit, and whether the interpretation goes beyond the evidence. Reject briefs that rely on an unverified contact or a guessed internal problem.

OpenAI's agent-building guide recommends human intervention when failure limits are exceeded or an action is high risk. See its guidance on human intervention.

For this proposed workflow, make human review mandatory before outreach. A reviewed research brief still doesn't establish the correct recipient or authorize a message. Keep those as separate decisions.

After the first batch, count the briefs you accepted, the ones needing correction, and the minutes you spent checking them. Note whether each accepted brief supported a clear next decision. Improve the most repeated failure before increasing volume.

Start with one offer, a small company list, and this worksheet. The useful result is a research process you can inspect, repeat, and fit into the time you actually have.