Company Lookup for AI Agents: Using MCP Tools in Workflows

A company lookup over MCP lets an AI agent turn a domain into company facts by calling a tool, instead of by browsing or guessing. The agent sends a domain, receives structured fields or a clear not found, and works from that. The benefit is grounded data; the risk is an agent that embellishes, so guardrails matter.

What is exposed

According to BusinessMCP’s company lookup page, the lookup is available over REST and as MCP tools: company_lookup for a single domain, company_bulk_lookup for batches, and email_pattern for a company’s address format. Its data API page adds that every tool is also an MCP tool usable from Claude, Cursor, VS Code or any MCP client, with the same key and the same prices as the REST API. The MCP project’s own introduction explains the protocol in general terms.

What an agent can usefully do

Task	Tool use	Human check
Qualify an inbound lead from a work email	company_lookup on the email’s domain	Review borderline cases
Enrich a spreadsheet of domains	company_bulk_lookup in batches of up to 100	Sample the output
Prepare for a call	company_lookup plus a web read	Verify claims before quoting
Draft outreach context	company_lookup, email_pattern	Compliance review before sending
Clean a CRM export	Bulk lookup, then write columns	Approve before overwriting

Key takeaway: an agent should report what the tool returned, including blanks, and leave interpretation clearly separate from data.

Guardrails that matter

1. Quote, do not paraphrase fields. Ask the agent to show the returned values, then summarise. This cuts silent embellishment.
2. Treat found false as final. Instruct the agent not to fill gaps from memory or to invent a profile. BusinessMCP says its API never invents a record to fill a gap, and your agent should hold to the same standard.
3. Cap spend. Each call can cost money under pay-per-call pricing. Use a dedicated key with a spending cap and a low-balance alert if the vendor provides them; the API reference mentions key caps as a reason a charge is refused.
4. Limit batch size. Tell the agent to process in small batches and stop on repeated errors.
5. Keep writes behind approval. Reading data is low risk; writing it into a CRM is not.
6. Log tool calls. You want a trail of what was asked and returned.

Cost and limits for agent use

Agents can call tools more often than a human would, so the bill can run away. At the listed rate of $0.02 per found company, a loop that re-looks-up the same domains repeatedly spends money for no new information. Tell the agent to check a local cache first, and design the workflow so each domain is looked up once per refresh period. Rate limits apply too: the API reference lists 600 calls per minute per key, and 60 per minute for heavier tools including bulk lookup. See company lookup API pricing for a worksheet.

Testing before you wire it in

BusinessMCP lists free tools that help with MCP setup, including an MCP server tester and an MCP config generator. Beyond that, run your own acceptance test:

1. Give the agent five domains: a large company, a mid-size one, a small one, a brand-new startup and a fake domain.
2. Check that it reports found and not found correctly.
3. Check that it never adds facts that are not in the tool output.
4. Check that it stops cleanly on an error such as insufficient balance.
5. Compare cost before and after.

If the agent fails any step, tighten its instructions and run again.

Prompting patterns that help

• “Use only the lookup tool’s output for company facts. If it returns found false, say so and stop.”
• “Show the raw fields first, then a two-line summary.”
• “Process at most 20 domains per run and report the count found and not found.”
• “Never email or write to the CRM without my approval.”

These are simple instructions, but they address the usual failure modes: guessing, runaway loops and silent writes.

Where agents fit with the rest of the workflow

An agent is a convenient front end for the same data. The underlying lessons about domain lookups, misses and REST calls still apply. Agents suit exploratory work and one-off research. For a nightly job over thousands of domains, a plain script with the bulk endpoint is cheaper and easier to audit.

Privacy and compliance

Company data is largely about organisations, but an agent workflow often mixes it with names and emails from other sources. Keep personal data out of prompts unless you have a lawful basis and a vendor agreement that covers it. The company lookup page says its data is firmographic and concerns legal entities rather than persons, which is the vendor’s statement, not a legal conclusion. Seek advice for your situation.

A short checklist

• Dedicated key, spend cap, low-balance alert.
• Instructions that forbid guessing and require quoting tool output.
• Small batches and a cache.
• Approval gate before writes or outreach.
• Acceptance test with a fake domain.
• A log of tool calls for review.

Example: qualifying inbound signups with an agent

Imagine a small team that receives forty signups a day. An agent reads each new work email, extracts the domain, calls the lookup tool and posts a short note to the sales channel: company name, industry, size band, country and a link to the company’s site. For free-mail addresses it posts nothing and tags the signup as personal. For a miss it posts “no record found” and nothing else.

Three design choices keep this safe. First, the agent never replies to the signup, so no one receives an automated message based on a guess. Second, the note includes the raw size band, not a rewritten figure. Third, a human decides whether to follow up. Daily cost is easy to forecast: at most forty lookups, charged only for those that find a company, at the listed $0.02 each.

Failure modes to watch for

Failure	Symptom	Fix
Embellishment	Summary mentions facts not in the tool output	Require quoted fields
Loops	Repeated calls for the same domain	Cache and cap calls per run
Silent errors	Agent reports success after an error	Require explicit status reporting
Over-reach	Agent writes to CRM unprompted	Approval gate
Stale data	Old record presented as current	Show the lookup date

Review a sample of agent runs weekly at first, then monthly. Most problems show up in the first few dozen runs, and tightening the instructions after each one is cheaper than discovering a pattern after a thousand.