The short version
A basic AI research agent wired through MCP, pulling company data, recent news, hiring signals and tech stack into one structured brief per account, doesn’t replace research, it compresses it. What used to be 15–20 minutes of manual tab-switching per account becomes a 90-second generated brief that a rep reads and sense-checks before a call.
The time saved doesn’t disappear; it should be redirected into more accounts covered per week or more thought per conversation, not just “doing less work.” The build itself is modest: an n8n workflow triggered when a new account enters the pipeline, calling two or three data APIs, and a prompt that structures the output consistently. Most of the engineering effort goes into prompt consistency, not the API calls.
Choosing sources and defining the schema
Start with three sources that answer different questions: recent news for triggers, job postings for priorities and growth, and technology signals for fit. Check the terms of each API and whether results can be stored. Resist adding a fourth source until the first three are reliably accurate, because every source adds failure modes.
Fix the output schema before writing a prompt. A workable brief has six fields: one-line company summary, the trigger event with date, likely stakeholders by function, one hypothesis on pain, risks or reasons not to call, and source links. Fixed fields allow reps to scan in seconds and let you compare accuracy across accounts.
Prompt and data discipline
Instruct the model to use only the supplied data and to write “not found” when evidence is missing. A sample line: “If a fact is not present in the provided sources, output NOT FOUND for that field. Never infer revenue, headcount or product names.” This single rule removes most of the confident guesses that damage credibility.
Require a source link beside every claim, and have the workflow discard any claim lacking one. Pass news items with their publication dates and tell the model to ignore anything older than 90 days. Store the raw inputs next to the brief, so when a rep says it is wrong you can see whether the data or the model failed.
Rolling out and measuring outcomes
Trigger the flow when an account hits a defined stage or list, and write the brief to a CRM note, so reps see it where they work. Add a “this is wrong” button that logs the field and reason. Review 30 briefs against manual research before wider rollout, scoring each field as correct, partly correct or wrong.
Track what the saved time buys. Compare accounts touched per rep per week, personalised messages sent and meetings booked before and after. If hours drop but output stays flat, reps are banking the time. Set the expectation up front that freed capacity goes into coverage or conversation quality, and check it monthly.
Worked example (illustrative)
A hypothetical team of four reps adds an automated brief for each new target account and audits 60 briefs in the first month.
What changed: Review time stayed in the process, which kept errors from reaching prospects. The extra meetings came from broader coverage, not from the briefs alone.
Do it: the steps
Tick them off as you go. Your progress stays on this device.
Traps to avoid
- ×
Measuring success by time saved alone. Measure what the saved time produced: accounts covered, meetings booked.
- ×
Giving the agent write access to the CRM beyond a note field. Keep outputs in a single notes field until accuracy is proven.
- ×
Changing the prompt without versioning it. When quality shifts you will not know which change caused it.
What to track
- Brief accuracy
- fields judged correct divided by fields audited on a sample of 30 or more; aim for 90% or above
- Research time per account
- median minutes including review, versus a measured manual baseline
- Coverage per rep
- unique target accounts touched per week; the number the saved time should raise
Benchmarks are typical planning ranges, not guarantees. Calibrate to your own data.
Questions people ask
Do I need MCP specifically, or are plain API calls enough?
Plain API calls in n8n work for a fixed three-source flow. MCP becomes attractive when you want the model to choose among tools or reuse the same connectors across several agents. Start simple, and adopt MCP when tool reuse or flexible lookups justify it.
How do I stop the agent inventing facts?
Constrain it to supplied data, demand source links, allow NOT FOUND, and keep temperature low. Then audit a sample weekly. Treat any invented claim as a defect to trace back to a prompt or data issue, not as acceptable noise.