How to Build an Airtable AI Agent for Lead Qualification

Short answer: An Airtable AI agent qualifies leads by chaining steps: read the new record, look up the company in a linked table, check enrichment fields, then write a score and a one-line reason. You define the steps and the output columns. Airtable credits pay for each model call, and agents are available on Business and Enterprise plans.

Build it after you have written down what qualified means. Teams that skip that step get an agent that confidently scores everyone 7 out of 10.

What is an Airtable AI agent, versus an AI field?

A field runs one model call per record. An agent runs a sequence: it can read several fields, look things up in other tables, call a model, branch on the result and write multiple outputs.

Airtable exposes agents from a base’s automation area. You get a step builder, not code. The practical ceiling is around six to eight steps; past that the agent becomes hard to debug and you cannot see intermediate values.

If your task is “categorise this ticket”, use an AI field. If it is “decide whether this lead deserves a call”, that needs the agent.

How do you define “qualified” before building?

Write a table on paper. Five criteria, each with a rule that a stranger could apply.

A workable example for B2B SaaS:

• Company size between 50 and 2,000 employees.
• Industry in your three target verticals.
• A job title that can sign or strongly influence a purchase.
• An explicit problem statement in the form, not a one-word answer.
• Not an existing customer and not a competitor domain.

Then decide the arithmetic: two of five criteria is a warm lead, four or five is a call this week, and a hard disqualifier (competitor, existing customer) overrides everything. The agent needs that arithmetic as instructions, and it needs it in the same words every time.

How do you build the agent step by step?

1. Trigger: new record in the Leads table, or a manual run on selected records during testing.
2. Step 1 — Read. Pass the form fields, the email body and the linked Company record into context.
3. Step 2 — Lookup. Query the Company table for employee count and industry. Do this as a lookup step, not a model call, so it stays deterministic.
4. Step 3 — Score. One model step with the five criteria and the arithmetic above. Ask for a score 1-10 plus a reason of 25 words or fewer.
5. Step 4 — Disqualifier check. A conditional step that sets a hard fail if the domain is on your competitor list.
6. Step 5 — Write. Output to number and long-text fields: score, reason, recommended next action.

Test on 30 historical leads you already know the answer for. That is the only honest calibration set. If the agent disagrees with your human judgement on more than about 5 of 30, the criteria are ambiguous, not the model.

What does this cost to run?

Plan cost first: agents sit on Business at $45 per user per month billed annually, with a workspace credit pool. Each agent run consumes credits for the model steps; lookup steps do not.

Rough math from a small base: 400 leads a month, five steps with two model calls, comes in well inside the Business tier pool for a team of ten. The problem case is re-running the agent on every record edit, which multiplies cost by however often your SDRs touch a row. Set the trigger to new records only and add a manual “re-score” button as a separate run.

Why do agents misbehave in production?

The predictable failures, from watching three of these deployments:

• Score inflation. Agents drift toward the middle and toward positive. A stated promotion risk is not the same as 9/10. Give explicit examples of a 3 and a 9 with reasons.
• Missing data treated as good news. A blank employee count gets read as “not disqualifying”. Add an instruction: unknown values reduce the score, never raise it.
• Tie-breaking on unrelated text. A company named “Group” will trigger industry rules. Use the lookup field, not the model, for anything factual.

The pattern is the same in all three: facts should come from lookups, judgements from the model. Mixing them produces confident nonsense.

Does the score replace an SDR?

No, and it should not be sold internally that way. What it does well is stop three people from manually reading a form that says “looking for a solution” with no other detail.

Where it earns trust is triage: score incoming, route 8s and above to a human within an hour, and put 3s into a nurture sequence without reading them. That saves 20-40 minutes a day per SDR on low-value reading. It does not find leads you would otherwise miss.

Who should build this

Build it if you process more than 150 inbound leads a month, you already pay for Business, and one person will own the criteria document. The economics are clear above that volume.

Do not build it if you are under 50 leads a month — read the forms yourself, you will learn more. Do not build it if your criteria change quarterly, because each change needs recalibration against a new set of known-good leads. And do not build it on Pro; agents are not available there.

Quick Summary

• Airtable agents chain lookups and model calls; use an AI field for single-call classification instead.
• Write five criteria plus a disqualifier rule before you open the agent builder.
• Keep facts in lookup steps and judgement in model steps — mixing them causes confident errors.
• Calibrate on 30 leads you already scored; more than 5 disagreements means the criteria are vague.
• Agents need Business ($45/user/month); trigger on new records only or credit spend multiplies.