AI Agents for Marketing Ops, Behind an Approval Gate

Where an agent earns its place: About four of the twenty-odd repeating decisions in a marketing operation are worth handing to a language model. Classifying free-text inbound: demo forms, event badge scans, support tickets that are really sales. Turning call notes into CRM field updates. Triaging conflicts between two enrichment sources. Drafting the exception note when routing sends a lead somewhere unusual. The rest of what gets sold as AI for marketing ops is a deterministic rule in a costume, and it belongs in code or in your CRM's own workflow engine where somebody can read it. I build the four. I decline the costume, and I say so in week one rather than after invoicing. How the work runs: Week one is a decision inventory. I sit with whoever makes each of these calls by hand today and write down what they actually look at, including the parts they do not say out loud. That document is worth having even if you stop there. Week two builds the evaluation set before it builds anything. Two hundred real historical cases pulled from your own CRM, labelled by your team, held out from the prompt. An agent that has never been scored against your own past decisions is a demo, not a system. From week three the agent runs on live data in shadow mode: proposals go to a queue, not to records. Your ops person clears the queue each morning and every disagreement is logged. Write access opens only when agreement on the held-out set holds above the bar we set in week one, usually between 90 and 95 percent depending on what a wrong call costs. Then one object type at a time, with the queue still in place as the fallback path. What it is built on: The Anthropic or OpenAI API, called from a Temporal or n8n workflow that owns retries and idempotency, against HubSpot or Salesforce through their REST APIs using a service account whose scopes are written down and approved by you. Prompts and evals live in your Git repository, not a vendor canvas. Every run writes a row: inputs, output, model version, cost, whether a human overrode it. When a provider ships a new model, we re-run the evaluation set before switching and you decide with numbers in front of you. Not included: Outbound copy that sends without a person reading it. Delete permissions. Anything touching deal amount, close date or stage, because those carry your forecast and a model has no business writing them. Chat widgets. Predictive lead scoring, which is separate work and needs more closed-won history than most companies have. Who this is not for: Teams whose CRM is not yet trustworthy. If the same company exists four times and half the contacts have no owner, an agent will make confident decisions on wrong inputs faster than any human could, and you will find out late. Fix the record layer first; I would rather sell you that. Also not for anyone who wants a count of agents for a board slide, because the number I hand back is smaller than the one you were hoping for and I will argue for it. What you are left with: A runbook in your own repository. A stop switch any one of your team can pull without calling me. A per-run cost you can read off a dashboard. And a standing agreement: if the queue stays empty and nobody overrides anything for a month, I tell you to turn the agent off and hardcode the rule.