Your AI marketing agent is reading whatever broke last quarter

A loyal customer of yours bought the same product three times last year, then went quiet six weeks ago. The lifecycle agent on top of your email tool looks at her profile, sees the gap in opens, sees the gap in purchases, and triggers the win-back sequence. A polite reminder. A 15% discount. A "we miss you" headline. The same customer, last week, called support twice about a defective unit and waited 38 minutes for a callback. Your support tool knows. Your returns dashboard knows. The email tool doesn't, because the integration that's supposed to push support tickets into the customer profile broke in February and nobody noticed. The agent didn't get it wrong. It made the right decision on the data it could see. The data it could see was wrong.

Most teams are about to install AI agents on top of marketing tools whose integrations haven't been checked in a year. The audit that finds those broken connections takes a Tuesday afternoon and a spreadsheet — and it has to happen before the agent makes its first decision, not after. The boring fix is upstream of the AI. It's mostly free.

What your email tool actually knows about a customer

Pick a real customer. Open her profile in your email platform. Make a list of what you see: email opens, clicks, purchase history, last campaign sent, segment memberships, maybe a custom field or two.

Now write a second list of what's true about that customer that isn't on the first list:

• The two support tickets she filed last week
• The product return she initiated and then cancelled
• The 14 sessions on your site in the last month, half of them on the help center
• The two retargeting ads she saw and ignored
• The reason her last order shipped late
• The fact that her household has a second account under a different email

The first list is what your email tool can see. The second list is what every other tool in your stack can see — and what your email tool can't, unless someone wired an integration that's still working today. The gap between the two lists is what the agent will be missing.

This gap was always there. For ten years it was a nuisance. A marketer reading the segments in her head knew the data was partial and adjusted — second-guessed the win-back trigger, asked the support team about the cohort first. When the data felt wrong, she stopped. The agent doesn't stop. It treats whatever's in the profile as the customer.

Why AI made this worse, not better

A marketer running off partial data is a marketer making partial decisions slowly. She compensates. She hedges. She asks around. An agent running off the same partial data is a marketer making partial decisions fast, at scale, on every customer in the segment, without hedging or asking around. The same broken integration that produced one awkward win-back email last year now produces ten thousand of them.

The agent is doing what you told it to do. The instructions and the model are fine. The customer profile it's reading is the variable. This is why "let's put AI on top of our marketing stack" is the wrong first move for most teams. The agent inherits every gap, every broken sync, every field that hasn't been filled in since the last admin left. The faster the agent runs, the more visible the gaps get — usually to the customer, in their inbox.

The audit, before the install

When we wired AI into TheNextGuide's tour catalog, the first month was almost entirely this exercise. No agent made a decision until the integration map was clean. Two hours of someone's time. A spreadsheet. No new tools.

Inventory the events first. List every customer-facing event you'd want the agent to know about. Purchase. Return. Support ticket. App session. Site session. Cart abandonment. Survey response. Loyalty redemption. Ad click. Whatever your business actually produces.

Map the wiring next. For each event, write down which tool is the source of truth and which tool needs to read it. Most events are produced in one tool — your store, your support tool, your app — and consumed in another. The line between them is an integration.

Then do the actual audit. For each integration, find out when it last successfully fired. Half are working. Some haven't fired in months and nobody noticed because nobody was looking at that field. The rest fire, but the data lands in the wrong field or the wrong format.

Last column is the decision. Either the integration gets repaired before the agent goes live, or you tell the agent in writing that this field is unreliable and not to use it. Both are acceptable. Pretending a broken field is good when it isn't is what breaks the install.

Most teams find five to ten broken or stale integrations they didn't know about. None of the fixes require new software. Most require someone to spend a Tuesday afternoon in the integration settings of tools they already pay for.

What this gets you

When the agent goes live after this audit, it goes live on data you've actually looked at. The win-back sequence doesn't fire on the customer who called support twice last week, because the support data is either in the profile or explicitly marked off-limits.

Most of the install pain isn't in the AI. It's in the integrations that broke last quarter and have been broken ever since.

Book a strategy call →