— Field Notes No. 14

The Customer Cannot Live in One Person's Head

The CRM knows the deal. The ad server knows what ran. Finance knows what got paid. The person holding the relationship knows why. Agents can reach every system and still get the customer wrong.

August 21, 2026·13 min read

Nobody can tell, from one place, whether the campaign is actually booked.

The signed order is in a shared drive. The seller confirmed a change in an email. Creative approval is sitting in somebody's inbox. The reservation is in the ad server. Finance has a different total. Ad operations knows which version everyone finally agreed to, because ad operations always knows, and none of the systems do.

I spent part of this week in Chicago at Local Media Fest, with people who carry real responsibility for local media businesses, and some version of that gap sat underneath every serious conversation about AI. It's a problem I've watched for years: much of what we call a working system is really a good employee remembering how to get the work through it.

Humans can live with that arrangement. We reconstruct the truth from the trail, ask the person who was on the call, and carry the answer forward. A machine can't safely do that. It can read the same trail and make a convincing guess, then act on the guess before anyone realizes it was wrong.

Which is why a new IAB Tech Lab effort called AAMP — Agentic Advertising Management Protocols — matters well beyond the people who write advertising standards. The project is building common ways for buyer and seller agents to discover inventory, negotiate, book campaigns and move them into ad servers. Its reference buyer can already take a brief, research sellers, negotiate and book a deal, with an approval before the money commits. It's still an alpha, not the way most advertising gets bought today. But the direction is clear enough to plan against. (IAB Tech Lab, August 12, 2026; IAB Tech Lab Buyer Agent)

The most important part of that work isn't the agent. It's the record underneath it.

The people were holding the systems together

IAB Tech Lab describes the direct-sold campaign exactly as an ad operations person would: its state is scattered across a dozen systems and informal channels. A human can look at the insertion order, the email, the Slack message and the trafficking record and understand that a campaign is reserved but not booked, or approved except for one creative. An agent needs the difference stated plainly.

Draft. Reserved. Booked. In flight. Finished.

A rate, a quantity, a start date, an approval, each attached to the same order and meaning the same thing to the buyer and the seller. Not inferred from whichever email happened to show up in the search results. (IAB Tech Lab, “The Contract Layer for Agentic Advertising,” April 21, 2026)

That should sound less like the future than a Tuesday. Take one local advertiser. The CRM knows the opportunity. The proposal knows what was recommended. The order system knows what was sold. The ad platforms know what ran. Finance knows what was billed and paid. The reporting platform knows what it counted. The seller knows the owner cares less about clicks than about filling the empty lunch hour, because he heard her say it across a table.

No system knows the customer.

The person does.

I spent more than twelve years running publisher-owned agencies. I can't count how often the clean client experience depended on someone knowing that the order was wrong, the invoice was right and the reason the client bought never made it into the CRM. We called all of that service. Some of it was. Some of it was a good employee hiding a bad system.

We've spent years calling that relationship management when a lot of it is systems integration performed by a human being. The seller, strategist, account manager or ad operations lead carries the context from one screen to the next: which promise changed the scope, which number is stale, which report the client believes, why the renewal is in trouble even though the dashboard is green. When that person leaves, the company finds out how much of the customer left with them.

This isn't a new problem. A 2020 Local Media Association data lab found the same complaints across major publishers: people working in systems that couldn't talk to one another, every platform arriving with its own reporting tools, too many dashboards and too much manual processing. The Seattle Times was working to eliminate four disconnected advertising systems. Lee Enterprises was building more than seventy reporting connections. (Local Media Association and Google News Initiative Data Lab, February 2020)

Six years later, the same weakness costs more. An agent can only carry work the business has made carryable. Every handoff needs a fact it can trust, an owner and a next state, and the next step can belong to a person, a system or an agent. If the record arriving with it is a guess, drawing more boxes around the work won't save it.

A pile of data is not a memory

For years I told anyone who would listen that we didn't have a data lake. We had a few hundred data puddles. Every one of them held something true. None of them agreed.

That's why this conversation always jumps to a warehouse. Move everything out of the native platforms. Put it in Snowflake. Put the files in an Amazon bucket. Index the documents. Connect an agent. Now the company owns its data and the problem is solved.

Some of that may be exactly right. A publisher should keep a usable copy of anything it can't afford to lose: customer and transaction history, contracts, product and pricing records, performance history, the audience relationships it collected directly, and the record of what was approved and what changed. If a vendor closes, doubles the price or simply stops fitting, the business shouldn't have to start over.

But moving the mess doesn't make it make sense. Put five versions of the same advertiser in a warehouse and you have five versions of the same advertiser in a more expensive room. Copy the proposal, order, invoice and campaign report into a bucket and you still have to know which number governs what. Build an index over every email and the machine can find the sentence where someone typed “looks good.” It still can't tell you whether that sentence approved the creative, the price, or lunch. The puddles came with you.

Storage preserves the record. It doesn't decide what the record means.

That distinction is finally showing up in the architecture guidance coming from the platforms themselves. AWS now describes the enterprise data problem as interoperability, not wholesale consolidation: some data gets copied, some stays in the system built to run it, and a shared catalog identifies what exists, what it means and which source is authoritative. Microsoft makes the same practical distinction between live transactions, structured queries and document search. A policy manual may belong in search. Current inventory belongs behind a live call to the system that owns it. (AWS, July 13, 2026; Microsoft Cloud Adoption Framework; Microsoft search and retrieval patterns)

You don't need one database to win. You need the business to know which system wins each question.

What did the client ask for? The discovery record.
What did we promise? The approved proposal and contract.
What can we sell? The current product and pricing catalog.
What did we book? The order.
What ran, and what did it earn? The delivery systems, then finance.
Who is allowed to change any of it? The permission and approval rules, written down before a machine asks.

That's a harder project than buying storage, and a much smaller one than moving everything. Start with one revenue line and the handful of facts the business keeps recreating to run it.

The screen is about to stop mattering

The software companies can see this coming. In April, HubSpot made a hosted connection generally available that lets compatible AI tools read contacts, companies, deals, calls, meetings, notes, tasks and emails, and write back to many of those records under the user's existing permissions. Salesforce did the same for enterprise customers, exposing data, queries, flows and business actions to outside AI clients. Its announcement opened with a question from co-founder Parker Harris: “Why should you ever log into Salesforce again?” (HubSpot, April 13, 2026; Salesforce, April 29, 2026)

That question is bigger than a new way to use a CRM. For thirty years, business software competed to become the screen where the employee spent the day. The next interface may sit above all of those screens. A seller asks for the account brief. An agent retrieves the history from the CRM, the pacing from the ad platform, the unpaid invoice from finance and the last three emails, then recommends what to do next. The employee may never open any of the underlying systems.

Good. Most of those screens were lousy places to spend the day, and the work was never the value either.

The last note argued that the valuable layer sits above the platforms: the one that knows the advertiser's actual objective, decides which agent can use which information, and measures whether any of it produced an economic result. That layer is only as good as the record underneath it. Whatever assembles the answer is deciding which sources to ask, which record to trust, what context to leave out and which action the user is allowed to take. If the publisher hasn't made those decisions, the agent, or the platform providing it, will make them by default.

We've seen this movie.

Local media spent years building audiences on platforms it didn't control. Then the platforms changed the distribution, kept the data and sold publishers access to the same people again. The industry's response has been to rebuild direct relationships through registrations, newsletters, subscriptions, events and other first-party connections. Local Media Association now calls zero- and first-party data strategically critical and recommends a lightweight infrastructure and governance model around it. News Product Alliance makes the cost of the old arrangement plain: when the platform changes its rules, the mediated relationship can disappear with it. (Local Media Association, February 3, 2026; News Product Alliance, November 4, 2025)

The next platform shift is moving inside the building. This time the risk isn't only losing the route to the audience. It's losing the only complete account of how the business serves a customer.

Keep the record. Replace what has to go.

None of this is an argument to build everything. It isn't an argument to keep everything either.

The build, buy or partner call still holds. Buy the commodity. Partner on the heavy shared systems. Build the short list where the operation is actually different and the economics justify owning it.

Then run the same honesty at what you already own, because some of it has to go. The system the operation outgrew three products ago, that everybody now routes around. The platform billing enterprise price for the ten percent of it anyone opens. The one that can't hand its own records to anything else: no usable export, no API, no way for the next system or the next agent to reach what's inside it. That last one used to be an inconvenience. In an operation where agents do the carrying, a system nothing can read from is a system the business can't use.

Keeping any of those isn't discipline. Nobody decided to keep them. The renewal just kept arriving.

And one item on the list is a build. Your own usable record of the customer across the systems is not something anybody sells you. Not a warehouse: the thin layer that says which system answers which question, holds the identifiers that tie an advertiser's opportunity to its order to its invoice, and keeps the approvals and the changes where both people and agents can read them. It's the smallest thing on this list. It's also the only one that has to survive every vendor you replace.

Every one of these choices needs an exit, including the systems you keep. Can you export the records in a form another system can use, or only a spreadsheet that looks complete until the migration starts? Do the stable IDs, relationships, timestamps, attachments, permissions and change history come with them? Can you read the data continuously through an API, or only request a one-time dump? Can another system perform the work, or does the capability exist only behind the vendor's screen? What happens to the history after the contract ends?

Those questions belong in the buying decision, not the postmortem.

Europe already turned the problem into law. The EU Data Act, in effect since September 2025, requires covered cloud and software providers to remove contractual and technical barriers to switching, provide open interfaces for many services and make data available in structured, commonly used, machine-readable formats. Europe didn't write those rules because switching was working fine. It wrote them because dependence on one provider had gotten expensive enough to require an exit by law. (European Commission; Germany's Bundesnetzagentur guidance)

American local media companies shouldn't need a regulator to ask the same questions.

The test isn't whether good people can make the current stack work. Good people can make almost anything work. The test is whether the operation can carry the customer from discovery through renewal without retyping the facts, reconstructing the promise or depending on one person's memory.

Keep the CRM if it works. Keep the ad server, the finance system, the email platform and the proposal software that earn their place. Replace the ones that don't, and stop calling the workaround a system.

But keep your own usable record of the customer across all of them. Keep the product definitions. Keep the pricing and approval history. Keep the evidence behind the report. Keep the permissions attached to the people, not buried inside a shared password. Keep a tested way out.

The machinery can change. The company should still remember.

Run the account test

Pick an account you can't afford to lose. Then take the people out of the exercise.

Using the systems alone, can someone who has never touched the account tell you:

What is the customer actually trying to accomplish?
What did we sell, and what did we promise?
What changed after the sale, and who approved it?
What ran, what did it cost us to deliver, and what did the customer pay?
Did it work, measured against the goal the customer named?
What has to happen before the renewal?

Not eventually. Not after three people search their inboxes. Not after the account manager explains which fields nobody uses anymore.

Can the company answer?

If it can't, an agent won't fix the problem. It will move the uncertainty faster: draft the renewal from the wrong scope, explain performance against the wrong goal, apply a current price to an old commitment, and put a polished answer on top of a business that still doesn't know.

If it can, the work changes quickly. The account brief assembles itself because the sources are known. The approved proposal becomes the order without somebody typing it again. Delivery problems surface against the promise that was actually made. The report answers the client's goal instead of reciting platform metrics. The renewal starts before the cancellation email, because the history survived the handoffs.

That isn't an AI strategy. It's a revenue operation that finally knows what it knows.

The agents are arriving either way, and so are the platforms eager to become the place where they work. Local media already learned what happens when somebody else owns the path to the people it serves. Don't make the same deal with the memory of the business.

Sources

Built Revenue · Field Notes No. 14 · https://www.builtrevenue.com/field-notes/the-customer-cannot-live-in-one-persons-head

— Get the next note

The full essay, by email, when it publishes. No filler, unsubscribe anytime.

You just ran the account test on one account, using what you already know about it. The AI Readiness Assessment runs the same question across the operation, with the answers coming from the people who do the work.

Scope an AI Readiness Assessment →

← All Field Notes Connect on LinkedIn →