This article is based on a Q&A with Yolm CEO and Co-Founder Ralf Haug to questions and Greg Kihlström for The Agile Brand Guide’s Expert Mode series.
Most marketing leaders have made a quiet peace with their data. The CDP calls a customer a loyal buyer while the support desk has her flagged for an escalated complaint, and somebody on the team just knows which version to trust. We’ve treated that as the cost of running a modern stack. Annoying, but survivable. The plan for AI, at most companies, has been to put agents on top of the tools we already own and let them get to work.
Ralf Haug, CEO and co-founder of Yolm, thinks that peace is about to break. Yolm builds business software that runs a company’s operations and reporting on a single live database, so Haug has an obvious stake in what follows. But his argument is plainer than the “business ontology” vocabulary consultants have been attaching to the post-ERP future. A clean data foundation, he says, means your business “exists in one place, live, instead of as a dozen partial copies that have to be reconciled.” The reason it’s suddenly urgent has less to do with the AI than with the people who’ve been covering for those copies.
Your Team Squints at the Dashboard. Agents Don’t.
Haug traces the urgency to a habit most marketers don’t realize they have. For two decades, fragmented systems worked well enough because people quietly corrected for them.
Why is it urgent now? Because for twenty years, people absorbed the inconsistency. An experienced marketer knows the dashboard runs a day behind and mentally adjusts. An AI agent doesn’t. It treats whatever it can see as the truth and acts on it at machine speed. The moment you put agents into the stack, every quiet disagreement between your copies becomes an input to an automated decision.
The human correction layer is almost never written down. Nobody documented that the “at-risk” segment lags by a day, or that returns don’t reach the CDP until a weekend batch job runs. Gartner forecast in January 2026 that 60% of brands will use agentic AI for streamlined one-to-one interactions by 2028. One-to-one is exactly where a stale record hurts most, since there’s no segment average to soften the error. Before any agent goes live, sit down with the people who “just know” which number to trust and get their workarounds on paper. Each one marks a spot where an agent will act on the wrong copy. That list is your readiness gap, and it’ll probably be longer than anyone expects.
Bad Facts Pass Straight Through Your Controls
Haug splits the failure into two kinds. The first is easy to picture: an agent recommends the product a customer returned last week, or sends a win-back offer to someone with an open escalation because the support desk isn’t connected to the CDP. “It isn’t hallucinating,” he notes. “It is faithfully reporting a copy.” The second kind worries him more.
A person makes that mistake once, notices the confused reply, and stops. Human hesitation was the safety net: the little itch that says check before you promise. An agent has no hesitation. It makes the mistake a thousand times before lunch, and every one of those touches is a real customer forming a real impression of your brand. Worse, none of it trips an alarm. An agent acting on the wrong copy doesn’t bypass your controls; it passes straight through them carrying bad facts, so everything downstream is logged, approved, and wrong.
Most AI governance conversations we sit in focus on the model — hallucination rates, mostly, and brand voice. Haug’s point is that a perfectly behaved model can still do brand damage at scale while the audit log looks spotless. Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. Haug reads that last reason as a foundation problem the agent exposes faster than any dashboard could. That’s his reading, not Gartner’s. Still, it argues for a new line in the risk review: not just “what if the agent says something wrong?” but “what if it says something accurate about a record that’s three days old?” Ask your martech lead to put the refresh schedule for every system an agent will read on one page. If the CDP and the email platform update on different clocks, you already know where the first incident will come from.
The Retrofit Holds Until the Agent Starts Deciding
Here’s where Haug’s commercial stake matters. Yolm’s proposition is that fixing the foundation beats bolting agents onto an existing stack, so we asked him to make the honest case for the retrofit first. He did, and it’s more generous than you’d expect from someone selling the alternative.
Your stack represents years of investment and trained teams. The integrations, however hard-won, work. … And a whole class of AI use cases runs fine on that stack today — drafting copy, summarizing calls, answering questions about a document. Those tasks tolerate imperfect data because a human reviews the output before it touches a customer. If that is where your ambitions stop, retrofit away, genuinely.
It breaks the moment the agent crosses from drafting to deciding.
A retrofitted stack can only feed a deciding agent a copy of the customer, and a copy, in Haug’s words, “can describe your business. It can’t run it.” Governance cracks too. Bolt one agent across six tools and you’ve created what he calls “a seventh actor” whose access is the sum of six permission systems, which makes role-based access control very hard to audit. His test for any vendor claiming to be AI-native is to ask where the agent reads from. An extract, a context layer, or “we sync with your systems” means a copy. Yolm’s architecture is built to pass that test, so weigh the advice accordingly. It’s still a fair question for every vendor on your shortlist, his included. So sort your AI roadmap into two columns. The drafting column can run on what you have. The deciding column is where the data architecture conversation with your CIO starts, and it’s where legal will ask what exactly the agent can see and do.
The Test Costs an Afternoon, Not a Rebuild
Most CMOs aren’t rebuilding anything this year. For them, Haug starts with one question: “Which version of this customer will the agent see — and how old is it?” After that, the advice gets concrete.
Then run the test on one real record. Pick a customer who did something yesterday — placed an order, opened a ticket, made a return — and trace that event through your stack. How many systems hold it? Do they agree right now? Which one would the agent have read? That exercise costs an afternoon, requires no rebuild, and tells you more about your AI readiness than any vendor deck. If the copies disagree, don’t let an agent act on them unsupervised. Scope it to drafting and keep a human between it and the customer until the foundation can carry it.
It’s cheap advice from a founder who sells the structural fix, and it holds up no matter what you buy next. We’d push it one step further. Run it on two customers, one with a recent return and one with an open ticket, with marketing ops, the CDP owner and someone from support in the room. Write down the lag between systems in hours. A CFO can actually weigh that number. Haug pairs the test with two follow-up questions: what exactly is this agent allowed to do, and where is the log? “If nobody can answer either one in a sentence, the governance isn’t there yet.” Until both answers exist, keep the agent in drafting mode, with human-in-the-loop review on anything a customer will see.
Haug has a line for the quiet peace marketing leaders have made with their data: “Your dashboard should not be a memory.” A dashboard running a day behind used to be something the team worked around. Agents take the working-around away.
So the AI conversation in your next planning cycle shouldn’t start with which agent to buy. Start with the trace. If a customer’s Tuesday return shows up the same way in every system an agent would touch, you’ve earned the right to let it decide something. If it doesn’t, the first AI line in next year’s budget belongs to the foundation underneath it.