In the vendor demo, the AI agent does something impressive. A customer asks for a refund, and the agent checks the account, approves it, and updates the customer's plan, all in a few seconds. The pitch says you can have this running in a matter of hours. The agent in that demo is real. What the demo leaves out is the thing it runs on: a single, clean, up-to-date customer record. Your company does not have one of those.
This is the gap that stalls most contact center AI projects, and it is rarely the AI itself. Modern models can hold a conversation and use tools well enough. The problem is the data they have to work from. This article explains why your customer data, more than the model, is the real blocker, and what to fix before you deploy.
The demo runs on data you do not have
Blake Morgan's interview with Salesforce on unifying the contact centre with agentic AI describes the destination accurately. Agents that resolve a problem instead of routing it. Setup that takes hours because the platform handles the wiring. All of that is true for the world the demo runs in.
Your world differs in one specific way. In the demo, the customer exists once. In your building, the same customer exists in a CRM, a billing system, a help desk, and a loyalty database. At least two of those four hold a different answer about which plan the customer is on right now. The gap between you and the demo is those four records that do not match.
An agent that cannot carry your details forward is not really helping
Mark Levy, who writes Decoding Customer Experience, makes the point sharply. Vendors love the word orchestration. It is meant to mean the systems work together behind the scenes. But if the customer still has to repeat their account number and their problem at every step, nothing was orchestrated. The customer is doing the joining-up themselves, from memory, in real time.
The customer repeats themselves because their details stayed behind. Orchestration, in the way that matters to the person on the other end, is really a data problem wearing a workflow costume. A handoff that carries no context is just a restart with a friendlier name.
Three layers, and the one that matters is at the bottom
Juan Martin Maglione, in the Service Matters newsletter, offers a clear way to break orchestration into layers. There is a process layer that runs the workflow, an agent layer that carries out the steps, and the model underneath. The split is useful. Vendors sell the agent layer, because that is the layer that looks good in a demo.
Every layer depends on the data below it. Point an agent at messy data and it does not fail quietly. It acts fast and confidently on the wrong information. It reads an out-of-date record, believes it, and gives the customer something they were not owed, or refuses them something they were. The model did its job. The data it was handed was wrong.
What data readiness actually means
Data readiness sounds vague. It is really four specific things, and each one has a clear failure when it is missing.
One, you can tell which customer you are dealing with. Miss this and the agent acts on the wrong account, or splits one person's history into two profiles. Two, you have one trustworthy source for what the customer is owed: their plan, what they paid for, what their cover includes. Get this wrong and the agent promises refunds it should not, or denies ones it should. Three, the agent can read the customer's history, so it works from what you already know instead of asking the customer to explain again. Four, access is controlled and logged, so the agent reads only what it should, and you can check later what it saw.
Where to start
Fix them in order. Start with identifying the customer, because everything else depends on knowing who you are serving. Then the entitlement data, because that is what the agent acts on most directly. History and access controls come after. None of this is exciting work. It is the work that decides whether the agent solves problems or creates them.
The vendor will keep showing you the agent layer, because that is what sells. The first phase of a real project goes where the demo never does: into the systems that hold conflicting versions of the customer. That data work is what decides whether the agent helps or harms.