Almost every stalled AI project we get called in to rescue has the same shape. The agent works. The demo was genuinely impressive. And it has been six months with nothing in production. The problem is nearly always three layers below where everyone is looking.
There are four layers to a go-to-market automation stack, and they only work bottom-up. You can skip a layer and still get a demo. You cannot skip a layer and get a system.
Layer one: data
Not “all your data”. Nobody has clean data and waiting for it is a way of never starting. What you need is a small, defined set of fields that the workflow in question depends on, and agreement on what each one means.
That second part is the one people skip. We have sat in rooms where sales and finance had different definitions of an active account and neither knew, because the disagreement had been living quietly inside two dashboards for three years. No amount of model quality survives that. The agent will confidently do the wrong thing, at scale, and the first person to notice will be a customer.
The test for this layer: can two people from different teams independently produce the same number from the same question? If not, you are not ready to automate on top of it.
Layer two: logic
Rules a human would agree with, written down before anything runs. If this account crosses that threshold, it goes to that owner. If a document fails validation, the client gets this message. If nobody replies in five days, stop.
There is a strong pull to hand this layer to a model, because writing rules is tedious and models are good at seeming to have judgement. Resist it for anything deterministic. A rule is inspectable, testable, cheap and identical every time it runs. A model asked to do a rule’s job is none of those, and when it drifts you will find out slowly.
Use the model where the rules genuinely run out. Not before.
Layer three: agent
Now the interesting part — and it is a smaller part than the market suggests. The agent handles what rules cannot: reading unstructured text, summarising a call, drafting something in context, deciding which of six situations this message resembles.
Three things make this layer hold up in production:
- Grounding. The agent answers from your documents and cites which one it used. If it cannot cite, it says so rather than filling the gap.
- Thresholds. Below a confidence line it hands to a human, with its reasoning attached. Handover is a feature, not a failure.
- Evaluations. A fixed set of real cases with known-good answers, run on every change. Without this you cannot tell whether last Tuesday’s prompt edit made things better or quietly worse.
The evaluation point is the one teams skip most and regret hardest. Prompt changes feel free. They are not free — they are untested deploys to production, and treating them as anything else is how quality erodes without a single alarm going off.
Layer four: surface
Where your team actually meets the work. This is the layer most projects get exactly backwards: they build a beautiful new interface nobody opens.
The best surface is almost always one your team already lives in. A Slack message in the channel they watch. A field populated in the CRM record they had open anyway. A draft sitting in the inbox they were about to check. If using the automation requires remembering to visit something, adoption decays — not immediately, but within about a quarter, and quietly.
Why the order is the whole point
Build top-down and every layer beneath becomes a constraint you discover late. The agent is impressive in the demo because the demo used three hand-picked records. In production it meets the field that means two different things, the rule nobody wrote down, and the interface nobody opens.
Build bottom-up and something useful ships at every layer. Fixing the data alone makes reporting trustworthy. Adding the rules alone removes most of the waiting. By the time you reach the agent, it is doing the small, genuinely hard part — which is what it is good at, and all you should have been asking of it.
If you want to know which layer you are actually standing on, that is what the first two weeks of our engagements are for. Book a call and we will tell you, whether or not there is a project in it for us.