AI Growth Architecture

Cost follows scope, not the size of your toolset. What drives it is how many source systems have to be connected and how clean their records are, how many decisions the system is allowed to take without a person, and how much of the work your own team picks up after handover. Data work is usually the largest and least visible item, because records that do not reconcile have to be corrected before any model can use them. Ask for a fixed price on the first module and a separate estimate for the data layer, so the two are not mixed.

Start with a diagnosis rather than a purchase. List the stages of your funnel and measure, for each one, how many people are lost there and how long the stage takes. The stage with the largest absolute loss is the one to work on first, even when another stage looks worse in percentage terms. Then check whether the data needed to explain that loss is already being recorded anywhere. Where it is not, recording it is the first piece of work, and the model comes later.

Records have to resolve to one customer. Where the same account exists under different names in different systems, a model counts one relationship as several and its output drifts accordingly. Outcomes have to be recorded as well, because a scoring model learns from deals that were won and lost rather than from opinions about why. And definitions need a named owner, so a qualified lead means the same thing in marketing as it does in sales. Without those conditions a model mostly reproduces the inconsistencies already sitting in the CRM.

Judge the funnel, not the model output. Useful indicators are the time between an inbound request and the first response, the share of scored leads that sales actually follows up on, and whether the scoring separates the groups at all: do high-scoring deals close more often than low-scoring ones? Record those values before you start, because only your own figures will show later what changed. Establishing which channel genuinely contributes is a separate exercise, and it needs a holdout: a share of the audience you deliberately leave untouched.

The most common failure is buying before designing. A platform is chosen, the problem it should solve is defined afterwards, and the result is one more disconnected system. The second is connecting everything at once, which stretches the timeline until nobody can point to a result and internal support runs out. The third is automating without an escalation path, so unusual cases are handled wrongly and quietly, and nobody notices until a customer does.