Once the data is unified and its definitions are settled, it can be queried in plain language instead of through a ticket to an analyst. This step wires that capability onto the layer built in the previous one and puts guardrails on it, so the assistant answers from your reconciled figures rather than improvising from whatever it can reach.
This phase is not sold on its own, and asking will get a no.
An assistant pointed at three systems that contradict each other produces fluent answers that are wrong, and a wrong answer delivered confidently in one sentence is worse than the three-day wait it replaced. The two phases above it are what make this one safe.
Every AI tool promises to answer questions about your business. None of them can, because the answer is spread across systems that contradict each other and no assistant can reconcile a contradiction it cannot see.
The work
What this phase does
- 01
Point the assistant at the reconciled layer only, never at the raw contradictory sources
- 02
Constrain it to the agreed definitions so it cannot quietly invent a fourth version of customer
- 03
Test it against questions with known answers before anyone relies on it
Outcome
A question answered in a sentence instead of three days, from figures that reconcile to your books.
What buyers ask
Why can our AI tools not already do this?
The model is almost never the blocker. The blocker is that the facts live in systems that disagree and nothing has decided which one is right. Fix that and the tools you already pay for become useful the same week.
Is this a chatbot?
No. It is your own data, queryable in plain language, answering from figures that reconcile. Nothing about it is customer-facing.
What stops it being confidently wrong?
It only ever reads the reconciled layer, it is held to the definitions agreed during the build, and it is tested against questions whose answers are already known. It will also say when it does not have the data, which is the behavior most implementations skip.
Do we need the previous two steps first?
Yes. This one is not sold on its own, because pointing an assistant at unreconciled data produces fluent answers that are wrong, which is worse than no answer at all.
Start with a reconciliation
Every engagement enters at the first phase, including the ones that already know which phase they want.
Send your site, your ad accounts, and what you spend. You get back a written finding on where your systems disagree, by how much, and which of your questions the data can answer today.
- Cost
- Fixed scope
- Turnaround
- Within ten business days