The most common AI request we get is not "build us an AI product". It is "we have a system that runs the business, and we think a model could take a third of the manual work out of it". That is a different engagement, and it fails for different reasons.
Start with the boring inventory
Before anything model-shaped, we write down four things: where the data lives, who is allowed to see it, which system is the source of truth, and what the current process actually costs in minutes. That last number is the baseline. Without it nobody can tell six months later whether the project worked.
The feasibility slice
One workflow, real data, two to three weeks. It answers three questions:
1. Is the data good enough for a model to be right often enough to be useful?
2. What does "wrong" cost here, and can we bound it?
3. What is the honest accuracy against a human doing the same task?
If those answers are bad, you have spent a small amount to avoid a large one. We have recommended stopping after feasibility more than once, and the client kept the write-up.
Sitting beside the system of record
The safest pattern is read-heavy: the model reads through an API or a read replica, produces a suggestion, and a person or a narrowly permissioned service applies it. Writes come later, one action at a time, each with an audit trail and a reversal path. Legacy systems rarely need a rewrite to accommodate this — they need one clean integration point and someone patient enough to respect their quirks.
Rollout is a delivery problem, not an AI problem
Feature flag per team, shadow mode first (the model runs, nobody sees it, you compare its output with reality), then a pilot group, then everyone. Keep the manual path working for the whole rollout. The failure mode we see most often is a big-bang switch with no way back.
After launch
Models drift, data changes, someone renames a field. The care tier exists for that: the evaluation suite runs on a schedule, cost and latency are monitored, and there is a named engineer when something looks wrong. Tiers and numbers are on the pricing page.
Frequently asked
Do we need to replace our existing system to add AI?
Almost never. In most engagements the model sits beside the system of record and reads through an API or a read replica. Replacement is a separate decision that should be made on its own merits.
How do you start an AI integration?
With a short feasibility slice: one workflow, real data, a measurable baseline and a working prototype at the end. If the numbers do not justify production, that is a good outcome and a cheap one.
Who owns the code and the prompts?
You do, from the first commit, in your own repository. That includes the evaluation set, which is often the most valuable artefact we hand over.