The AI part is often the easy part
Clients are surprised by how much of an AI project is ordinary integration work. Extraction from an invoice might be a fortnight; getting the result into the accounting system reliably, with duplicate protection and reconciliation, can be longer.
That is not a complaint about the estimate. It is the shape of these projects, and a proposal that does not reflect it is understating the work.
What reliable writing requires
- Idempotency — the same record written twice does not create two
- Retry with backoff for transient failures, which are most failures
- A durable queue, so nothing is lost when the target is unavailable
- Reconciliation — a daily comparison of what we sent against what arrived
- Traceability — every written record carries the source document reference
Ask the API question first
Before anyone quotes an AI project, ask your systems' vendors in writing: do you provide an API for creating these records, is it documented, and does it cost extra? That answer moves the price more than the AI does.
Where AI output should not go directly
Anything financial, anything customer-facing and anything irreversible should pass through a rule or a person before it lands. The AI proposes; something deterministic disposes.
That boundary keeps the audit trail intelligible and keeps a probabilistic component out of the path of decisions that must be defended.
Reconciliation catches what monitoring misses
A daily job comparing counts and totals between the AI pipeline and the destination system finds the quiet divergences: records that failed silently, duplicates, values transformed unexpectedly.
It is cheap to build and it is the difference between finding a problem the same day and finding it in a quarterly review.