Published
Choose the task before choosing the technology
Good first candidates are high volume, rule-adjacent, tolerant of review, and currently consuming skilled staff time: triaging enquiries, summarising documents, extracting data from invoices, drafting first-pass responses.
Poor first candidates are low volume, high consequence and hard to verify. Those are the pilots that get quietly shelved after a bad answer reaches a customer.
Ground the model in your own information
A general model that has never seen your documentation will produce confident, plausible and wrong answers. Retrieval-grounded systems answer from your governed content and can cite the source of each claim.
That citation requirement is what makes the output auditable, and auditability is what makes staff willing to trust it.
Design the escalation path first
Decide up front what the system does when it is unsure: hand off to a human with full context, rather than guessing. Confidence thresholds and explicit hand-off rules matter more than model choice.
Govern data, access and retention
Know what data goes to which provider, where it is processed, whether it is retained, and who inside your organisation can see the outputs. Write it down before the first pilot, not after the first incident.
Measure against the baseline you had
Record the current cost, handling time or error rate before deployment. Without that baseline, every AI project can be declared a success and none can be justified for renewal.
The short version
Narrow scope, grounded answers, explicit escalation, measured baseline. That sequence converts AI from an experiment into infrastructure.