Integration
Integration is most of the AI project
Years of partner integration work taught me that the model is rarely the bottleneck. What decides success is getting the right data to the right place, with a named owner.
For more than six years I have worked with technology partners on how their products connect to a large enterprise platform. The pattern is consistent. The products that deliver lasting value are rarely the ones with the most impressive features. They are the ones that fit cleanly into the systems and habits customers already have.
AI makes this more true, not less. A model is only as useful as the context it can see and the actions it is allowed to take. Most of the real work in an AI project is plumbing: access, data shape, triggers, permissions and what happens when something fails.
Five questions to ask about any AI integration
- 01
What is the system of record?
Name the one place the truth lives. The AI should read from it and write back to it, not work from a copy that drifts out of date.
- 02
What triggers the work?
An event (a new ticket, a signed contract), a schedule, or a person asking. Each needs a different design and a different failure plan.
- 03
Who is the AI acting as?
It should never have broader access than the person it is helping. Reuse your existing identity and permissions instead of creating a super-user account.
- 04
What happens when it fails?
Retries, alerts and a human queue for the cases it cannot handle. If the answer is "it won't fail," the plan is not finished.
- 05
Who owns it in six months?
A named person with time and budget to maintain it. Integrations without owners quietly break and take trust in the project with them.
Signs the integration plan is too thin
- "We will export a spreadsheet every week and upload it."
- "IT will sort out access after the pilot."
- "The vendor says it connects to everything."
- No one has said what happens when a record is missing or wrong.
- The success metric is "it works," not a number from the initiative.