Blog · AI for small and midsize businesses
Data and integrations: find dependencies that can delay an AI project
An AI project can be delayed by a missing permission or unavailable system owner long before model quality matters. A dependency register turns vague integration assumptions into named checks, required evidence and decisions that can be resolved before committing to delivery.

In this article
“Connect the assistant to our system” is an assumption until someone verifies access, supported interfaces and usable data. A working demonstration with an exported spreadsheet does not establish that the live integration is available. Record dependencies before estimating delivery.
Build a register that requires evidence
| Dependency | Evidence required | Accountable owner |
|---|---|---|
| System interface | Supported method, required fields and a successful permitted test. | System owner. |
| Identity and permissions | Approved roles and confirmation of what each may read or change. | Access administrator. |
| Data availability | Representative records, quality checks and update frequency. | Data owner. |
| Test environment | Safe sample records and a place to verify failure behavior. | Technical lead. |
| External support | Named contact, response route and any necessary service work. | Supplier relationship owner. |
Add status, next action, required date and consequence if unresolved. Use statuses such as verified, awaiting evidence and blocked. “API available” is insufficient; an application programming interface is a controlled way for software to exchange data, and the particular endpoint may not expose the field your workflow needs.
Trace the exact information path
For each proposed answer, identify the authoritative source, identifier used to find the record, required permission and freshness requirement. Check missing and duplicate identifiers. A customer name alone may not identify the correct account; the workflow needs an unambiguous selection rule.
In a hypothetical distributor, an inventory assistant needs product availability. The system owner confirms a read interface exists, but the technical test reveals it returns on-hand quantity without reservations. The dependency is not resolved: operations must define which figure answers the business question and whether the system exposes it.
Until then, the assistant can explain general availability policies from approved documents, while live stock questions remain with staff using the current system. That narrowed scope is preferable to presenting incomplete stock data as a reliable answer.
Check uncertainties before estimating
Separate build work from dependency resolution. A supplier's approval delay should not be hidden inside a precise development estimate. Use a short technical investigation to test the highest-impact uncertainty first. Record what was verified and what remains assumed.
Deloitte's complexity framework includes integrations and data variety. The NIST profile provides a broader reference for documenting system conditions and evaluation.
A useful dependency review ends with a choice for each unresolved item. For example, if the supplier cannot provide a test account, the sponsor may approve a limited offline experiment but should keep live integration outside its acceptance. Record the temporary boundary and the evidence needed to remove it. Recheck dependencies when a system owner or interface changes; a successful test from an earlier environment may no longer describe the current connection.
The sponsor reviews unresolved dependencies and chooses to obtain evidence, narrow scope or change timing. Bring your register and proposed data path to an AI integration assessment before treating a live connection as a settled requirement.