“Can we ask AI whether this customer is active?” sounds straightforward until three systems provide three reasonable answers.
The billing system might mean paid up. The CRM might mean someone is still pursuing the relationship. The customer portal might mean the login is enabled. None of those definitions is necessarily wrong. Combining them without a decision is where the trouble begins.
A technical leader’s first job here is to make the business question precise and identify who can settle it. Adding an AI interface before that work risks making disagreement much easier to access.
The issue has a local context, too. Refresh Miami’s October 2024 report on Matia’s $10.5 million seed round described the startup’s focus on bringing fragmented business data together.1 The investment is an industry signal, not proof that one platform solves every company’s data problem.
Start with the decision someone is trying to make
For a fictional service business, “active customer” might determine whether a customer can request work through a portal. That is a policy decision with operational consequences.
The answer needs an owner, a definition and a route back to the information used to reach it. Think of a technical leader as helping write the map legend before installing satellite navigation. A very precise voice can still send everyone confidently toward the wrong meaning of “open.”
Separate status labels from business meaning
Ask the service team what they need to decide. Then ask finance, sales and operations how their records contribute.
Perhaps portal access requires an enabled account and an approved service agreement, while an overdue invoice triggers a review rather than an automatic block. That is a possible policy for our fictional business, not a rule every company should use.
Document the agreed meaning in ordinary language. If departments disagree, make that disagreement visible to the person responsible for resolving it. A field mapping cannot substitute for a business decision.
We build AI agents with clear jobs and sensible limits. See how.
Trace where the answer gets delayed
Once the meaning is clear, follow the information between systems. A correct source record may still reach the portal late because a scheduled exchange failed or because the receiving service hasn’t applied the update.
Record when each relevant fact was last refreshed and how delays are detected. A timestamp without a definition can be misleading: was it the time the record changed, the time it was copied or the time someone opened the dashboard?
The team should be able to distinguish “the customer is not eligible” from “we cannot currently establish eligibility.” Those answers lead to different actions.
Make the next investment about a visible constraint
This is where fractional CTO work can be useful: connecting the technical detail to a decision about sequence, ownership and budget.
The next step may be an AI assistant. It may instead be a missing integration, a better exception view or a policy that has never been written down.
Decide what to repair and what to contain
An old integration may be awkward but dependable. Another may fail frequently and create manual work every week. Treating both as a single modernization problem makes prioritization harder.
Martin Fowler’s technical-debt discussion is helpful because it distinguishes how design compromises arise and asks whether the cost of carrying them justifies repayment.2 Not every imperfection needs an immediate rewrite.
For our sample business, fixing the unreliable agreement sync may matter more than replacing a stable reporting tool. Agree on the outcome to improve and the evidence that will show whether the change worked.
If this was useful, try Search your business knowledge before asking AI to improvise.
Give the eventual AI answer boundaries
If the underlying decision becomes reliable, an AI assistant can explain it or help staff locate the relevant information. It should cite the records or policies behind the answer and acknowledge when evidence is missing.
The application’s permissions and business rules still need to control access and actions. An assistant explaining an account restriction should not gain authority to remove it simply because the customer makes a persuasive case.
Keep a practical handoff to staff. An honest “this needs review because the agreement information is stale” is more useful than a polished answer built around an assumption.
I like AI projects that begin with this kind of groundwork. It helps the business discover whether the obstacle is interpretation, connectivity or an unresolved decision.
By the time the assistant arrives, someone can explain what the answer means and who stands behind it. That is a much better welcome than giving it access to three contradictory systems and wishing it a pleasant first day.