The Comfortable Version of Adoption
The easiest AI project to approve is one that asks the company to change very little. Give people better tools. Help them write, search, and prepare their work faster. Keep the responsibilities and approvals familiar so the rollout does not become a fight about how the business operates.
There is nothing strange about wanting this. A company has customers to serve while it experiments. Changing a working process carries risk, and nobody wants an AI pilot to interrupt the business.
But that reasonable constraint can quietly become the entire strategy. The company wants a large improvement in performance while declaring most of the things that determine performance out of scope.
The Task Is Easier to See
Consider a support request that needs a billing correction. Writing the response is visible work. Finding the account history is visible work. Both are easy places to demonstrate an assistant: here is what the employee used to do, and here is how much faster it happens now.
The request may still wait for Finance to inspect the charge, return it for missing information, and process the correction in its next batch. A faster answer from Support does not necessarily bring that correction any closer.
Each team can improve its own part and still leave the customer waiting. Support has answered. Finance has followed its procedure. The AI team has shown that the model produces useful work. Nobody has had to change the sequence connecting those activities.
I think this is a more useful explanation for disappointing AI projects than saying people need better prompts. The project was defined around a task someone could improve locally. The result the company wanted depended on a process that project could not change.
The Existing Process Is a Proposal
Business processes contain assumptions about what is expensive. Reading a long conversation takes time, so someone summarizes it. Another team cannot access the necessary information, so someone collects it. A specialist has limited availability, so requests accumulate until there are enough to review together.
AI can change some of those costs. An agent may be able to interpret the request, gather the relevant records, and identify what is missing before another person touches it. That gives the company a reason to reconsider where work begins, which steps can happen together, and when a specialist is actually needed.
For the billing request, the new process could check the charge as soon as the customer reports the problem. Cases that meet an agreed correction rule could go directly to execution. A specialist would receive the cases that need judgment, with the conflicting evidence already available. The response would follow the correction or explain the specific decision still pending.
Some of that would use AI. Some would use ordinary software. Part of it would be a decision to stop batching work that no longer needs to wait. The design has to account for all three, even if only one looks impressive in a demo.
The old process should be treated as a proposal made under different constraints. It may still be the right one. But familiarity is a weak reason to preserve it without examination.
Small Does Not Have to Mean Local
This does not require redesigning the whole company before shipping anything. That can become another way to spend months preparing for value that never arrives.
Start with one recurring problem and follow it to an actual result. A single category of billing correction is enough. Include the work across Support and Finance, the cases that go wrong, and the effort needed to recover. The scope is narrow, but it contains the outcome you want to improve.
Then the difficult questions become concrete. Which corrections can run without review? Which checks must remain? Does preparing evidence actually reduce the specialist's work, or do they still need to repeat the investigation? A process that relies on capabilities the agent has not demonstrated is another bad process.
Local productivity improvements are useful too. Saving someone an hour can be worth paying for. The mistake is assuming those savings will automatically produce a different level of business performance without changing anything around them.
What Is Allowed to Change?
An AI team can build the system and still be unable to remove an approval, change a queue, or ask two departments to share responsibility for a result. Someone with authority over that process has to participate in the work. Otherwise every difficult decision comes back as a requirement to preserve the current behavior.
That is the question I would bring to an AI project before discussing models: what is the company actually willing to change if the technology works?
If the answer is only the software, the company has already limited the result it can expect.