AI value starts with the workflow

The first question I ask about an AI initiative is what will be better if it works.

That might be a customer getting a useful answer sooner, a team completing more worthwhile work or a business acting on an opportunity it previously missed. The answer should describe a change in the operation.

A demonstration is a useful beginning. It is not the whole case.

Choose a process someone owns A useful initiative has an owner who understands the current workflow and can act on the output.

Who receives the result? What happens when it is wrong? Where does a person take over? What does the customer or colleague experience?

These questions shape the design. They also help identify when a simpler automation would do the job.

Measure completed outcomes Generating an answer quickly is different from completing useful work.

Include the effort spent checking, correcting and integrating the output. Watch whether the team is doing more valuable work or merely producing more material.

Measure the actual workflow and compare it with a credible baseline. A useful first release should tell you where time is saved, where work moves to and what is happening to quality. A headline productivity percentage from another setting cannot answer those questions for your organisation.

Count the operating cost Model usage is one part of the cost. Integration, evaluation, support, exceptions and ongoing maintenance also need an owner and a budget.

Some of those costs already exist in the manual process. Some are new. Compare the whole operating arrangement rather than a model invoice with a person's salary.

Avoid building a financial case around fixed ratios. A low-volume internal assistant and a customer-facing production service have different demands.

Design the follow-through In commercial outreach, for example, finding an interested prospect matters only if the organisation can respond appropriately. The handoff, customer-management process and team capacity are part of the product.

The same applies elsewhere: an issue identified by a model needs a route to resolution; a suggested action needs someone authorised to take it.

Keep learning after launch Use a bounded first release to learn about quality, exceptions, uptake and useful outcomes. Agree what would justify expansion, a change of approach or stopping.

Model capabilities and tools will keep changing. A clear purpose, an owned workflow and evidence of results make it easier to decide when those changes matter.