Better baselines and end-to-end measures improve AI ROI, says Teachers Mutual Bank’s CIO

Many AI programs can show momentum, but not what actually changed for the business.

In this ADAPT Insider conversation, Dan Chesterman, CIO at Teachers Mutual Bank Limited, argues that the issue usually starts earlier.

Organisations need a clearer view of what they are trying to improve, how that process performs today, and which outcome would prove the investment worked.

Listen to the full episode on Apple Podcasts, Spotify, and YouTube.

Key takeaways:

• AI ROI becomes clearer when organisations start with a strong baseline and measure business improvement against it.
• Use cases, model choice, and governance need to match the decision or process being improved.
• End-to-end outcomes give a more useful picture of AI value than local productivity gains alone.

ROI starts with a baseline

Dan’s view is that AI performance is hard to judge when the starting point is vague.

If an organisation cannot describe current process performance, customer experience, or another measurable business outcome, it becomes much easier to celebrate implementation than improvement.

That is why he puts so much weight on baselining.

Once the current state is visible, AI can be measured against something real.

That also makes it easier to separate activity metrics from business metrics.

Token consumption, model usage, and go-live dates may show movement, but they do not show value on their own.

Clear use cases matter more than broad AI ambition

Dan also points to overapplication as a common problem.

AI can easily become a catch-all label for projects that have weak purpose, poor fit, or an inflated cost base.

Organisations can end up applying more expensive models than the task requires and calling that progress.

His framing is practical. Start with the use case, the decision, or the workflow that needs to improve.

Then choose the level of intelligence, cost, and governance that fits it.

Some tasks need prediction. Some need precision. Some need human judgement at key points.

Once that context is clear, AI becomes easier to apply with discipline.

End-to-end outcomes matter more than local gains

Dan is equally clear that improving one step in a process does not guarantee a better result overall.

A faster task can still create friction elsewhere, or save time that never becomes reusable because it disappears into other work.

That is why he pushes leaders towards end-to-end measures.

Completion rates, customer outcomes, process performance, and cost per successful outcome all give a better picture of whether AI is helping the system work better.

In that sense, AI value comes from what changes across the journey, not from one isolated efficiency win.