The Biggest Mistake Banks Make When Deploying AI in Workflows

There’s a pattern we keep seeing across financial services. A team spots a workflow that looks automatable. Something repetitive, document-heavy, rule-adjacent. They bring in a vendor, run a pilot, generate some internal enthusiasm. Six months later the thing quietly stops being used. Or it gets used, but nobody trusts it. Or it handles 80% of cases fine and creates expensive chaos for the rest.

The technology is rarely the problem. The sequencing of approach usually is.

Automation is not the starting point

When organisations think about AI in workflows, they tend to think in binaries. A human does it, or a machine does it. The question becomes: can we automate this? And if the answer looks like yes, the pressure is to move fast.

But there is a lot of territory between “human does it” and “machine does it,” and most organisations skip straight over it.

Assist, augment, automate. That is the spectrum that matters, and the order is not arbitrary.

Assist means AI supports a human who still owns the decision. It surfaces information, flags anomalies, drafts a first pass. The human reviews, adjusts, and acts. Augment means the human and the system are genuinely sharing the work, each doing what they are better at, with clear handoffs. Automate means the system handles end-to-end, with human oversight reserved for exceptions and escalations.

Most firms try to start at automate. That is where the cost savings look biggest on a slide. It is also where things tend to break.

The judgment problem

Automation works when you understand a workflow well enough to encode what it requires. The difficulty is that most financial services workflows carry more embedded judgment than anyone has ever had to articulate, because a human was always there to apply it without being asked to explain it.

Take KYC reviews or client onboarding. On paper it looks procedural: collect documents, verify against data sources, flag discrepancies, approve or escalate. But anyone who has worked in it knows there is more going on. Edge cases that experienced analysts handle on instinct. Patterns that fall outside the rulebook but that someone with a decade of context would recognise immediately. Relationship considerations that quietly shape what good enough looks like for a particular client.

When you automate before you have surfaced that judgment, one of two things tends to happen. The system is too rigid and creates friction for legitimate cases. Or it is calibrated too loosely and becomes a liability when something goes wrong, and then you are explaining to a regulator why no human was more involved.

Neither outcome is a technology failure. It is a design failure.

Why assist is not the cautious option

There is a tendency to treat assist mode as the conservative choice. Something you do while waiting for the technology to mature enough to really automate. That framing misses what it actually gives you.

When you deploy AI to assist rather than replace, you build a working relationship between your people and the system. Analysts start to understand where the AI performs well and where it does not. You accumulate real evidence, decisions made, overrides logged, accuracy tracked, that tells you whether augmentation or automation is genuinely appropriate and for which parts of which workflows.

You also build institutional confidence. The people inside the process feel in control. They are not being bypassed. They are being supported. That matters for adoption, and it matters when things get audited.

In financial services, that audit trail is not a nice-to-have. “A human reviewed and approved this” is a defensible position. “The model decided and nobody checked” is not.

The question most AI projects never ask

Most AI deployment projects start with a technology question: what can this do? The more useful starting point is: what does this workflow actually need?

That requires understanding the workflow before trying to change it. Where does judgment live? Where are the edge cases? What does good look like, and who decides? When something goes wrong, who is accountable?

These are not technology questions. They are design questions. And they tend not to get asked, not because teams are careless, but because the pressure to move quickly and show results makes it feel like slowing down is the wrong instinct.

It is not. Taking the time to understand a workflow properly before deploying AI into it is what separates implementations that hold up from ones that get quietly decommissioned.

Helping teams ask the right questions before they build

This is the work we focus on at Experience Haus. Not selling tools, not running proofs of concept that demonstrate what the technology can do. Helping teams ask the right questions before they build, so that when they deploy, they are building something that actually lasts.

The firms getting AI deployment right in financial services are not necessarily the ones with the most sophisticated technology. They are the ones who understood their workflows well enough to know where AI belongs in them, and at what stage.

If you are working through that question, we would be glad to think it through with you. Read more about our Rewired, Reimagination and Realise approach here.