Why AI deployments in finance stall

June 2026

A deployment lives or dies on one thing: how well you understand the business you're dropping it into. Most die because everyone skips that bit. Someone buys a clever model, points it at a vague problem, and waits for a return that never turns up.

In a big financial firm the build is rarely the hard part, and it's rarely yours to do. You really don't want an investment desk writing its own software. What you want is someone who can drag out what the team actually needs, which is usually something they can't put into words, decide what's worth building, and run the build around them. That's the bit everyone gets wrong.

Then there's the other problem: the people paid to say no. Risk, legal, compliance. Their objection is reputational risk, and nobody wanting to own a decision that might blow up. Almost never the model itself. So half the job is making the decision safe to take. You build the consensus, and you hand the people with a veto something they can sign without losing sleep.

I've worked both ends. I took digital assets into Baillie Gifford from nothing, dealing with the FCA directly and with risk, legal and compliance in-house. And I built the systems that run my own company, so I know what a build actually costs.

The order that works is boring, and it almost always holds. Understand the business. Find the real requirement. Then build. Lead with the tool and you get a nice demo and very little else.

I advise financial firms on putting AI to work. me@yussefrobinson.uk