Walk into a private bank today and the model is already in the room. It drafts the client’s quarterly report, suggests a rebalancing, summarises a decade of the relationship before the review meeting. The relationship manager reads it, mostly agrees, and sends it on.
At some point a client will ask a plain question about a number or a recommendation: who decided this? The honest answer has to be a person. Most banks have wired AI into the front office faster than they have settled who that person is.
The real risk is undefined decision rights
The recurring failure I see has little to do with the algorithm. It comes down to undefined decision rights. What is the tool allowed to do on its own? Who reviews anything that reaches a client? Who answers when the client pushes back?
Go deeper with GlobalData
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
A capable model sitting on top of unclear ownership produces fluent output that nobody in the building can stand behind. My SSRN work on AI governance keeps landing in the same place. These systems fail for organisational reasons, because the firm never decided who owned the result.
Adopt AI in the right order
Banks tend to deploy AI into client-facing work and sort out the governance afterward, usually once something has gone wrong. The better sequence starts before the tool is switched on.
Map the workflow and rank tasks by what is at stake. Client-facing judgments, a performance figure, a suitability call, a tax assumption that flows into advice, get a hard human check every time. Internal drafts such as a meeting recap or a first-pass research summary can run with lighter monitoring and a periodic audit.
Decide who owns each client-facing output before the tool goes live. Settling that after the first complaint is too late, because by then the client is already holding a number that no one in the firm chose to put a name to.
What good governance looks like in practice
Good governance in a client-facing setting is concrete, and it has little to do with grand policy documents. Favour tools that show their reasoning and cite a source over tools that simply assert an answer. When an output is going to a client, verify it against that source. A fluent answer and a correct answer are different things, and the gap between them is where reputations go.
Keep a record of how an AI-assisted answer was produced, so it holds up months later when a client, or a regulator, asks how the firm arrived at it. Put a named person on sign-off for anything that counts as advice, with dual control where the sums warrant it.
Watch for automation complacency, the slow drift where a tool is right often enough that the team stops checking. I have written about that drift and its companion, skill erosion. Both arrive quietly, and both are governance problems before they are technical ones.
Build, buy or partner follows accountability
A firm can buy a model, or bring in a partner to run one. It cannot outsource the client relationship. When the advice is challenged, the firm answers for it. The supplier stands behind a contract, at most.
So, accountability should drive the sourcing decision rather than follow it. Build where the bank needs control of the logic and an audit trail it can defend in front of a client or a regulator. Buy or partner where the task is lower-stakes and well bounded, a document summary, an internal search, a first draft that a person will rework anyway.
I have watched the pendulum swing. Some firms that rushed to buy are moving client-facing work back in-house, once they price in who carries the liability when it goes wrong. Control of the advice and control of the system tend to travel together.
The questions to ask before going live
If a wealth firm takes one practical thing from this, make it a short list of questions to put to any AI supplier before it touches client work. Can the system show its reasoning and cite the source behind an answer? Can we log every input and output, and keep that record for as long as the advice lives? What are its known failure modes, and have they been tested on data that looks like ours? Who signs off on client-facing output, and can a person stay accountable inside the loop rather than rubber-stamping it? Where does our client data go, and who else can see it?
A supplier who cannot answer these cleanly is telling you something. The firms that ask early tend to design better systems, because the questions shape what gets built.
Accountability is the first design decision
Come back to that client asking who decided. The bank that settled who owns the decision and who answers for it, before it chose a tool, can give a straight answer and show its working. The bank that bought the tool first is still hunting for the person whose name goes on the advice.
AI will keep moving into the front office. The firms that do well will be the ones that treated accountability as the first design decision, the thing they fixed before they argued about which model to run.
Dr. Leigh Coney is the founder of WorkWise Solutions.
