Capability Without Consequence: How to Manage AI in a Finance Team

Originally published on LinkedIn.

AI has arrived in finance on a wave of hype, much of it pushed by technology vendors whose core claim is that these tools will replace finance professionals. That framing has done real damage, because it has forced the profession into arguing about the wrong thing.

Most finance people I speak to are still stuck at the same question.

Can AI be trusted with complex finance work?

The caution behind that question is not irrational. Hallucination is real, and I have seen it produce work that looked finished and was wrong underneath. Safety researchers have published findings on models behaving in ways their developers did not intend, and those findings get reported widely. Finance professionals are trained to spot risk from a distance and to assume the worst case is the one that will land on them. Given what has been promised and what has actually been delivered, scepticism is a reasonable default.

Just over a year ago I decided the debate was not going to resolve itself from the outside, so I stopped reading about it and started testing. I put Claude on real use cases across modelling, consolidation, data transformation and reporting.

I have now spent more than a thousand hours producing finance output with it. Output I sign, and output that goes to clients and boards.

Here is what that year gave me. Trust is not a property of the tool. The trust question is the wrong question, and it is wrong for a reason that took me a long time to see.

Why trust is the wrong question

AI does something no technology in my career has done before. It thinks. It reasons. It takes calls.

That is where its value comes from. Not from retrieving, not from calculating, but from working out how to approach a problem it has not been given complete instructions for.

Anything that can think cannot be objectively right all the time.

That applies to me, it applies to every professional I have worked with, and it applies to this. You cannot have a thing that reasons and also demand that it be correct every time, because the reasoning is what creates the possibility of error. The two arrive together.

So asking whether AI can be trusted is like asking whether a person can be trusted in the abstract. It is not a question you can answer, and it is not how you have ever run a team. You do not resolve trust with a colleague once and then stop looking. You brief them, you review their work, and you own what leaves the building with your name on it.

Why the profession finds this so hard to accept

Finance has spent thirty years getting good at a completely different kind of technology.

ERPs, accounting platforms, reporting tools. All of them run on rules somebody wrote in advance and hard coded. They have no ability to think, and that is the point of them. They provide structure around data, and the controls that finance cannot function without. They fail the same way every time, and every failure traces back to a line that can be found. Someone in IT fixes the bug and life goes on. The risk is low and, more importantly, the thinking has already been done by a person before the system ever runs.

I spent 2019 and 2020 on a Deltek Maconomy implementation, leading the AP and treasury modules for a multinational group, and by the end I was training teams in other countries to run them. In three years nobody asked whether Maconomy was accountable for a number, because the question is absurd. A person wrote the logic, a person signed the configuration, a person entered the data, and a person signed the report.

That is the model the profession carries into every new technology. Scope it, configure it, write the policy, train the team, sign it off. It is one of the few technology disciplines finance genuinely owns, and it is the reason we get this one wrong. We reach for the process we are good at.

AI needs a different model. Not a different implementation plan. A different mental model of what we are dealing with.

Capability without consequence

I have come up with a model that works for me and my team, and I call it Capability without Consequence.

It has three components.

First, AI has capability that exceeds a finance professional's on specific tasks. Speed, breadth, recall, first-pass construction. Not judgement, not context, and not accountability, but on the mechanical and analytical work that fills most of a finance week it is faster and often better than we are.

Second, it carries no accountability whatsoever. None. It cannot be fired, cannot be promoted, cannot sit in front of a client and defend a number, and cannot be sanctioned by a professional body. I hope that remains the case, because it is a large part of what makes us valuable.

Third, because of the first two, someone has to own the output. Not review it loosely. Own it, in the way you own anything you sign.

That combination is genuinely new. Everything else in a career pairs capability with accountability, and the pairing tightens as you move up. A senior manager owns their calls. An auditor signs an opinion and stakes a licence on it. Even the intern answers to somebody. This is the first thing I have worked with that has one and none of the other.

Once you accept that as the starting condition rather than a flaw waiting to be fixed, the management response becomes obvious.

It widens the gap between senior and junior

The common prediction is that AI will compress the difference between an experienced practitioner and an inexperienced one. Having now trained a number of finance teams, I do not see that at all. I see the opposite.

The more skilled you are at your craft, the better your output with AI. Every time.

Senior people produce better results because they know what good looks like before they see it. They know which numbers to interrogate, where a group structure hides its problems, and which answer is plausible but wrong. They have been through the ups and downs of their own craft, so they know exactly what to check.

I could teach Claude a ten entity consolidation because I know consolidations. Someone who had never closed a set of group accounts could have spent the same hours and produced something that looked correct and was wrong underneath, and would not have known the difference.

The capability is available to everyone. The ability to direct it and to catch it is not.

Which is what worries me about the people coming into finance now. My judgement came from manual work and from getting things wrong in front of an audience. They will have these tools from their first week, and I am not sure where the equivalent education comes from.

I think that responsibility sits with us, the people leading finance teams. It is how I train my own juniors. They get the tools, but they do the work manually first, so they know what an answer should look like before something produces one for them. It also means explaining why we are making a decision rather than just making it, so judgement gets handed down deliberately instead of being absorbed by accident.

I do not have all the answers on this one. It needs more attention than it is getting.

The right question is how do we manage it

So the right question was never whether to trust it. The right question is how to manage something that has the capability I want and none of the accountability I need.

Management of it starts with one shift.

Treat AI as a new hire, not a new tool.

That is the single most important idea in this whole framework. Specifically, treat it as a colleague with the capability of the most senior person in your team and the accountability of the most junior one.

Hold both halves at once and three things change immediately.

How well you brief it. A great deal, in my case. You would not tell your most junior colleague "do the forecast" and expect back the thing that was in your head. You would explain what it is for, who is reading it, what decision hangs on it, and what must never appear in it. That is not prompting technique. It is delegation, and every finance leader already knows how to do it.

What work you give it. The senior half of the framing tells you the scope can be ambitious. There is no reason for me to spend eight hours building a budget model when I can build a better one with Claude in a fraction of that time. But judgement stays with me. When a finance manager in my team brings me a judgement call, I encourage it, and a fair proportion of the time I still change the decision. That is the right relationship here too. Senior colleague, yes. More senior than you on the call itself, no. AI can form a view, and often a good one, and I will use it to shape my own thinking. What I will not do is pass on a judgement I have not independently arrived at. Take the help, own the call.

How much you check it. The junior half of the framing answers this one precisely. However much you would review the work of your most junior colleague before it left the building, review this the same amount. On complex technical output, that is a lot.

There is an upside here that people miss. A correction you give a person has to be remembered, and it competes with everything else they are carrying. A correction you capture properly in the setup of an agent stays captured. You never tell it twice, and it does not drift. So the error rate does not stay where it starts. In my direct experience it falls sharply, provided you are actually building the corrections in rather than re-prompting your way around the same problem every time.

Go far enough with that and something strange happens. The agent starts catching your mistakes. Believe me, it happens more often than I would like to admit.

Why implementations fail

I have watched AI projects fail inside finance functions, and the cause is almost always the same. The tool gets treated as software, so it gets implemented, and the expectation attached to it is that every prompt returns a perfectly accurate answer. The first time it does not, the conclusion is that the technology is unreliable, and within a month everyone is back in Excel.

But look at what that expectation is actually asking for. Guaranteed accuracy every time means no judgement, and no judgement means no thinking. Strip the thinking out and you have removed the only thing that made it worth having. What you are left with is a worse version of software you already own.

If you need deterministic output, use deterministic tools. That is what they are for. If you want reasoning, accept that reasoning comes with a review obligation attached, permanently. That obligation is not a defect waiting to be engineered away. It is the job.

Intelligence cannot be implemented. You can work with it, direct it, learn from it, and get an extraordinary amount out of it. What you cannot do is install it.