Are Humans Optional in AI in Finance? The 3 Gates Every Finance Team Needs

Full transcript

Lightly edited for readability. The audio is the original.

0:00Why AI in finance is not delivering the ROI it promised

AI is failing. 80% of projects not only have no ROI, they have a negative ROI. And the reason? We've been skipping everything that has made finance successful for 30 years.

Nobody in finance approves a $4 million capital project by reading the final page. You wouldn't do it. Neither would your CFO.

Capital projects have stage gates. At each one, there's a named person or a committee, a test they apply, and the right to stop the whole thing. And that's how anybody trusts the delivery of a number that big. And nobody in finance questions that. It's just how we work.

Now go and look at the AI workflow your team turned on last month. Human in the loop is one person at the end reading the output, and that's the entire control. So in this episode: what happens when you put stage gates on an AI workflow, and who should be standing at each one?

Welcome to Excel at AI in Finance, the show for finance pros who want to get home sooner and be the ones their company can't afford to lose. I'm Mike, I run F9 Finance, and I'm a senior finance leader working in the trenches. Each week, I share the AI tools, workflows, and mindsets that get you where you want to be. All right, let's get into it.

1:14The year the optimism vanished from the AI conferences

So every year I have the great opportunity to listen to a bunch of smart people talk about the latest developments in AI and finance automation. I really look forward to it. Now, two years ago, it was all optimism about how generative AI was going to change everything we do in finance and data analysis. Then last year, it was all optimism about how AI agents were going to be able to run these multi-step processes and take so much work into the tool and off our plates.

And this year, I was waiting for the optimism and I got something else entirely. Presentation after presentation about how AI isn't delivering. A lot of ways to fix what wasn't being delivered. But all in all, it was about human connection, of all things, and then where AI is going off the rails. The optimism was completely gone.

2:11The client deck workflow that broke at every stage

And it got me thinking about some projects that I've seen go completely off the rails. I was working with a client who shall remain nameless for this purpose, and I was helping them with an AI workflow that just wasn't delivering the results they were looking for. They were looking to really push a button and have an entire deck pop out on the other end.

Now, this deck would consolidate three different sources of data together, totally doable by AI. It would do the analysis and the judgment around what needed to go in the deck. And then it would produce the first pass of the deck before going to a human.

Their expectation was still human in the loop. It needed to be polished. It needed to be cleaned up. But that was the workflow.

What they were getting back was rarely anything that was even usable. So going back to it, I started off by looking at the data that was coming in. Because they'd stopped having a person looking at the data, they were missing some major issues that were feeding in from the data sources. They were no longer looking.

And because the AI was trained that this is the data, it was taking that as the rule, as the guidance, as the context. But the context had gone off track and no human was looking at it. Then it was taking this data that had gone completely off track, and it was making judgment calls on what to include, what not to include, and the story to tell.

And again, at that point, we not only had data that wasn't where it needed to be, now we had a story being built on that data that may not have always been the story, or the way that that story needed to be communicated to people. And again, no one was looking at that point. It was fully produced, and then it went into a deck handed off for someone to look at.

So we've got data that doesn't make sense. We've got a story that's gone off the rails based on data that doesn't make sense. And then we bring the human in to try to resolve this.

And this is a perfect example of what everyone was talking about with how AI is letting us down. We've asked too much of this technology that still has a lot of ways to grow in its development cycle. We've stopped paying attention to the things that really matter because the tool can probably do it. Or honestly, we were promised the tool could do it. All that optimism I was talking about.

And every message I kept hearing was about how the more of a process you ask the AI tool to automate, the less you're going to like the results. So how do we get all of the automation we're looking for, but also like the result that comes out? And that's where we go to a concept that finance has been working on for a really long time. It's a completely different way of thinking about human in the loop, and I'm going to walk you through it step by step.

5:20Human in the loop is the least specified idea in AI

Now, human in the loop is the most agreed idea on AI right now. I'm not saying it's not there. Every vendor says it, every policy says it, and it is also one of the least specified ideas I've come across in finance. When a finance team says human in the loop for AI, almost every time they mean that there's one person looking at a mostly finished thing at the very end, just like the client I mentioned earlier.

5:42What a stage gate is, and how finance already runs them

Now, finance worked this out on capital projects a long time ago. You don't hand somebody a $4 million check and then check the results afterward. And a gate is four things: there's a fixed point in the process, a named owner, a test that the owner applies, and the complete authority to stop it right there.

Now, in a capital project, we have three gates that are pretty much universal across every project. The first gate is the concept gate. Before any amount of significant time or money is spent, the concept: is this worth anybody's time at all? Should we be doing this?

Does this add business value? Does this add more business value than other concepts we have? Do we want this? Do we need this?

There's a business case gate. Now that we've done some initial development, we've done some research, maybe we've spent a little bit of money as a proof of concept. It's the business case gate. Do the numbers hold up? Do they substantiate making our investment in this business case versus another business case?

And then the funding gate. Are we committing? And who signs the check? And in bigger capital projects, there could even be additional gates.

We stage gate capital projects to make sure that they're delivering as we go. It's a way of helping eliminate sunk cost fallacy, where we over-index on the fact we've already spent money and we are afraid to walk away if things aren't working out. So sometimes there could be multiple funding gates or multiple business case gates, where we keep reviewing as the project builds out to make sure we're on the right track.

At every single one of these gates, the committee or the person in charge of the project, depending on the size and how the company works, has the authority to stop the project. At every stage, someone is asking questions, looking at the budget, looking at the development, looking at maybe changes happening outside the company, if it still makes sense. People are evaluating and can shut it down, and are given the authority in advance to do so.

This process works. It's why, as I'm describing these steps, if you've worked with capital at all, you're nodding your head along, because these are pretty much always present. It's universal across finance for anyone doing large-scale capital projects.

8:02The 3 gates every AI workflow needs

So with that in mind, let's apply the same thing to AI. So for AI, I want to talk about three gates. Now, just like capital, you can have more, but these are three gates that need to exist in every single AI workflow: the input gate, the logic gate, and the release gate.

Now let's start with the input gate. This is about the source, the data. Is the data what we think it is? Is the data coming through as we expect?

Is it the right data? Is it the right scope of data? Are the systems functioning?

Now for this gate, I'm not saying that you need to check this every single time the tool runs. If this is something that's running hourly or daily, that doesn't make sense. But what it means is that there's someone who's accountable, that the data flow coming into the AI tool is consistently being checked, consistently being audited.

The things we think are going in are going in. The quality is there. The accuracy is there. The AI tool, as it's designed, can rely on this data. Somebody needs to have the authority at any point to pause or stop the AI workflow when this isn't the case.

If it's an AI workflow that someone is physically running, if you're sitting in front of a Claude Cowork or a ChatGPT Work and you're running the process, build this into all of your processes. After the data collection and curation phase, review it from a high level. If you're looking at your company and you do $5 million of revenue a year, make sure that you're looking at a summary view of a full year and it's $5 million, or 6 months and it's $2.5 million. You see where I'm going with this? Build these into the processes to make sure that the data is giving you what you need.

Now, the next one is the logic gate, and this is the one that is most often skipped. This is the owner of the context.

Does this match what we know? Does this match the people? Does this match the storytelling? Is this bringing in what we need?

This is the phase where we now know and can assume and trust the data that's coming in, but it's about the story. It's about the way the work is being approached. Again, you don't always have to have this running every single time the process runs if it's a frequent hourly or daily process.

If you're running this, again, in Claude Cowork, or you're running this in ChatGPT Work, this is where you need to be checking the output before it does the generation. If you're going to build a PowerPoint, this is where you're asking for an outline, so you can review the outline and make sure the context is there.

This is where, if it's a recurring process, somebody's checking the context: that the business context hasn't changed, that the people in the room haven't changed, that there's not one-off events that we need to take out of the model's knowledge or put back in the model's knowledge. Does this match what's really happening? And somebody needs to own that and be able to pause at any point.

Then there's the release gate. And this is most often where we have the gate today.

When the work is produced, is this ready to be released? What happens if it's wrong? And what amount of reviewing do we do to make sure it's not wrong? It's the most critical part of the process.

We often talk about how AI should shoot to do 80 to 90% of a task. It's not really capable of 100. That was loud and clear as I listened to AI experts and field experts talk about this over the last week. AI is not delivering 100% because there's too many places where the humans still come in, where the messy data still comes in, where the context changes, that they just can't reliably deliver that.

So this is where you decide. Is the AI work good enough to do final tweaks and proceed? Is it accurate? Is it ready to go? Or do we need to go back and rerun this through a process, or in some cases hand it to a human entirely?

This is the person with final accountability to the quality and accuracy of what's coming out, just like the final person who allows the money to be released to move forward on a capital project. We have that accountability elsewhere in finance. It's time we put it here, so AI tools can start delivering the value that we're all looking for.

12:17A model can produce a number, only a person can own it

A model can produce a number, but only a person can own it. Finance worked this out on capital projects 30 years ago. Nobody approves $4 million at the end. You put gates in the middle, a name on each one, a test at each one, and the right to stop at each one.

An AI workflow should be treated like a capital project that runs continuously. Three gates, three names, that is human in the loop. And finance has been building it this way for 30 years.

And this is hard for a reason that has nothing to do with technology. Drawing the three gates takes 20 minutes with a pen. Telling somebody they're allowed to stop your workflow, and then telling your boss it can be stopped, takes something else entirely.

12:593 actions to take this week

Okay, so what do you do with this today? Three things, about 20 minutes, and they'll tell you whether you have a loop or a signature. Action one: take your highest stakes AI workflow and draw it as three points instead of one (where the data goes in, where the answer comes out, and where it reaches somebody who's gonna act on it).

Action two: put a name on each point, and next to the name, write the one question that person is answering. Is the data what we think it is? Does this match what we know about the business? And what happens if this is wrong?

And action three: tell one of those three people, out loud and in front of whoever runs the workflow, that they are allowed to stop it. A gate that nobody can stop is a status update.

The finance pros who build those gates don't just keep their seat. They become the ones their company can't afford to lose.