We’ve crossed the chasm.
A year ago, AI in finance was a demo. Now it’s in the workflow. We can ship - decks, models, commentary, first drafts of almost anything - faster than we can read what we shipped. Context engineering turned out to matter more than prompt tricks. And the same tools that let us produce more have quietly filled our drives, our inboxes and our feeds with something we didn’t have a word for two years ago: slop.
That’s the real tell that the experiment is over. Not the wins - the mess. When a technology moves from novelty to daily operations, the interesting question stops being what can it do and becomes what is it actually for.
The bill lands at desk level first
Let me be precise about the scale I mean here. Not enterprise AI strategy. Not agentic transformation of the close. I mean desk-level, individual usage - the thing each of us actually reaches for on a Tuesday. Because that’s where the cost shows up first, and quietly.
The pattern is familiar by now. You adopt an AI tool for one genuinely useful thing. It works. You lean on it. It becomes part of your day - more advanced, more recurring, more embedded. And precisely as that happens, the pricing follows you there: more tokens, more credits, a higher tier for the features you now depend on. Somewhere along the way, a headcount line quietly became a licence line.
That can be a perfectly good trade. But if the usage underneath it stays casual - generate, generate, generate - you end up paying real money for piles of output you then have to review. And a model’s output with no professional skepticism applied to it isn’t an asset. It’s a liability with good grammar.
The encouraging half of the picture
Because there’s another half, and it’s just as real.
The most interesting AI work I see isn’t coming from the people chasing the newest tool. It’s coming from people with deep business context - years of accumulated judgment about how their organisation actually makes money, where the numbers lie, what a stakeholder genuinely needs. Not tenure for its own sake. Diversified, hands-in experience. Those are the people rolling up their sleeves, getting into the tools themselves, and building genuinely clever proofs of concept.
The tool isn’t the differentiator - the context is. AI amplifies judgment you already have; it doesn’t manufacture judgment you don’t.
Which is exactly why we’ve reached the next stage. The experiment produced results. Now the harder work begins: scaling and structuring what worked - and deciding, honestly, where AI makes a difference and where it doesn’t.
Technology as an accelerator
Jim Collins made this point a quarter of a century ago and it has aged remarkably well. In Good to Great, his team studied companies that sustained exceptional results consistently over the period of fifteen years and asked (amongst other questions) what role technology played in getting them there. The finding: used well, technology is an accelerator of momentum, never the creator of it. None of the great companies began their transformation with pioneering technology. They got disciplined about their strengths first, then applied carefully chosen technology to what was already working. If a technology didn’t fit, they ignored it.
For us in finance, that discipline has a name: judgment. Our job has never been to deploy the most tools. It’s to decide how to spend the organisation’s scarcest resources - people, money, time - to enable performance and deliver value to the people who rely on our numbers. AI is one line in that allocation. It is not the whole budget.
So before reaching for a new solution, it’s worth visiting the good one already sitting in your Excel ribbon.
Three tools we’re too quick to replace
None of these is glamorous. All three are quietly excellent at exactly the jobs we’re now tempted to hand to an agent.
Power Query: repeatable data cleaning, consolidating files, joining and reshaping exports. You build the steps once; every month you press refresh and get the same result, every time, for free. It’s built into Excel, and every transformation is recorded.
Power Automate: rules-based workflows: approvals, notifications, moving files, scheduled handoffs between systems. It runs identically every time and leaves a run history you can inspect.
Python in Excel: the heavier analytics native Excel strains at: forecasting, scenario modelling, statistics, large-dataset cleaning, richer visuals. It needs a licence and runs in Microsoft’s cloud, but the code sits in the cell - transparent, traceable, and reviewable.
Notice the word that keeps surfacing: auditable. Same input, same output, and you can show your working. That is not a nice-to-have in finance. It’s the job.
Now hold that against AI in the same grid. Copilot’s agent mode in Excel became generally available in early 2026 - it can plan and execute multi-step tasks directly in your workbook, and you can even choose which underlying model runs it. It’s genuinely useful. But it is non-deterministic - the same prompt can produce different work each time. It is metered - you pay per run. And its output has to be verified every single time. Microsoft is now bolting verification layers onto it precisely because confidently-wrong answers are the known failure mode.
Put the two columns side by side and the rule almost writes itself.
The rule
Reach for Power Query, Power Automate or Python when the task is deterministic, recurring and has to be auditable - month-end mechanics, reconciliations, the reports that must run the same way forever.
Reach for AI when the task is ambiguous, exploratory, one-off or language-shaped - drafting commentary you’ll edit, interrogating a messy dataset you don’t yet understand, getting a first version of something you can’t yet structure.
Sending an agent to re-clean the same file every month is paying a consultant’s day rate to press a button you already automated.
The point isn’t less AI
It’s the opposite. Being deliberate about where AI earns its cost is exactly what lets you spend the budget (and your attention) where it genuinely compounds. Technology as an accelerator, not a creator. The momentum still has to be yours.
This issue’s action
Block one hour this week, and protect it - no meetings, no notifications, just you and a blank page. List out the things you actually do with AI at your desk right now, and put each one into one of two columns: the deterministic work, where the same input should always give the same output and you need to be able to show your working, and the non-deterministic work, where the task is genuinely ambiguous, exploratory or one-off. Then, honestly, decide which tool belongs to which - where ‘classic’ solution is quietly the better answer, and where an AI model is worth what it costs. You won’t get every line right first time, and that’s fine. The point isn’t a perfect map. It’s that you’ve stopped defaulting to AI for everything and started spending it on purpose.