Last autumn I opened a terminal window and had very little idea what I was doing. I had spent thirty years running charity finance teams, and a black screen with a blinking cursor was not where any of that experience had prepared me to be. It is, in hindsight, the single most useful thing I have done for my working life.

AI Appreciation Day falls on 16 July each year, and its organisers describe it as a global moment of pause, to reflect on the minds we have created. A day set aside for appreciation could easily become a day for advertising, and the sector has enough of that already. So this is a reflection rather than a pitch: what genuinely changed, what I would still be careful about, and what I think the next stretch looks like for charity finance.

It started in a terminal

The tool was Claude Code, and it was not designed for me. It was designed for developers. I spent the first few weeks doing things badly and slowly, which turned out to matter, because working that close to the machine teaches you something that a polished interface hides.

The lesson was this. The quality of what AI gives you is set by the context you give it, far more than by which model you happen to pick. Your chart of accounts. Your fund structure. Your coding rules. The way your organisation talks about restricted income. The judgement you apply without noticing you are applying it, written down so something else can follow it. Get that right and most capable models will do creditable work. Get it wrong and none of them will, however impressive the demonstration looked.

That idea is not mine originally — I owe it to Phil Hewlinson at Multiply Academy, who put it more sharply than I could — but nine months of practice has made it the thing I would defend hardest. It also explains why so many finance teams try AI once, get something generic back, and quietly conclude it is not for them. They were not given a poor model. They were given no context.

Most disappointing AI output is not a failure of the technology. It is a conversation with something that has never been told anything about you.

Then it became work

Around the turn of the year that way of working expanded into Claude Cowork, and the difference was immediate — not because the underlying capability changed, but because a finance team could reach it without living in a terminal. I still use both, and the choice between them is a question of who is doing the work rather than of one being better. The context underneath is the same either way, which is rather the point.

Over the last six months the output has stopped feeling like a demonstration. A bank reconciliation, from raw files through to a drafted journal, with the exceptions flagged rather than buried. A payroll journal split by cost centre and fund. Accruals and prepayments. A management accounts pack with variance commentary that reads like a person wrote it, because it was built from the organisation's own numbers and its own way of explaining them. None of that is speculative. It runs.

The pattern underneath it is consistent. The work is designed once, carefully, with a finance professional's judgement encoded into it. Then it runs monthly, and the review takes a fraction of the effort the original process took, because you are no longer re-performing the work — you are checking that it agrees to a control total, that it has flagged what it was unsure about, and that the narrative says what the numbers say.

What I appreciate is not the model

If AI Appreciation Day is about reflecting on what we have created, then my honest answer is that the part I value most is not the technology at all. It is what happens in a room when a finance team sees this land on their own numbers for the first time.

Charity finance teams face two distinct barriers, and neither is scepticism. The first is time. Teams are too busy running month-end the way they have always run it to find the headspace to improve it, which is as circular a problem as finance has. The second is imagination — not a shortage of it, but a shortage of evidence. Most people I speak to are already persuaded by AI in principle. What they have not seen is it working in a finance function that looks like theirs, on data as awkward as theirs, under a governance regime as real as theirs. Without that picture, it can be genuinely hard to know where to start.

The moment the picture arrives, the question changes. It stops being "could this work?" and becomes "what else could we do?" I have watched that switch flip more times now than I can count, and it remains the most satisfying part of the job.

What I would still be careful about

Appreciation is not the same as enthusiasm without limits, so it is worth being plain about the parts that need care.

Where I think this goes

Predictions age badly, so I will keep these to the ones I would still stand behind if I am wrong about the details.

The tools will keep leapfrogging one another, and it will matter less than people expect. Anyone certain that today's best model is permanently best is selling something. The context you assemble — your rules, your structures, your worked examples — transfers between tools. That is the durable asset, and it is the one charities should be investing in.

The close becomes continuous rather than monthly. Once reconciliation and coding run to a schedule rather than to a heroic week, the argument for a five-day close as a fixed feature of finance life becomes hard to sustain. What replaces it is not a faster close. It is finance teams spending their month on the questions that were always more valuable and never had room.

Governance catches up, and the charities that move early will be the ones setting the terms. A one-page AI use policy naming the approved tool, the data it may touch, and who signs off is not bureaucracy. It is the thing that lets you say yes with confidence.

The agentic shift lands in finance later than the headlines suggest, and more usefully. Not an AI that runs your finance function, but a set of well-scoped pieces of work that run reliably, hand back exceptions, and know when to stop. That is a considerably less exciting story than the one being told, and a considerably better one for anyone accountable for the numbers.

The ambition underneath it

I should be direct about why any of this holds my attention, because it is not the technology for its own sake.

Every charity I have worked in has faced the same arithmetic. Income is hard-won and often falls short of what the work needs. Costs rise regardless. And a real share of what is raised is consumed by the machinery of running the organisation — necessary machinery, staffed by people doing careful work, but machinery all the same. Every hour a finance team spends re-keying, re-checking and reconciling by hand is an hour funded by donors who gave for something else.

My ambition is straightforward. Help charities adopt smarter ways of working, so that more of what they raise reaches the people they exist for.

That is not an argument for smaller finance teams. I have never once made it and would not. It is an argument for finance teams whose time goes to judgement rather than to keying — better forecasts, sharper questions to budget holders, a clearer picture in front of trustees before a decision rather than after it. In a sector where the difference between a good and a poor financial decision can be measured in services delivered, that is not an efficiency story. It is a mission one.

Nine months ago I could not have told you any of this with a straight face. I would have been guessing. What has changed is that I have now done the work, in real finance functions, with real constraints, and I am confident about what the tools can do and clear-eyed about what they cannot. If there is something worth appreciating today, it is that a Finance Director with no technical background can now reach this at all — and that the barrier to entry for a small charity is closer to curiosity than to capital.

If you are somewhere near the start of that, I would genuinely enjoy talking it through — including the parts where I would tell you to be careful.

Sources and verification: AI Appreciation Day details (16 July annually; described by its organisers as a global moment of pause, to reflect on the minds we have created) verified 16 July 2026 against aiappreciationday.org. Charity AI adoption, policy and governance figures are from the Charity Digital Skills Report 2026 (Zoe Amar Digital / Think Social Tech), published 13 July 2026 on 807 responses. This piece originally quoted the interim March 2026 findings (88% adoption, 60% policy); it was updated on 22 July 2026 to the final published figures. The context-over-model argument draws on Phil Hewlinson's work at Multiply Academy. Anthropic's plan, product and data-processing terms change frequently — confirm the current position at claude.com before relying on it. Review times and exception profiles reflect my own practitioner experience rather than benchmark research. Nothing here is legal, tax or data protection advice.