Burn Tokens, Not Headcount
What if your next hire is a context window?
Part 2 of The Self-Improving Firm, a nine-part series on what AI-native looks like for a UK accounting firm
Daniel Lawrence, CEO Bots For That
“Burn tokens, not headcount.”
this is a slogan. It comes from Silicon Valley. It belongs in the same family of provocations as “move fast and break things” and “ask for forgiveness, not permission.” Like all good slogans, it carries something true and something useful, and it also carries a lot of baggage that does not translate cleanly to a UK accounting firm.
The thing it carries that is true is this. For the first time in the history of the profession, the choice between adding capacity by hiring a person and adding capacity by spending on computers is a real, specific, comparable choice. It is something a finance director can put on a slide. It is something a managing partner can argue about in a quarterly meeting. It is no longer a metaphor.
That is what makes the slogan worth taking seriously, even with all the baggage. It forces a question the profession has not had to ask before, and that most firms are still not asking honestly.
Where the slogan comes from
In a software startup, headcount is the dominant cost. Engineers are expensive, slow to hire, slow to onboard, and difficult to scale. A founder who can solve a problem by spending five thousand pounds a month on API calls instead of hiring a sixty+ grand engineer has made an obvious decision. The slogan exists because that decision is now genuinely available to people building software, in a way that simply was not true three years ago.
The slogan does not translate cleanly into accounting for several reasons.
A first-year trainee in a UK mid-tier firm is not the equivalent of a software engineer. They are cheaper. They are part of a three-year training contract that is also a regulatory pipeline into qualified status. They generate billable revenue from very early in their first year. And they are doing work that is, for the firm’s traditional model, both economically essential and educationally essential. They are how the firm makes money on compliance and how the firm produces its future managers and partners.
You cannot simply, not hire them and rent GPUs instead. The training pipeline collapses. The qualification numbers fall. Five years later, there are no senior managers. Ten years later, there is no succession.
So the slogan does not work as a literal instruction. Anyone trying to sell it to you that way is either inexperienced or selling something.
What the slogan does work as, properly understood, is a forcing question. The question is not “should you replace your trainees with tokens”. The question is what proportion of your firm’s capacity should come from each, what each is genuinely good at, and whether the current ratio is the one you would choose if you were starting today.
What is actually comparable
The almost unavoidable fact, is that compute and headcount are now denominated in the same currency. They are both line items. They both have unit economics. They can both be increased or decreased month to month. That was not as easily true to say a few years ago. It is true now.
A trainee in a UK mid-tier firm costs the firm somewhere in the region of forty to fifty thousand pounds a year, all in, by the time you have factored in salary, qualification costs, study leave, supervision time, and the proportion of that year they spend learning rather than producing. A qualified accountant costs more. A manager costs significantly more. A partner is a different economic animal entirely.
A meaningful enterprise-grade AI deployment, capable of producing first-pass working papers, drafting routine correspondence, answering client queries, summarising meetings, and reasoning across the firm’s data, costs somewhere between a few thousand and a few tens of thousands of pounds a year per active user, depending on the tools, the data integration, and the volume of usage. The cost is falling roughly an order of magnitude every couple of years (and potentially getting faster).
These are not equivalent units. A trainee does things AI cannot do. AI does things a trainee cannot do. But they are now in the same conversation, on the same spreadsheet, in the same monthly P&L review. That is the new fact.
The honest question for a managing partner is this. If next year’s plan involves hiring four new trainees, two new qualifieds, and one new manager, what would the same money buy in compute? And what would the firm look like if the answer were three trainees, one qualified, one manager, and a serious AI deployment? Or two trainees and a larger one? The right answer for most firms is not zero of either column. But the current answer is, for most firms, the same answer as last year, with the AI line item added on top rather than substituted in.
That is the legion structure from Part 1, expressed as a budget.
What does not appear on the spreadsheet
A budget comparison only takes you so far. The harder things are the things that do not show up as line items.
The training question is the largest. It was flagged at the end of Part 1 as the hardest unresolved problem in the entire series, and it appears here because Part 2 is where it bites. If trainees spend less of their time on the basic preparation work that AI is now doing, what are they spending their time on? What are they learning by? How do they reach competence in three years? The ICAEW and ACCA syllabuses have not been redesigned around this question (to my knowledge). The training contract structure was built for a different production model. And the risk is firms that get this wrong could possibly produce a generation of qualified accountants who cannot do the work the qualification implies they can.
That is not a reason to keep doing things the old way. It is a reason to be deliberate. A firm that simply removes trainee preparation work and assumes the rest of the training pipeline will work itself out is making a serious mistake. A firm that thinks carefully about what its juniors should be learning by, designs a deliberate alternative pathway, and adjusts its mix accordingly is not.
The client trust question is the second. Most accounting clients, particularly in the SME segment, hire their accountant for the relationship as much as for the work. They want to know who is doing their accounts. They want to be able to ask their accountant a question and get a sensible answer. A firm that has aggressively substituted compute for people may produce identical output, on paper, but the client relationship looks different. That is not a reason to avoid the substitution. It is a reason to be honest about what is being kept human, and to make sure the human part is genuinely visible to clients.
The regulatory question is the third. The professional bodies are still working through what AI-assisted work means for review, for sign-off, and for professional liability. The PI insurers are doing the same. A firm that gets too far ahead of where the regulators are, without thinking carefully about audit trail and accountability, is taking on risk that does not appear on the compute-versus-headcount comparison.
None of these problems is a reason not to ask the question. They are reasons to ask it properly.
The honest comparison
Here is the partner-level conversation this series is asking firms to have.
Look at next year’s planned recruitment budget. Look at next year’s potential compute budget. Treat them as the same currency, because they now are. Look at the work the firm does and ask which of it should sit on each side of the line.
For most mid-tier UK firms, the answer is not zero recruitment. It is also not the same recruitment plan as last year with an AI tool added on top. It is somewhere in between, and the question is where. That is not a question with a single right answer. It is a question that depends on the firm’s mix of compliance and advisory work, on its client segments, on its succession plan, on its appetite for operational change, and on the partners’ honest view of what their firm is for.
The slogan, properly translated, is not “burn tokens, not headcount.” It is “be deliberate about what mix of tokens and headcount actually serves your firm and your clients, instead of inheriting last year’s mix and bolting AI on top.”
That is a less catchy line. It is also a more useful one.
A closing observation
The Roman legion did not solve the cavalry problem by spending more on horses. It solved it, eventually, by reorganising around a different idea of what an army was for.
A firm that approaches the AI question as a budget allocation exercise on top of the existing pyramid will get the modest productivity gains that come with that. A firm that uses the budget conversation as an excuse to ask what shape it actually wants to be in, and then makes the allocation accordingly, will end up somewhere different.
The next three parts of the series are about that shape. The sensor layer. The decision layer. The working papers factory. Each is a piece of what an AI-native firm actually looks like when you take the comparison seriously.
The question for now is whether the comparison is on your spreadsheet at all.
Daniel Lawrence is the CEO and co-founder of Bots For That and creator of the work automation operating system. He has spent more than a decade deploying enterprise automation and AI in regulated industries including accounting and professional services. The Self-Improving Firm is a nine-part series exploring what AI-native operations look like for mid-tier and large UK accounting firms.
Part 3, The Sensor Layer, asks what becomes possible when every client interaction, email, support query, filing deadline, and late payment is captured and made legible to a system that can reason across it.
