The Self-improving Firm
A nine-part thought leadership series for mid-tier and large UK accounting firms Companion arc to Making Accounting AI Daniel Lawrence, Bots For That Ltd
Strategic Positioning
Making Accounting AI established the case for the profession. The Self-Improving Firm takes the next step. It stops talking about AI as a productivity tool inside the existing firm and starts asking what the firm itself should look like if you built it around AI from the floor up.
The framing borrows from two source pieces. The first argues that most companies today are organised like Roman legions, with humans acting as conduits for information moving up and down a hierarchy, and that bolting AI onto that structure delivers a 20% productivity bump on a fundamentally outdated operating model. The second goes further: AI should not be a tool inside the company, it should be the operating system the company runs on, with closed-loop systems replacing open-loop decision-making and a small number of clearly defined human roles replacing the management pyramid.
This is deliberately uncomfortable territory for a mid-tier firm. Partner structures, manager-senior-junior pyramids, billable hours, training pathways, succession planning, every load-bearing element of the modern firm rests on the legion model. That is what makes the series worth doing. It is also why the tone matters. This is not a demand that firms blow themselves up. It is a serious question, asked with the profession, not at it, about which parts of the legion are load-bearing and which parts are habit.
The series also lands cleanly against BFT’s standing positioning. The AI-first, brought into accounting line is the entire argument for why Daniel Lawrence is the one asking these questions. Everyone else in the market is asking how to put AI into the firm. The series asks how to put the firm into AI.
Core Argument
A mid-tier UK accounting firm in 2026 has three structural choices. It can run AI as a productivity tool inside the legion, a 20% improvement on the existing model. It can pretend the question does not apply to it. Or it can become an AI-native firm: closed-loop, queryable, organised around DRIs and ICs rather than managers, with partners deployed where human judgement is genuinely irreducible. This series is about what the third option actually looks like, honestly, for a real UK firm with real clients, real regulators, and real partners who did not sign up for a startup.
The Nine-part Arc
Each part poses a specific what-if question rooted in the source ideas, then explores what an honest answer would look like for a UK firm of 50 to 500 staff. The arc moves from provocation through dismantling to reconstruction.
Part 1 - The Roman Legion Problem
What if the org chart is the bottleneck?
Opens the series with the central provocation. The modern accounting firm is structured to move information up and decisions down through layers of humans. That structure made sense when human judgement was the only mechanism for processing complex financial information at scale. It made less sense when spreadsheets arrived. It makes less sense again now.
The question is not whether the partner-manager-senior-junior pyramid still works. It is whether it would be the structure anyone designed if they were starting today.
Part 2 — Burn Tokens, Not Headcount
What if your next hire is a context window?
The most quotable line from the source material, applied to the most labour-intensive profession in the British economy, and reframed as a P&L conversation. UK firms have spent two decades arguing about offshoring, near-shoring, and the talent pipeline. This part asks a harder question: what work in a mid-tier firm is currently done by people because it has to be, and what work is done by people because that is how it has always been done?
The framing is deliberately concrete. Recruitment cost versus compute cost is a real, quantifiable comparison. Trainee intake, qualified accountant salaries, manager bonuses, these are the numbers a firm runs on. The part asks partners to honestly compare next year’s recruitment budget against next year’s potential compute budget, and to be clear about which mix actually serves the firm and its clients.
Part 3 — The Sensor Layer
What if your firm could see every client signal in real time?
The first of three parts on the recursive loop. A typical mid-tier firm has client data scattered across Xero, IRIS, CCH, Outlook, a practice management system, WhatsApp groups, and a partner’s memory. The sensor layer concept asks what becomes possible when every client interaction, every email, every support query, every filing deadline, every late payment, is captured, structured, and made legible to a system that can reason across it.
This is where MTD Command Centre and Client Insight do real work, but the part is not about products. It is about the principle: if it is not recorded, it does not exist to the firm brain.
Part 4 — The Decision Layer
What if every policy in your firm were written down?
The hardest part of the series for most firms. Mid-tier practices run on partner judgement, manager experience, and a thick layer of unwritten convention. The decision/policy layer concept demands that the rules be made explicit: what can the AI do without asking, what must it escalate, what is the partner’s specific authority, what is the manager’s.
Most firms cannot answer these questions today. That is not a failure of AI readiness. It is a failure of operational clarity that AI happens to expose.
Part 5 — The Working Papers Factory
What if seniors stopped preparing files and started specifying them?
The most original part of the series. Software teams have started running specs-and-tests factories, where humans write the specification and the success-defining tests and AI generates the code until it passes. Some teams now maintain repositories with zero handwritten code, only specs and tests.
There is a direct accounting analogue worth taking seriously. The working papers model, with seniors preparing, managers reviewing, partners signing off, is a human implementation of the same idea. The question worth posing is what happens when spec-writing and review remain human but file preparation becomes generative.
This is genuinely provocative for a mid-tier firm because it inverts the current career pyramid. Seniors do not graduate into managers by doing more files faster. They graduate by getting better at writing specifications and tests. The training implications are significant. The economic implications are larger.
Part 6 — Closed-Loop, Not Open-Loop
What if your firm got smarter every night?
Reframes the learning mechanism in language accountants instinctively understand. The distinction between open-loop and closed-loop operations is exactly how internal controls work. An accounting firm already lives in a partially closed-loop world for audit and assurance.
The question, then, is which of the firm’s own operational decisions are currently open-loop. Which client onboarding judgements, which fee proposals, which staff allocations, which review notes go out the door and never come back to inform the next round? A learning firm analyses its own work, every query, every objection, every misfiled return, every difficult conversation, and updates its rules and tools accordingly.
Annual partner reviews and post-mortem files are the legion-model version of this. The recursive version is continuous. For an accounting firm, this is the difference between a practice that compounds knowledge and one that simply rotates staff through the same mistakes.
Part 7 — The End of Middle Management
What if coordination weren’t a full-time job?
The sharpest provocation in the series, and the one that needs handling with the most care. A typical mid-tier firm has a substantial layer of senior managers and directors whose primary function is coordination, moving information between juniors and partners, between departments, between the firm and the client.
Rather than blunt elimination, the part offers a more honest framing borrowed from Jack Dorsey’s work at Block: every AI-native company runs on three archetypes. Individual Contributors, the builder-operators across all functions who bring working prototypes rather than pitch decks. Directly Responsible Individuals, strategic coordinators focused strictly on customer outcomes with clear, undivided responsibility. And the AI Founder Type, leaders who lead by example and refuse to delegate their AI strategy.
The question for the firm stops being are your managers redundant and becomes which of your senior people are coordinators, and which are DRIs. That is a question a managing partner can engage with seriously. The push-back is part of the conversation, not a reason to soften it.
Part 8 — Where Humans Still Belong
What if the partner’s job became more human, not less?
The counterweight to Part 7, and the part that makes the rest of the series defensible. There is work in a firm that AI cannot, should not, and probably will not touch for years. The difficult conversation about a failing business. The ethical judgement on a marginal disclosure. The negotiation with HMRC. The sensitive succession discussion. The new-client pitch that turns on trust.
The series argues these are not residual tasks left over after automation. They are the irreducible core of the profession, and they are also the highest-value work the firm does. Done well, this part repositions the partner not as a coordinator of work but as the firm’s DRI for its most important relationships, the firm’s interface with reality.
Part 9 — Building the Firm in this New Shape
What if you were starting a firm today?
The closer. The series ends not with a transformation roadmap, those are the cheap, easy, vendor-led version of this conversation, but with a thought experiment.
If a serious, experienced partner left a Top 50 firm tomorrow to start their own practice, knowing what we now know about AI, what would they build? What would they not build? What would they record, what would they automate, what would they refuse to automate, and how many people would they hire in year one?
The answer is the implicit benchmark every existing firm should be measuring itself against. The most interesting firms over the next decade in UK accounting are probably not the Big Four, and probably not the practices still doing manual bookkeeping. They are the mid-tier firms that already have scale, client trust, and capital, and the willingness to rebuild operationally.
