The Decision Layer
What if every policy in your firm were written down?
Part 4 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
Ask a managing partner at a mid-tier UK firm to write down the rules by which their firm operates, and you may get one of two answers.
The honest answer is “I am not sure I could”. The other answer is the firm’s manual, which is a real document, but which everyone in the firm knows is not actually how the firm operates day to day. There is the manual, and then there is what people actually do, and the two are related but not identical. The space between them is where the firm really runs.
This is not a criticism. It is how professional services firms have worked for as long as there have been professional services firms. Partners exercise judgement. Senior managers exercise lower-grade judgement they have absorbed from partners. Juniors do what the senior in front of them tells them to do. Conventions accumulate. Different partners do things differently and the firm tolerates the variation because the partners are senior and have earned the right to do things their way.
This is also the thing AI cannot work with.
What the decision layer means
In the source framework, the decision layer, sometimes called the policy layer, is the set of rules a system uses to decide what to do, what to escalate to a human, and what to refuse. It is the difference between an autonomous system that runs your firm into a regulatory wall and one that knows when to stop and ask.
In a technical setting, this looks like a configuration file. In an accounting firm, it looks like answers to questions the firm probably cannot answer today.
What can a manager sign off on without partner involvement? Specifically, by client size, by engagement type, by risk profile? Where is the threshold? Is the threshold the same across all partners or does each partner have their own version?
What can a senior do without manager review? What can a junior do without senior review? Are there situations where a junior should escalate directly to a partner, bypassing the chain? When?
When can the firm extend credit to a client? Under what circumstances does an unpaid invoice become a stop-work decision, and who makes that call?
What is the firm’s position on engaging with crypto clients? Cannabis clients? Adult industry clients? Politically exposed persons? Are these positions written down, or are they “we know it when we see it”?
When a junior spots something that looks like a potential money laundering signal, what do they do? What does the senior do? At what point does the firm have to file? Who decides?
When a client gets aggressive with a member of staff, what is the firm’s response? What does the partner have to do? At what point does the firm fire the client?
These are not exotic questions. They are routine questions a mid-tier UK firm faces multiple times a year. They are also questions most firms cannot answer in writing, in a way that two different partners would agree with, off the top of their heads.
That is the decision layer problem.
Why this is the hardest part of the series
Most firms can be talked into the sensor layer. It is straightforward and it is mostly an infrastructure problem. The case for it is intuitive: more data, better signals, fewer surprises. Partners may grumble about meeting notes but they will, eventually, come around.
The decision layer is different. It demands something firms have spent decades avoiding. It demands that the rules be made explicit. And the moment you start writing the rules down, several things happen at once, none of which are comfortable.
The first thing that happens is that the partners discover they disagree with each other. Partner A signs off on advisory engagements at a fee threshold that Partner B would never agree to. Partner C accepts client work from sectors Partner D has refused. These disagreements have always existed. They have been managed through partner autonomy and a polite refusal to compare notes. The decision layer exercise drags the disagreement into the open, and the firm has to decide whether to align or to formally codify the variation.
The second thing that happens is that managers and seniors realise the firm has been less coherent than it presented itself as being. The “firm’s position” on a particular issue turns out to be five partners’ personal positions, none of which were written down, all of which were enforced inconsistently. This is uncomfortable for people who have spent years acting as if they understood the firm’s policy. It is doubly uncomfortable for the partners, because the inconsistency was visible all along to the people who reported to them, and the polite fiction that it was not visible is now broken.
The third thing that happens is that some of the firm’s flexibility goes away. Partners are used to being able to make a discretionary call on a Tuesday afternoon, and the freedom to do that is part of how they have always defined seniority. A written decision layer constrains that. Not entirely. There is always a category of decision marked “escalate to partner judgement”. But it constrains it more than current practice does. Some partners will resist this. The resistance is honest. Discretion is real and it is sometimes valuable. The question is whether the firm has more of it than it actually needs.
The fourth thing that happens is the unwelcome discovery that the firm has been carrying a lot of risk it had not properly looked at. When you write down the rules for engaging with potentially high-risk clients, you also discover the clients on the books today who do not fit the rules. When you write down the threshold for filing a SAR, you discover the historical cases that probably should have been filed. The exercise produces a small mountain of regulatory housekeeping. This is a feature, not a bug, but it does not feel like one in the moment.
This is not really an AI problem
The point worth holding on to is that none of this is about AI.
AI is the thing forcing the conversation, because you cannot deploy an autonomous system that touches client work without telling it what it can and cannot do. But the underlying problem, the gap between the manual and what the firm actually does, predates AI by a century. It is the operational equivalent of running a business without management accounts. Most mid-tier firms have done it for decades and got away with it, because the firm’s partners knew what was happening and could fill in the gaps with their own judgement.
This is not sustainable for two reasons that have nothing to do with AI. The first is succession. Every firm losing a senior partner is losing a decision layer that nobody else can fully reproduce, and the loss is unmanaged because the layer was never explicit. The second is scale. Firms growing from twenty to fifty staff, or from fifty to two hundred, hit a wall that the partner-knowledge model cannot cross. The firms that scale through it are the ones that, often without naming it as such, build a decision layer along the way.
AI just brings the deadline forward. A firm that wants to deploy AI on client work has to answer the decision layer questions explicitly, because the AI cannot read the partner’s mind. A firm that does not deploy AI can put off answering the questions for another five years. In neither case do the questions go away.
What becomes possible
A firm with an explicit decision layer is a different firm.
Onboarding new staff is faster, because the rules are written down rather than absorbed by osmosis. A new senior can be productive in months rather than years, because the senior’s job is now bounded by something they can read, not a culture they have to infer.
Partner meetings become more useful, because disagreement is about substance rather than precedent. When two partners disagree on a marginal call, they are disagreeing about the underlying judgement, not about which partner gets to set the convention.
Risk is more honestly distributed. The firm can see, in writing, what it has decided is acceptable and what it has not. When something goes wrong, the conversation is about whether the rule was right, or whether the rule was followed, rather than about whose unwritten norm was supposed to apply.
AI becomes possible to deploy on real work, because the system has something to follow. Not “do whatever a partner would do”, which is incoherent. But “do this when the engagement matches these criteria, escalate when it matches these criteria, refuse when it matches these criteria”. That is something a system can act on, and something a partner can sign off on the system acting on.
What should stay judgement
The decision layer concept is not “codify everything”. A firm that codifies everything produces a firm that cannot respond to the unusual situation, which is often the situation that matters most. Some decisions genuinely depend on context, relationship, and intuition that cannot be reduced to a written rule, and a serious decision layer is honest about which decisions those are.
The trick is knowing what to codify and what to leave to judgement. The honest version of the answer, for most firms, is that there is far more that could be codified than is currently codified, and a smaller residue of genuinely contextual judgement than partners typically claim. The series will revisit this in Part 8, when it asks where humans still belong. For now, the question is simpler. What does the firm currently leave to “partner discretion” that is really just convention, and what is genuinely a partner-level call?
Most firms have not separated the two. The decision layer exercise is the separation.
A closing observation
The Roman legion ran on a written rulebook. The Manual of Military Discipline. Centurions did not improvise. Junior officers did not exercise vague judgement. Every decision had a rule, and the rule was knowable to anyone who had been trained in it. The legion’s effectiveness depended on that legibility. Two centurions in two different legions made the same call in the same situation, because the rule was the rule.
The modern accounting firm is the opposite. The rules vary by partner. The conventions vary by office. The “firm policy” is, often, an aggregation of partner habits with a manual stapled to the front.
This worked when the firm was small enough that partners could see each other’s work. It works less well at fifty staff. It does not work at two hundred. And it does not work at all when the firm wants to deploy systems that need to know what they are allowed to do.
The decision layer is not a bureaucratic exercise. It is the firm finding out what it actually believes. Most firms have not had to do that, in writing, in their entire history. The ones that do it well will have an advantage the others will struggle to close.
The question for now is simpler than the answer. If a new partner joined your firm tomorrow, could they read a single document and know how the firm makes decisions?
If the answer is no, the decision layer is the work.
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 5, The Working Papers Factory, asks what happens when seniors stop preparing files and start specifying and reviewing them, and what that means for the career pyramid that has organised the profession for a century.
