The Sensor Layer
What if your firm could see every client signal in real time?
Part 3 of The Self-Improving Firm — a nine-part series on what AI-native looks like for a UK accounting firm
A partner at a mid-tier UK firm knows things about their clients that nobody else in the firm knows.
They know the managing director is selling next year and has not told anyone. They know the FD is unhappy and is interviewing elsewhere. They know last year’s bank covenant breach was discussed in a corridor and never written down. They know which clients pay late on purpose and which pay late by accident. They know the new manager is going to struggle with the chairman of the audit committee because the chairman is a particular kind of difficult.
None of this is in a system. It is in the partner’s head. When the partner retires, most of it leaves.
This is the firm’s most important sensor. It is also the most fragile.
What firms actually have today
Most mid-tier firms believe they have client data. What they actually have is files.
There is a folder structure in SharePoint or a cloud equivalent. There are working papers from the last three audits. There are last year’s accounts as a PDF. There is an engagement letter, somewhere. There is the practice management system, which knows about jobs and deadlines but very little else. There is Xero or QuickBooks or Sage on the bookkeeping side, which knows about transactions but not about conversations.
There are emails. There are a lot of emails. They are scattered across the inboxes of every person in the firm who has ever spoken to the client, with no shared visibility and no structure.
There is a WhatsApp group with the client, probably, because that is how mid-market clients increasingly communicate with their accountant. The partner is in it. Sometimes the manager. The senior is not. The firm brain does not exist in this thread at all.
There is a Teams chat about the client between the engagement team. There are notes from last week’s meeting in a notebook on someone’s desk. There are voice notes a partner made walking back from a client visit, on their phone, that nobody else has ever heard.
This is not a data infrastructure. It is a collection of artefacts produced by a firm whose information-moving function still runs through people. Each piece of it makes sense in isolation. None of it can be reasoned across by anything other than a human being who happens to be familiar with all of it, which means in practice nobody can reason across it. Not even the partner. Even the partner has forgotten half of it.
This is what the source material means by “if it is not recorded, it does not exist to the company brain”. For most firms, the answer is that most of it does not exist. The firm runs on it anyway, because human beings are remarkable at extracting signal from incomplete information. But the firm is operating well below the capability it could be operating at, because the system that would otherwise be reasoning across everything has been given almost nothing to reason with.
What the sensor layer actually means
The technical concept is borrowed from robotics and from agentic AI. A system that interacts with the world needs to take in information about the world before it can do anything useful with it. The sensor layer is the architecture of that taking-in: what signals are captured, how they are structured, where they are stored, and what the rest of the system can do with them.
For an accounting firm, the sensor layer is the answer to a specific question. What does the firm know about each client, in a form that a system can read?
The honest answer for most firms today is “very little”. The optimistic answer for most firms in five years is “a great deal more, if they have been deliberate about building it”.
A serious sensor layer for a UK accounting firm pulls in signals from at least the following sources. Bank feed activity from the client’s bookkeeping platform, with transaction patterns analysed rather than just stored. Filing deadlines from HMRC and Companies House, with the firm’s response status against each. Practice management system events. Email correspondence with the client, with permission, parsed for sentiment, urgency, and topic. Meeting notes and call transcripts, generated automatically. WhatsApp and Teams messages, if those are the channels in use, with appropriate consent. Document exchange events. Late payment patterns, both to the firm and to the client’s suppliers where visible. Industry-specific signals if relevant: VAT registration threshold proximity, payroll growth rates, sector benchmarks.
Each of these is a sensor. Each produces a stream of structured signals that, in isolation, says something small. Together, they say something the firm could not previously see.
What you can do with it that you cannot do today
The point of building this is not the sensor layer itself. The point is what becomes possible once it exists.
You can spot the client who is heading into trouble before the trouble lands on your desk as an emergency. Late payments accelerating, supplier complaints in email, a quiet pattern of director loan movements, a VAT return that suggests the business is hitting a wall. Any one of these is missable. Together, they are a pattern. A firm that can see the pattern can call the client three months earlier, with something useful to say.
You can spot the cross-sell opportunity that nobody has noticed. A client whose business is growing in a particular way is heading toward a regulatory threshold, or a financing decision, or an acquisition the firm could advise on. The signals exist. They are not currently visible to anyone except, occasionally, the partner who happens to have been paying attention.
You can spot the client who is unhappy before they leave. Response times slowing. Tone in emails shifting. A WhatsApp group going quiet. Meetings rescheduled twice. None of this is hidden information. It is just not aggregated anywhere a system can read it.
You can run the firm as a firm. A managing partner who can ask the firm brain a question and get an answer drawn from across every client engagement is operating with a different level of awareness from a managing partner who has to call a partner meeting to find out what is going on. This is the recursive loop concept in action. The firm becomes legible to itself.
This is also where key AI automations also do real work. They are sensor layer implementations for specific slices of the firm’s data. The principle, though, is bigger than any product. Any firm that takes this seriously, with whatever tools, will end up building something like a sensor layer. Any firm that does not will be running blind to key signals.
The hard part
The technical work is real but mostly solvable. The hard part is operational discipline.
The sensor layer only works if the firm actually records things. That means partners writing meeting notes, in the system, after every client interaction. It means voice memos getting transcribed and stored, not left on phones. It means WhatsApp conversations being mirrored into a shared workspace where the firm brain can see them. It means the practice management system being used as a system of record, not a job-tracking afterthought.
Most firms will recognise this is a long way from how they currently operate.
There are also serious questions to answer that the technology does not solve for you. Client consent for AI-assisted processing of their data is not optional. GDPR considerations are not optional. Decisions about which calls are recorded, which messages are mirrored, which transcripts are retained and for how long, are not optional. The firm needs a position on each of these, agreed at partner level, written down, and reflected in client engagement letters.
These are not reasons not to build the sensor layer. They are reasons to build it deliberately. A firm that turns on transcription across every meeting without thinking about consent and retention is making a serious mistake. A firm that thinks carefully, builds the consent architecture properly, and is honest with clients about what is being captured and why, is not.
The third hard problem is that the sensor layer changes what people are accountable for. If the system can see that a manager has not responded to a client email in nine days, the manager can no longer hide that fact in the volume of their inbox. If the system can see that a partner’s clients have lower satisfaction signals than another partner’s, that fact exists, in the firm, in a way it did not previously. Some firms will find this clarifying. Others will find it threatening. The firms that succeed with the sensor layer are the ones that lean into the clarity rather than away from it.
A closing observation
The hierarchy described at the start of this series existed because information had to be physically carried by humans. Every layer of the pyramid was, among other things, a packet of memory. A junior remembered the detail of the file. A senior remembered the pattern across the file. A manager remembered the engagement. A partner remembered the client. Each layer compressed and held what the layer below could not.
The sensor layer is what replaces that as the firm’s memory. Not the people. Not the files. The continuous, structured, queryable record of what the firm knows about its clients, accessible to humans and to machines, surviving the departure of any individual, and growing more useful the more it accumulates.
A firm that builds this will, over time, know more about its clients than any individual partner ever could. That is a strange thing to say about a profession that has run on partner memory for centuries. It is also true.
The question for now is what your firm actually records, and what it lets fall away. Most firms have not asked this question deliberately. It is overdue.
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 4, The Decision Layer, asks what becomes possible once the firm’s policies are explicit enough that an AI system, and the people working alongside it, know what to do without asking.
