What if the partner’s job became more human, not less?
Part 8 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
Picture this: A partner in a mid-tier UK firm sits across from a client and tells them their business is not going to make it.
The client has known, deep down, in the way people know these things without admitting it to themselves out loud, for about six months. The deals not closing, costs higher than they should be, and the slow decline in cash flow. The client has kept paying themselves the same salary because they didn’t know how or when to stop.
The partner has known for about three months, from the numbers. They’ve waited for the right moment because they knew what the conversation would do to the client. The right moment comes in a monthly management accounts meeting. The partner walks the client through the numbers and then says the sentence out loud that they’ve both been avoiding.
The client cries. The partner watches and waits. Then, over the next hour, they work through what the client needs to do next. Not what the numbers say, that was clear enough. What the client, as a person, needs to do first, in what order, to protect their family, their staff, and what dignity they can keep out of what comes next.
This is what a partner does. It’s one of the most important things a firm does. And it’s not something an AI system is going to do at any stage.
The residual and the irreducible
There is a bad version of this scenario being played out in the media, which this series is now going to distance itself from.
And it goes something like this. AI will handle all the routine, mundane, repetitive stuff. Humans will focus on higher-value advisory. Partners will spend more time on relationships. Everything gets more meaningful.
In theory this is somewhat comforting, and it’s partly true, but it’s not sufficient. It is comforting because it lets partners and their firms believe that the AI conversation is happening to somebody else. It is partly true because relationship and advisory work is genuinely less automatable than compliance preparation. It is not sufficient because it treats the human role as whatever is left over after the automation, which frames human work as a residual category, rather than a category with its own logic.
The residual framing is dangerous for two reasons. First, it does not tell the firm what to invest in. If human work is defined as whatever AI cannot yet do, the definition changes every twelve months or less, and the firm has no basis for deciding which capabilities to build in its people. Second, it does not honestly answer the question of what the human is actually contributing. “The AI cannot do it” is not a value proposition. It is a placeholder.
The better framing, and the one this piece will explain, is that some human work in an accounting firm is not residual. It is irreducible. It is the thing the firm is actually for. Everything else, including the compliance production line, exists to support it. The AI conversation is not about identifying what humans can still do. It is about clarifying what humans are for, and reconfiguring the firm to do more of that and less of the rest.
Four kinds of irreducible work
There are at least four categories of work in a firm that will remain human by necessity rather than by convention, for the foreseeable future. It is worth being specific about each, because the specifics are what firms should be investing in.
The first is high-stakes conversation. The failing business conversation described at the start of this piece. The succession conversation with a founder who does not want to let go. The bereavement conversation when a client dies and the firm is talking to a spouse who has never seen the accounts. The redundancy conversation when a client is closing part of their operation and the accountant is the person helping them plan it. These conversations are not primarily about accounting information. They are about the accountant being a human being alongside another human being at a difficult moment. AI is very good at many things. Being a human being is not one of them.
The second is high-stakes judgement under ambiguity. The marginal disclosure call where the guidance is unclear and the auditor has to decide what to do. The tax position where the technical answer is defensible but the reputational answer is different. The valuation question where the number depends on assumptions that only the partner is qualified to make. The independence question where the firm has to decide whether it can take on a piece of work. These are judgements the firm’s professional bodies, insurers, and regulators require to sit with a named human being, and where the value of the judgement is inseparable from the accountability of the person making it. AI can inform these judgements. It cannot make them. Trying to make it make them is a category error that will produce firms that fail regulatory review and end up in front of tribunals.
The third is relationship trust. Not “relationship management,” which is a coordination category. The specific, personal, cumulative trust that a client places in the individual accountant they have worked with for years, that survives fee increases, difficult conversations, and the accountant’s inevitable mistakes, and that is the reason the client stays with the firm through changes of staff and platform. This trust is built through consistent presence, honest communication, and evidence that the accountant sees the client as a person rather than another set of numbers. And it is portable: when the accountant leaves the firm, the trust often goes with them. It is not reducible to any set of information the intelligence layer can hold. It is the accountant’s own asset.
The fourth is institutional memory that is not yet operational. Part 6 of this series argued that the closed-loop firm captures organisational memory in a form the intelligence layer can reason across. That is true for most operational memory. But there is a category of memory that remains stubbornly human: the sense of what the firm stands for, what it has refused to do in the past and why, what the founding partners meant when they set the culture, and what the current partners feel would betray something important. This is can’t be fully articulated in a decision layer, at least not yet. It is held by the people who have been in the firm long enough to have absorbed it. When they leave without transmitting it, the firm loses something it cannot easily recover.
Each of these four categories has one thing in common. The value of the human in the loop is not that they are performing a task the AI cannot yet perform. The value is that a human being is doing it, and the human quality of the doing is the thing that matters.
The critique from Part 7, addressed
Part 7 of this series acknowledged a serious critique of the AI-native model. Hierarchy performed one function the intelligence layer does not yet replace: it created natural aggregation points where judgement was compared, debated, and preserved across the firm. Remove the layers, and the aggregation goes with them. The firm moves faster and forgets faster.
Drowning in manual, repetitive work? Tell us the task and we’ll show you what to automate.
This critique is where the four categories above start to earn their keep. If the coordinating middle layer is gone, and the intelligence layer routes information without aggregating judgement, then the aggregation has to happen somewhere else. The Player-Coach role from Part 7 is one part of the answer: senior people still doing the work, still mentoring, still in the flow of engagements, still comparing notes as a matter of course. The DRI structure is another part: named individuals who are accountable for outcomes, whose judgement is legible and whose reasoning can be preserved as they exercise it.
But the deeper answer is that the aggregation of judgement is itself an irreducibly human function, and the firm needs to invest in the practices that produce it. Partner conversations that are actually about difficult calls, not about status updates. Case reviews where partners sit together and work through what a marginal decision should have looked like, not to punish but to align. Written reflections from senior people about what they have decided this year and why, captured in a form the firm can hold and future partners can read. Mentorship structures that are not just about developing juniors, but about junior partners spending real time with senior partners on the hard cases.
None of this is new. Good firms have always done some version of it. But most firms have done it informally, on the assumption that the pyramid did the aggregation automatically. In an AI-native firm, that assumption is no longer safe. Judgement aggregation has to become an explicit practice. This is the specific work that has to be added, not removed, when the coordination layer thins.
What this means for the partner role
The partner role, in this framing, is not “less operational, more strategic.” That is the residual framing, and it is not quite right. The partner role is where three of the four irreducible categories concentrate. High-stakes conversation. High-stakes judgement. Relationship trust. The fourth, institutional memory, is spread across the senior population but is often held most deeply by partners.
The AI-native firm, if it takes this seriously, is not a firm where partners spend more time on advisory. It is a firm where partners spend more of their time on the specific kinds of work that only a partner can do, and where the rest of the work, the routine correspondence, the file preparation, the coordination, the reporting, is either automated or absorbed by ICs, Player-Coaches, and DRIs.
This may have an odd implication. Partners in an AI-native firm could actually end up having fewer clients, not more. Their time on each client is likely to be deeper, not shallower. The economic model could actually shift from partners as leveraged supervisors of large teams to partners as high-density human interfaces to a small set of client relationships that matter. The compliance revenue is still there, but it is largely produced by the infrastructure. The partner revenue is produced by the partner being the partner.
For most mid-tier firms today, this is a considerable inversion of how partnership economics currently work. Partners are, on average, over-leveraged and under-focused. They have too many clients they cannot really pay attention to, and they spend too much of their time on work that could be done by systems or by other people. The AI-native firm is not a firm where partners work less. It is a firm where partners work on the right things.
The training implication, closed
Part 1 of this series flagged training as the hardest unresolved problem in the series. Part 5 gave a partial answer: juniors will learn by reviewing rather than preparing, by seeing patterns at scale, by developing judgement applied to output. That answer is incomplete on its own. This part completes it.
The other thing juniors will learn by, in an AI-native firm, is by being present at the high-stakes conversations. By sitting in the room, when they can, when a partner has a difficult conversation with a client. By observing the marginal judgement calls being made, and being invited into the reasoning behind them. By spending real time with the senior people whose institutional memory they will eventually inherit. This is not a new training pathway. It is an old one that most firms may have quietly stopped doing, because the pyramid was busy producing files and there was no time.
If the pyramid stops producing files, some of that time can be recovered. The Player-Coach role from Part 7 makes this explicit: senior people still in the flow of the work, mentoring by working alongside. The AI-native firm has, in principle, more capacity for genuine apprenticeship than the legion-model firm did, because the routine work that used to consume the mentoring hours is done by the infrastructure.
Whether that capacity gets used for apprenticeship, or whether it gets used for more billable partner hours, is a choice each firm will make. Firms that make the wrong choice may find, in about a decade, that they don’t have a bench of partners coming through. Firms that make the right choice will hopefully find they have produced the first generation of accountants trained natively in what the profession is actually for.
A closing observation
The Roman legion had a term for its most experienced soldiers, Evocati. Veterans who had served their full term and been retained, or recalled, because their judgement was worth more than their arms. They did not carry the same load as regular legionnaires. They advised. They handled the situations the centurions could not fully specify in advance. They were the legion’s memory of what worked and what did not, walking around in bodies.
Every legion had them. No legion could function properly without them. And they were not middle management, and they were not coordinators, and they were not, in any modern sense, managers. They were the human residue of a system that had learned that some things could not be codified into rulebooks, no matter how sophisticated the rulebook became.
The accounting firm has always had its own evocati. The senior partners who have seen enough to know how a difficult situation actually plays out. The client relationship holders who have been through a decade of a business together. The technical specialists who have accumulated the kind of judgement that is not in the manual. Every firm has them. Every firm depends on them more than it usually admits.
The AI-native firm is not a firm without evocati. It is a firm that has built the infrastructure to let the evocati do more of what only they can do, by taking the routine work off their plate. That is the point of this whole series. The infrastructure is not the goal. The goal is a firm where the humans are doing what only humans can do, at the intensity and quality that only humans can bring to it.
The question for now is what your firm’s evocati are actually spending their time on. If the answer is “the same things a competent senior manager could do”, that is the work of the next five years. Not to replace them. To free them.
Daniel Lawrence is the CEO and co-founder of bots for that and creator of the Automation Operating System (AOS). 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 9, Building the Firm in this New Shape, is the closer. It asks what a serious partner would build if they left a Top 50 firm tomorrow to start again, knowing what we now know.
