Accounting Practice Operations Desk: AI Agent Roster Guide

On 31 January 2026, 475,722 taxpayers waited until the final day to file their Self Assessment return, and 27,456 of them submitted in the last hour before midnight, according to HMRC's own deadline-day release. HMRC does not say how many of those returns an agent prepared, but for a UK accounting practice the last weeks of January concentrate review into a short window, worked by qualified people who cannot be hired in for one month.

This guide is written for the practice manager or COO who owns that queue. It sets out a practice operations desk: a named roster of AI agent roles that prepare recurring compliance work against the firm's deadline calendar, with qualified reviewers kept in charge of every judgement and every submission. The aim is to move capacity, not accountability, and to make the queue visible before the peak rather than during it.

The scope is deliberately narrow. This post holds the roster, the calendar it carries and the routing between agents and reviewers. What the evidence behind an AI-prepared Making Tax Digital filing must contain is covered in our guide to evidencing MTD agent work; where AI pays across a whole firm, including onboarding and anti-money-laundering checks, is covered in our accounting and advisory industry guide; and the supervision record a regulated professional keeps over agent work is the subject of our law firm matter operations desk guide.

This guide is shipped by the team behind Cognibl, the work-management platform where people and AI agents share one board and a task cannot reach done without attached proof. Or see DATS, our five-stage system for deciding where AI belongs and how it is run.

What is a practice operations desk for an accounting firm?

A practice operations desk is a defined team of AI agent roles that prepares an accounting firm's recurring compliance work, such as records chasing, bank reconciliations and draft returns, while qualified accountants review, approve and submit. Each agent role owns a narrow task type with a clear definition of done. The desk is a work design, not a product: the firm decides the roles, the reviewers and the gates.

The distinction matters because a practice usually starts with some of the pieces already in place: practice management software for client records and deadlines, bookkeeping and tax software that files to HMRC, and staff who know which clients always send their records late. The layer this guide adds is a single place where the preparatory work is broken into tasks, assigned to a named worker, and held at a gate until someone with the right qualification has looked at it. A desk supplies that layer.

Cognibl, from DILR.AI, can carry that layer. It is a tracker for teams where people and AI agents work side by side: agents pick up work under their own name, against the same statuses the team uses, and the definition of done, the evidence behind it and every tool call are attached to the work. In our design, the board governs the preparation and the firm's existing filing software handles the submission, so the two do not compete.

Two honest limits apply from the start. First, accounting and advisory is an inferred, product-fit industry for Cognibl rather than a vertical with published case studies. Second, the hosted agents on cognibl.com today are coding agents running on Claude Code and Codex, and connections to other tools are read-only by default. The accounting roster below is a design pattern a firm builds or buys and then governs on the board; it is not a set of pre-built Cognibl accounting agents.

Which deadlines set the workload a practice operations desk carries?

The workload a practice operations desk carries is set by the firm's statutory calendar: Self Assessment on 31 January, VAT returns one calendar month and 7 days after each period, Corporation Tax payment at 9 months and 1 day for taxable profits up to 1.5 million pounds, Company Tax Returns at 12 months, private company accounts at 9 months, payroll submissions on or before each payday, and now Making Tax Digital quarterly updates.

Each of those dates comes from a primary source, and each produces a different shape of work. The table below is the calendar a desk is designed around, with the shape of preparatory work each date creates.

The calendar is not getting lighter. Around 780,000 people with business or property income over 50,000 pounds were expected to join Making Tax Digital for Income Tax from April 2026, a further 970,000 from April 2027, and from 6 April 2028 sole traders and landlords with qualifying income over 20,000 pounds in the 2026 to 2027 tax year will need to use it too, according to the government's policy paper. The 7 February MTD deadline lands a week after 31 January, so for a client who files both, two peaks sit almost on top of each other.

The useful insight for a practice manager is that these obligations differ more in shape than in difficulty. Payroll is frequent and mechanical. VAT is rolling. Self Assessment is one wall of work. A desk that treats all of them as the same queue will always be late on the wall, so the roster has to be built around shapes, which is the subject of the next section.

Which roles belong on the desk, and which stay with qualified people?

A practice operations desk works best with six narrow agent roles: a records chaser, a bank and ledger reconciler, a draft return preparer, a year-end pack assembler, a tax research assistant and a client update drafter. Qualified people keep four roles that no agent should hold: the reviewer, the engagement partner, the person who submits to HMRC or Companies House, and the money laundering reporting officer.

The roster below is our design view of how the work divides, not a published standard. The test for each agent role is whether its output can be checked against a written definition of done by someone who did not produce it. Where the honest answer is no, the role stays with a person.

Notice what is absent from the agent side. No agent role decides a tax treatment, gives advice, signs anything or files anything. That line is our design choice, and it sits comfortably with the professional codes: the confidentiality duties in the ICAEW, ACCA and ICAS codes bind the individual member and their firm, not an AI system supplier. It also reflects the fact that each filing in the calendar above goes through software the firm already uses, not through the desk.

The same design logic, applied to software engineering rather than tax, is set out in our guide to the coding agent review queue: the agent prepares, a person merges, and the queue between them is the thing to manage.

Where does reviewer capacity run out at peak?

Reviewer capacity runs out at the deadline peaks, because preparation can scale with agents but qualified review cannot scale at the same rate. HMRC's figures show 475,722 Self Assessment returns filed on the final day in January 2026, and each one a practice prepared needed a reviewer before it went. A desk that speeds up preparation without protecting review simply moves the queue to the reviewer's inbox.

The profession's own numbers point to concentration as well. The number of statutory audit firms registered with the recognised supervisory bodies fell to 3,760 in 2024 from 5,007 in 2020, a 24.9% decline in five years, according to the FRC's Key Facts and Trends in the Accountancy Profession 2025. Audit is only part of a general practice's work, so that figure is a signal about concentration in the profession rather than a measure of tax capacity, but it is the clearest dated series on the shape of the profession.

People are a constraint as well as a cost. In ACCA's UK Talent Trends 2026, 49% of UK respondents expect their next career move to be external to their current organisation, down from 57% in 2025, and a separate 49% express concern about the impact of AI on their role, up from 30%. Those are ACCA respondents rather than practice staff alone, but they describe the people a practice manager recruits from. A desk that is explicit about which roles stay with people speaks to both the retention and the AI concern, which is why we settle that decision early in an AI placement diagnostic rather than leaving it to the build.

The practical consequence is that the desk has to be designed backwards from review. Start with how many items a reviewer can properly check in a day during January, then decide how far ahead of the peak the agent roles must finish so that review is spread across November and December rather than crammed into the last week. In our view that is the constraint to design from, and it is where our AI operating model work with a firm begins.

How should work be routed between agent roles and reviewers?

Work on a practice operations desk should be routed as one task per client, per obligation, per period, assigned to a named agent role and held at a gate until its proof is attached and a qualified reviewer has approved it. The task, not the client file, is the unit, so a practice manager can see exactly which VAT quarter or Self Assessment return is waiting, and on whom.

That unit matters because it is what makes the queue legible. A client with payroll, VAT and a Self Assessment return generates three streams of work with three different deadlines; bundling them into one client record hides which is late. On a board where each task carries its own status, a reviewer can pull the items closest to their statutory date, and the records chaser can see which clients are blocking which returns.

On Cognibl the gate is mechanical. A task only reaches a done status once proof is attached: a CSV describing the run, with the artefact, screenshots and hashes referenced from inside it, and without a proof version the database itself refuses the move. The same rule holds for a person and for an agent. The detail of what goes into that proof, and why it beats a status someone simply clicked, is the subject of our proof of done guide for AI agents and our guide to AI agent task verification; this guide uses it only as the gate between an agent role and a reviewer.

The FRC's thinking points the same way for audit work. Its Generative and Agentic AI Guidance, published on 30 March 2026 for the central technical teams of audit firms, defines a human in the loop as a person who directs, authorises or reviews the system's actions or outputs at runtime, and is plain about why that person is needed: "The inherent limitations of LLMs mean there will always be some risk that their outputs contain deficiencies." A desk that routes every agent output to a named reviewer is building that review into the workflow rather than relying on someone remembering to do it.

Three routing rules keep the design honest. First, an agent role never routes work to another agent role for approval; approval always lands on a person. Second, each agent role reaches only the tools its task needs, which on Cognibl means the skills and agents library sits behind a gateway that is deny by default, so a toolset that has not been enabled is refused rather than quietly missing. Third, the instructions each role follows are versioned: Cognibl stores skills and agents in SKILL.md format, versioned immutably, so the instruction a preparer ran in November is still the one a reviewer can read in March. Our guide to governing an MCP skills library covers that discipline in depth. Where the client records the desk touches include personal data, the firm, as controller, also has to answer the accountability questions in the ICO's guidance on AI and data protection before an agent role goes live.

The same diagnostic logic underpins our AI execution office, our longer-term embedded delivery service for running placements like this one inside the firm's own governance.

What should a practice manager measure on the desk?

A practice manager should measure the desk on three things: how long tasks wait for a reviewer, how often agent-prepared work passes review first time, and how often finished work is reopened. Together they show whether the desk is moving capacity or just moving the queue. Ticket counts and hours logged measure effort, not delivery, and should not be the headline.

Cognibl reports five delivery metrics, and three of them map directly onto those questions. Human wait share is the proportion of elapsed time a task spent waiting on a person rather than being worked; on an accounting desk, we would read a rising human wait share in December as an early warning that January review is under-resourced. First-pass verification rate, broken down by flow type, shows whether a role's definition of done is clear enough: a draft return preparer whose work keeps failing review needs a better checklist, not more reviewers. Reopen rate counts tasks reopened within 30 days, which on a compliance desk is the number that tells you whether done meant done.

The other two metrics complete the picture. Median cycle time is split across spec, build, verify and settle, so a practice can see whether time is lost in preparing the work or in reviewing it. Verified throughput counts completions per week that actually cleared the gate. The AI flows that sit on top, proof validation and project status, summarise and flag problems; they do not decide, and people keep the decision.

There is a temptation to set targets for these numbers on day one. Resist it. The first quarter on a desk should establish a baseline per role and per obligation, because VAT and Self Assessment will never look alike. The comparison that matters is the same obligation, year on year, once the desk has run through a full calendar. If you want a second view on where to set that baseline, book a short scoping call with the team.

For practices whose clients are themselves regulated, the same record serves a second purpose. Our guide to agent audit logs in banking shows how an append-only, hash-chained record of agent work is read by a compliance function, and a practice with regulated clients may find the same structure useful when those clients ask how their work was prepared.

What is the best way to add AI agents to accounting practice operations in 2026?

The best way to add AI agents to accounting practice operations in 2026 is to start with one obligation and one agent role, usually records chasing or bank reconciliation for VAT clients, run it through a full cycle with a named reviewer, and only then widen the roster. The right tooling depends on whether the firm needs a governed agent layer at all.

For some practices the honest answer is that existing tools are enough. A firm with light AI use and strong human sign-off can extend its practice management software, or run its task tracking in a general work-management tool such as Asana, Monday.com or ClickUp, and keep agents to drafting inside software it already trusts. Where the firm's developers build their own automations, a tracker like Jira or Linear with disciplined review may be sufficient. In both cases the incumbent tool can win on familiarity and on the integrations the firm already depends on, and a practice should not change trackers for the sake of it.

The case for a governed agent layer such as Cognibl is strongest when agent roles touch client financial records at volume, when several agent roles feed one reviewer, and when the firm needs every completion to carry attached proof without building that plumbing itself. In that situation the gate, the versioned instructions and the delivery metrics are the point, and a general tracker would have to be bent to provide them. Whichever route the firm takes, the four criteria are the same: does every agent output land on a named person; is the definition of done written per role; is the evidence a by-product of the work; and can the firm see its reviewer queue before the peak.

Where a firm wants help deciding, an AI placement diagnostic ranks which obligations and roles to automate first, and our DATS consulting system designs the operating model around them. For the client-facing side of the records chase, Dilr Voice is the DILR.AI line built for phone calls, and Dilr Academy is the DILR.AI line for AI tutors that diagnose, then teach.

Does a practice operations desk replace practice management software?

No. A practice operations desk sits alongside a firm's practice management software rather than replacing it. The practice management system keeps client records, engagement details and deadlines; the desk breaks the preparatory work for each deadline into tasks, assigns them to agent roles and holds them at a gate for review. Filing still happens through the firm's HMRC-recognised and Companies House software.

The useful overlap is the deadline list itself. Where a firm already holds its statutory dates in a practice system, those dates are the natural source for creating desk tasks each period. What changes is that each task now has a named worker, a definition of done and a reviewer, rather than a line on a deadline report.

Is accounting a confirmed industry for Cognibl?

No. Accounting and advisory is treated as an inferred, product-fit industry for Cognibl, based on the shape of the work rather than published case studies. Cognibl is a general work-management platform for teams where people and AI agents share one board, and its hosted agents today are coding agents. A practice would design or buy its accounting agent roles and govern them on the board.

That is worth stating plainly because the hosted agents on the product's own site are coding agents. The governance mechanics, the done-gate, the deny-by-default gateway and the delivery metrics, are general; the accounting roster is our design applied to them. For the plain-language product summary, see what Cognibl is.

How small can a practice be and still run a desk?

A practice of any size can run a practice operations desk, because the design scales down to a single agent role and a single reviewer. A sole practitioner might automate only records chasing for VAT clients and review every item personally. What a small practice should not do is skip the reviewer step, since that is the part that keeps the professional duty where it belongs.

The difference between a small and a large desk is mostly the number of reviewers and how the queue is shared between them. A ten-person practice and a regional firm face the same January wall in proportion; the larger firm simply has more reviewers to route to and more reason to watch human wait share closely.

Want to see this in practice? Read about the Cognibl platform, explore our AI solutions for professional services, browse more industry guides, or read about our approach to placing AI inside regulated firms.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.