AI Agent Governance for Recruitment Agencies in 2026

Two dates now sit on every UK recruitment agency's back office. Since 6 April 2026, the agency that holds the contract with the end client has been responsible for making sure PAYE is operated correctly when an umbrella company employs its workers, and HMRC's guidance on labour supply chains says it can recover any underpayment from that agency. Since 1 October 2026, right to work checking has reached beyond employees to workers, individual sub-contractors and online matching services. HMRC's own policy paper expected the umbrella change to have a significant impact on approximately 30,000 agencies.

Both changes land on the same people: the compliance lead, the payroll and finance team, and the operations director who signs the supplier list. Both also create the same kind of work. It is not glamorous work. It is checking, chasing, reconciling and filing, every week, for every umbrella and every worker on assignment, with a record kept of each step. It is also the kind of repetitive, rules-based work an agency might want to hand to an AI agent, and it is exactly the work where an agent that says "done" without evidence is a liability rather than a saving.

This guide is about governing that agent work. It sets out how a placement back-office desk can split the HMRC due diligence checks and the right to work follow-ups between agents and accountable people, what a done record for each task should reference, and where the current limits of Cognibl, from DILR.AI, sit for a recruitment agency. It deliberately cedes two things. The legal detail of the umbrella rules and the right to work extension lives in our guide to where AI pays in UK recruitment, and the work-seeker paperwork at registration lives in our first-call work-seeker terms guide. Here the subject is the desk itself.

This guide is shipped by the team behind Cognibl, a work-management platform where people and AI agents share one board and a task reaches done only once proof is attached. Or see DATS, our five-stage AI consulting system for placing AI inside regulated operations.

What does AI agent governance mean for a recruitment agency's back office?

For a recruitment agency, AI agent governance means every back-office task an agent touches, from an umbrella due diligence check to a right to work follow-up reminder, has a written definition of done, a named person who accepts it, and a stored record of what the agent actually did. The test is simple: could the compliance lead show HMRC or a client auditor the evidence behind any task marked complete?

Most agencies already run some form of this for people. A compliance administrator ticks off an umbrella's VAT registration on a spreadsheet, and a manager spot-checks the file before an audit. The difference with agents is volume and confidence. An agent can run the same check across every umbrella on the preferred supplier list overnight, and its own summary of the run is not evidence that it opened the source it was meant to read. Governance is the layer that refuses to accept the prose and asks for the evidence instead.

That is the core idea behind the proof of done approach to agent work management: work is finished when it is proved, not when it is claimed. For a recruitment back office it translates into three practical rules. First, the agency writes down what each recurring check must produce before any agent is allowed near it. Second, an agent's output is a draft until an accountable person accepts it. Third, the record of each run is kept somewhere the agent cannot quietly edit.

The ICO's guidance on AI and data protection frames the same expectation from the data side: accountability and governance, impact assessment and transparency sit with the organisation running the AI, which for candidate and worker data is usually the agency as controller. None of this is new law aimed at agents. It is ordinary accountability, applied to a new kind of worker. More examples of this pattern across regulated sectors sit in the industries category of the blog.

Why did 6 April 2026 turn umbrella due diligence into a records problem?

From 6 April 2026, a UK recruitment agency that contracts with the end client is jointly and severally liable for PAYE an umbrella company in its chain fails to pay, and HMRC can pursue the agency first. Because HMRC's guidance asks agencies to keep a record of their umbrella due diligence, the question an agency must answer shifts from "do we check?" to "can we prove what we checked, and when?"

HMRC's policy paper on the umbrella company market states its intent plainly. In its words, "This measure is intended to encourage increased due diligence among businesses that choose to use umbrella companies to engage workers". The same paper estimates around 700,000 individuals work through umbrella companies, and the HMRC Employment Status Manual entry ESM2410 confirms the rules bite on qualifying payments made on or after 6 April 2026, not before. HMRC's supply chain guidance also lists arrangements the rules do not normally cover, such as workers engaged through their own personal service company where the off-payroll conditions are met.

The practical guidance sits in HMRC's page on responsibilities for employment businesses working with umbrella companies. It is worth quoting the record-keeping line exactly, because it is the sentence that turns a compliance habit into an evidence duty: "You should keep a record of the due diligence checks that you have undertaken on an umbrella company." The same page says the checks should be regular, reasonable and proportionate for the business, and, on involvement in non-compliant supply chains, it adds a warning every compliance lead should pin to the wall: in most circumstances, claiming that you were not aware of non-compliance is not a defence.

HMRC's December 2025 Employer Bulletin told agencies to prepare by reviewing their labour supply chains and conducting their own due diligence, including reducing the risk of using an umbrella that operates a tax avoidance scheme. That is the workload this post designs a desk around. We cover the wider liability mechanics, and who the duty binds in a chain with no agency, in the recruitment industry guide; if you are mapping a broader governance model first, our AI operating model work starts there.

Which umbrella checks can an agent prepare, and which must a person sign?

Most of HMRC's named umbrella checks can be prepared by an AI agent, because they are lookups and comparisons: Companies House details, VAT registration, insurance and accreditation status, and payslip-to-assignment-rate reconciliation. Deciding whether an umbrella stays on the supplier list, and judging a suspected skimming or avoidance pattern, should stay with an accountable person in the agency, with the agent's evidence attached.

The table below takes the umbrella checks HMRC names in its guidance that suit agent preparation, then adds the supplier-list decision they feed. For each row it shows what an agent can prepare, what the done record should reference, and who accepts it. It is a design proposal, not an HMRC template. HMRC lists the checks; how an agency divides the labour is its own decision, and the agency remains the party HMRC can pursue.

HMRC's list goes further than the table: it also asks agencies to understand how workers are engaged and paid, to look for National Minimum Wage underpayment, to check a GLAA licence where the sector requires one, and to check that umbrellas carry out and record right to work checks, which the next section covers. Two design choices in that table matter more than the rows. The first is that every row produces a record of what was read, not just a pass or fail. A green tick with no source behind it proves nothing when HMRC or an end client asks what was actually checked. The second is that the decisions sit in the judgement rows and the final row: whether a deduction pattern is skimming, whether a mini umbrella warning sign is real, and whether an umbrella stays on the list are human calls. An agent that flags is useful. An agent that decides, on a tax liability that falls on the agency, is not something we would recommend any agency build.

HMRC also says to keep supply chains as short as possible and to train staff to spot non-compliance. A short chain makes every row cheaper, whoever runs it, and the training point is a reminder that the person accepting an agent's check needs to know what a wrong answer looks like. If your agency is still deciding where agents belong at all, an AI placement diagnostic ranks the candidate workloads before anything is built.

How does the 1 October 2026 right to work extension change what the desk records?

Section 48 of the Border Security, Asylum and Immigration Act 2025, in force from 1 October 2026, extends the illegal working provisions to people engaged under a worker's contract, individual sub-contractors and online matching services. For a recruitment desk this widens the population whose right to work evidence must be tracked, but it does not let an AI agent perform the statutory check itself.

The commencement is set by SI 2026/683, and the scope is in the text of section 48 itself, which inserts a new section 14A into the Immigration, Asylum and Nationality Act 2006. The duty binds the business doing the engaging. GOV.UK's page on penalties for employing illegal workers puts the civil penalty at up to £60,000 for each illegal worker, which is why this is a board-level number rather than an administrative one.

The prescribed check is narrow and specific. GOV.UK's guidance on checking a job applicant's right to work allows three routes: an online check with the applicant's share code, a check of original documents, or an identity service provider using identity document validation technology. British and Irish citizens cannot get a share code. Copies are kept during employment and for 2 years after, the date of the check is recorded, and time-limited permission triggers a follow-up check before it expires.

None of those routes is "an agent read a photo of a passport". That is why we keep the agent's role on the right to work side deliberately small:

• Before the check, an agent can chase a worker for a share code, confirm an appointment slot, and make sure the right person is booked to run the check. Our candidate screening guide covers the calls that usually start this.
• After the check, an agent can confirm that the record exists in the agency's own compliance system, that the check date is filled in, and that the expiry date for time-limited permission has a follow-up task scheduled.
• For umbrella-employed workers, HMRC's guidance says employment businesses should check that umbrella companies are carrying out and recording right to work checks. An agent can request and log the umbrella's confirmation for each worker, so a gap shows up as an open task rather than a surprise.

What the agent never does is make the eligibility decision or store the identity documents in its own task record. That boundary also keeps the agency inside the employment law limits set out in our guide to AI decisions in UK recruitment law.

What should a done record contain for a placement back-office task?

A done record for a recruitment back-office task should reference the source the agent read, the date and time of the run, the result against the written definition of done, any exceptions found, and the person who accepted it. It should point to identity documents held in the agency's own compliance system rather than copy them into the task, so the evidence trail never becomes a second store of sensitive data.

This is where the mechanics of a tool start to matter. In Cognibl, a task reaches a done status only once proof is attached: per the Cognibl product page, that proof is 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, which matters on a mixed desk where a compliance administrator and an agent may close the same type of task on the same day.

For a recruitment desk, the CSV is the useful part. A Companies House check, for example, produces a row per umbrella with the lookup date, the fields compared and the mismatches found. A reconciliation run produces a row per worker sampled, the assignment rate, the statement figure and the variance. The proof references the files read; it does not need to contain a payslip or a passport. On cognibl.com, the product tour adds that Cognibl checks the proof against the task's definition of done before the task can move to done, and only a passing proof moves it.

Two cautions keep this honest. First, the record of tool calls is not a full transcript. Cognibl's connections page says every call through the gateway, whether allowed, refused or failed, is written to the project's trace, append-only, hash-chained and redacted at the moment of capture. Redaction is the right default for candidate data, but it means the trace shows that a call happened and how it ended, not every value the agent saw. The CSV proof is where the reviewable detail belongs. Second, a done record proves what the agency checked. It does not prove the umbrella paid HMRC, and it does not stop HMRC recovering underpaid PAYE from the agency where the rules apply.

We go deeper on why an agent's own report should never be the record in our guide to verifying AI agent tasks, and the banking version of the same audit question is in our post on agent audit logs in banking.

Where do Cognibl's current limits sit for a recruitment desk?

Cognibl, from DILR.AI, is today strongest as the board and the done-gate for agent work, not as an agent that writes into an agency's payroll or compliance system. Its live pages say connected tools are read-only by default, with Google Workspace the only tool where an admin can allow writes, and a staging queue that holds a write until approval is not built yet. A recruitment desk should be designed around those limits.

We would rather an agency hear this from us than find it in a pilot. Cognibl's homepage leads with coding agents on Claude Code and Codex, each opening a pull request with proof of work that only a person can merge. Its connections page links GitHub, Google Workspace, Linear, Notion and 90 more tools through the Model Context Protocol, and each project chooses what its agents can reach. For a recruitment agency that means an agent can read sources and produce evidence, and the agency's own staff still update the ATS, the payroll system and the supplier list.

One limit deserves particular care. Where an admin does allow Google Workspace writes, Cognibl's connections page is explicit that writes go straight to the tool, with nothing held for approval. On a recruitment desk, a Workspace write could be an email to a worker or an umbrella. Our recommendation is plain: keep Workspace at read level for any project that touches worker or umbrella correspondence until your own process for reviewing outbound messages is in place. On dilr.ai, the same platform's gateway is described as deny by default, so a toolset that has not been enabled for a project is refused rather than quietly missing, and every write is attributed by key name.

What Cognibl does provide fits the recruitment problem well. Recurring tasks can run hourly, daily, weekly or monthly, and each run is a new task with its own proof, which suits a weekly umbrella sweep or a monthly reconciliation sample. Two AI flows, proof validation and project status, summarise and flag problems but never decide. And the skills and agents library is versioned, so the instruction an agent ran in March is still the instruction you can read in September, and items archive rather than delete.

Recruitment is also an inferred industry for Cognibl rather than a proven one. We have mapped it from product fit with the work, not from a published recruitment customer, and this post says so plainly.

How should an agency pilot a governed back-office desk?

An agency should pilot a governed back-office desk on one recurring check, one umbrella cohort and one accountable person for four to six weeks, measuring how often the agent's first proof passes review and how long tasks wait for a human. Only when the record holds up under an internal audit should the desk take on a second check or a larger share of the supplier list.

The sequence we would use, in order:

1. Pick the check with the clearest source. Companies House matching is the usual first choice, because the source is public, the comparison is mechanical and a wrong answer is easy to spot.
2. Write the definition of done first. What sources, what fields, what counts as an exception, and what the CSV must contain. The compliance lead owns this document, not the agent vendor.
3. Run agent and person in parallel for two cycles. The person does the check as today; the agent does it into the board. Differences are the most useful data the pilot produces.
4. Move to agent-first with human acceptance. The agent runs, the person accepts or rejects each proof, and rejections go back with a reason. Acceptance is the agency's own review step, recorded next to the proof; it is a process the agency runs, not a separate approval feature in the tool.
5. Audit the file, not the dashboard. Before expanding, pick five umbrellas at random and ask someone outside the desk to reconstruct each decision from the accepted records alone.

Cognibl's five delivery metrics include first-pass verification rate by flow type, human wait share, verified throughput per week and reopen rate within 30 days. On a compliance desk these line up with the questions a pilot has to answer: how often an agent's proof passes first time, how long work sits waiting for a person, how much verified work the desk completes, and whether accepted work is later reopened.

There is a sector reason to insist on that discipline. APSCo's Recruitment Index 2025 found 34% of UK staffing companies trialling specific AI tools and 32% actively integrating AI, yet just 6% said it was demonstrating tangible benefits. A pilot that measures verified output rather than activity is one way to find out quickly which side of that gap a back-office desk lands on. The operating rhythm for running pilots like this, with senior practitioners embedded rather than advising from outside, is how our AI execution office works.

The same desk pattern has been worked through for other regulated professions. Our practice operations desk for accounting firms is built around filing deadlines, and the matter operations desk for law firms is built around supervision duties. The recruitment version differs in one important way: the record is not mainly for the agency's own quality control. It is the evidence an agency would want to show HMRC, a Home Office officer or an end client asking how its supply chain is checked.

Where do the other DILR.AI lines fit around the desk?

Around a recruitment back-office desk, the other DILR.AI lines handle the work the desk should not. Dilr Voice handles candidate calls, such as share code chases and first-call screening, while DATS designs the governance model and the evidence layer. Dilr Academy trains staff who accept agent output, and Dilr Mira, a class of private clinical small language models, is not offered for recruitment documents today.

On the front of the desk, Dilr Voice handles the high-volume calls that feed the back office, including the first-call work-seeker conversation and share code reminders. Our post on voice AI for staffing shift-fill shows how that call layer works on a temp desk. The governance design and the evidence layer behind both are DATS work, delivered through our approach of placing AI inside the systems a firm already runs.

HMRC's guidance tells agencies to train staff to identify non-compliance in their supply chains, and the person accepting an agent's proof needs the same skill. Dilr Academy is how we would build that AI fluency for compliance and operations teams. Dilr Mira is a class of private clinical small language models built for healthcare documents, and we do not offer it for recruitment paperwork today. DILR Studio does not apply here: we see no named recruitment back-office use for it.

What is the best way to govern AI agents in a recruitment agency in 2026?

The best way to govern AI agents in a UK recruitment agency in 2026 is to put every agent task on a board with a written definition of done, a proof record per run and a named human who accepts it. Cognibl, from DILR.AI, fits agencies wanting that done-gate for read-and-check work; agencies wanting agents to write into payroll or ATS systems should wait or look elsewhere.

How we would judge any option, ours included:

• Does a task's completion require evidence, enforced by the system rather than by habit? This is Cognibl's design centre. Mainstream work-management tools such as Monday.com, Asana and Jira are mature boards with their own automation; buyers should check for themselves whether any of them refuses a status change until a proof file is attached, because that is the property this whole desk depends on.
• Does the agent reach your recruitment systems directly? Agencies running Bullhorn or Vincere should look first at what their ATS can automate in place, because that keeps the work where the data already lives. For simple in-system reminders, that can be the better choice.
• Can an outsider reconstruct a decision from the record? Whatever the tool, run the five-umbrella audit test from the pilot before trusting it.
• Who owns the definition of done? It should be the agency's compliance lead, written down, versioned and reviewed when HMRC updates its guidance.

The honest concession: an agency with a handful of umbrellas and one compliance administrator may not need agents at all. HMRC says checks should be proportionate, and for a short chain a well-kept spreadsheet with a monthly manager review can be the proportionate answer. Agents earn their place when the supplier list, the worker count or the frequency of checks outgrows what one person can evidence. If you want a second opinion on where that line sits for your agency, book a call and we will tell you plainly.

Can an AI agent carry out a right to work check?

No. An AI agent should not carry out the statutory right to work check. GOV.UK allows an online share code check, a check of original documents, or an identity service provider using identity document validation technology, carried out by or for the employing business. An agent can chase share codes, schedule checks, confirm the record exists and schedule follow-up checks for time-limited permission, but the check and the decision stay with the agency.

Does a clean due diligence record protect an agency from umbrella PAYE liability?

Not on its own. Under the umbrella rules in force since 6 April 2026, where they apply to a supply chain, HMRC says it will recover underpaid PAYE from the agency or end client. A record of regular, proportionate checks shows what the agency checked and helps it spot a failing umbrella early, which is the practical protection, but HMRC does not present it as a way to avoid the debt.

Should candidate documents be attached to an agent's proof record?

No. A proof record for a recruitment back-office task should reference documents held in the agency's own compliance system, such as the check date, method and file location, rather than copy passports, share code results or payslips into the task. Keeping sensitive data in one controlled store makes retention, access and the two-year right to work copy rule far easier to manage.

Want to see this on your own back office? Explore the Cognibl work-management platform, read what Cognibl is and how it works, see the DATS consulting system, or learn more about DILR.AI and the team behind it.

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.