AI can identify unmatched items, assemble evidence and draft variance commentary. It cannot make a weak close reliable by writing a persuasive explanation over unreconciled numbers. The correct target is a shorter path from source transaction to reviewed balance, with every adjustment, judgement and sign-off still visible.
This guide is current to 31 July 2026 and addresses UK finance teams preparing management and statutory information. Applicable accounting frameworks depend on entity type, size, group status and listing. Tax, charity, financial-services and public-sector requirements add separate duties. Directors, finance leaders, auditors and advisers retain their responsibilities; this is not accounting or legal advice.
Break the close into control units
Do not deploy one agent across the general ledger. Map tasks by input, control objective and authority:
| Unit | Safe initial AI role | Authoritative output | Owner |
|---|---|---|---|
| Bank reconciliation | Suggest matches and exceptions | Approved reconciliation | Cash accountant |
| Accrual support | Locate purchase and service evidence | Reviewed journal proposal | Cost-centre owner and finance |
| Intercompany | Compare counterparty records | Agreed difference and resolution | Entity controllers |
| Variance commentary | Draft from locked numbers and drivers | Signed management narrative | FP&A lead |
| Close status | Identify overdue tasks and dependencies | Approved close dashboard | Group controller |
| Statutory disclosure | Check completeness against a checklist | Accounts approved under governance | Directors and reporting team |
Every unit needs a state model: not started, prepared, exception, reviewed, approved, posted and reopened. Prevent the AI from marking its own work reviewed or bypassing a segregation-of-duties rule.
Freeze definitions before measuring speed
Baseline at least three ordinary closes and one difficult period if available. Record elapsed days, staff hours, late journals, reopened reconciliations, post-close adjustments, audit requests, spreadsheet failures and overdue dependencies.
Agree:
- close start and finish events;
- materiality or investigation thresholds by account;
- who prepares and who approves each control;
- acceptable supporting evidence;
- treatment of late invoices and subsequent events;
- journal classes permitted for automated preparation;
- hard posting cut-offs and reopening authority; and
- the source of each management metric.
The measure is not “days to first pack”. Track days to a pack that survives review. A faster close with more late corrections has only shifted work beyond the reported finish.
Useful balancing measures include unreconciled value, aged reconciling items, manual journals, reviewer rejection rate, time spent answering evidence queries and adjustments discovered after approval.
Protect the accounting record
UK companies must retain sufficient records to show and explain transactions. Current government guidance on company and accounting records lists money received and spent, assets, liabilities, stock and supporting material. Companies House’s June 2026 accounts preparation and filing guidance also states that every company must keep accounting records.
An AI-generated summary is not source evidence. Preserve the invoice, contract, bank statement, goods receipt, payroll report or approved calculation. Link every suggested match, accrual and explanation to immutable source identifiers and the source-system timestamp.
Use this evidence hierarchy:
- external statement or counterparty document;
- controlled subledger or operational system;
- approved calculation with source inputs;
- corroborated correspondence;
- model-generated suggestion; and
- unsupported narrative, which cannot close an exception.
Do not let a vector index become the only route to old evidence. Retention, legal hold and audit export must work independently of the model supplier.
Configure the applicable reporting framework
The FRC’s current FRS 102 page notes that the Periodic Review 2024 amendments have a principal effective date of 1 January 2026. It also lists later amendments with their own commencement dates. Teams must configure by accounting period, not assume “latest” means applicable.
Build a framework register containing:
- entity and consolidation scope;
- reporting framework and effective edition;
- company-size assessment and exemptions;
- accounting policies and elections;
- materiality used for management and statutory work;
- disclosure owner;
- auditor or adviser interpretation where relevant; and
- transition adjustments and comparative treatment.
When a rule changes, version the checklist and show which entities and periods are affected. A language model can retrieve relevant paragraphs, but the technical accounting conclusion belongs to a qualified reviewer.
HMRC’s July 2026 guidance on events after the reporting period distinguishes adjusting and non-adjusting events under FRS 102. Configure a post-balance-sheet event workflow that records condition date, discovery date, evidence, materiality and approval rather than asking AI to decide from a news feed.
Use AI for reconciliation without losing control
Start with a high-volume, low-judgement reconciliation such as known bank receipts or supplier statement lines. Define deterministic matches first: identifiers, exact amount and permitted date window. Use AI only for residual cases where text or document context adds value.
For every suggestion, show:
- source and ledger lines;
- amount, currency and date difference;
- extracted reference;
- rule or model version;
- confidence calibrated on comparable cases;
- prior linked items;
- reason for the proposed match; and
- reviewer action and timestamp.
Block many-to-many matches above a threshold, old reconciling items, unusual currencies, related-party items and possible duplicates for specialist review. Do not train on approvals without distinguishing later reversals; an initially accepted match may have been wrong.
Recalculate the balance outside the AI service. The signed reconciliation should include opening balance, movements, source closing balance, ledger closing balance, reconciling items and subsequent clearance.
Draft commentary from locked numbers
Variance commentary is a useful generative task when its data contract is strict. Supply the approved current period, comparison, currency, scale, threshold, dimensions and named driver evidence. Require the draft to cite those cells or records.
Ban vague filler such as “market conditions” unless a verified source supports it. Separate:
- arithmetic contribution;
- operational explanation;
- uncertainty or missing evidence;
- corrective action;
- owner and due date; and
- expected future effect.
The model must not calculate an alternative total in prose. Render numbers from the reporting cube after the text is approved, and rerun consistency tests immediately before publication.
Review for unsupported causality. A cost rose after volume increased, but the relationship may involve price, mix and timing. Keep “observed”, “management believes” and “forecast” distinct.
Control journals and privileged access
During the pilot, AI may prepare a journal package but cannot post it. The package must include entity, period, accounts, amounts, currency, description, supporting evidence, preparer, approver, reversal date and policy basis.
Enforce segregation of duties in the ERP, not only in an orchestration layer. Use least-privilege service accounts, time-bound elevated access and multi-factor authentication. Prevent the model from changing vendor bank details, chart-of-account mappings, accounting periods or approval limits.
Pause journal automation for:
- management override or unusual top-side entries;
- revenue recognition or contract modifications;
- provisions, impairments and valuations;
- related parties and intercompany disputes;
- tax, pension or share-based payment judgements;
- acquisition, disposal or restructuring;
- suspense accounts above threshold; and
- entries requested through unverified email.
The FRC’s March 2026 guidance for audit firms using generative and agentic AI is directed at audit engagements, not company close teams, but its focus on confidence in output quality and wider quality management is instructive. A client system should never imply that using AI creates audit assurance.
Align internal control with accountability
For companies applying the UK Corporate Governance Code 2024, the FRC’s updated Corporate Governance Code guidance is the relevant context. Other entities can still use its control principles proportionately without claiming Code compliance.
Maintain a control register with objective, risk, frequency, evidence, preparer, reviewer, system dependency and failure route. Record overrides and reopened tasks. A model update is a system change: assess it, regression-test it and approve it before the close window.
If the AI drafts board reporting, management remains accountable for completeness and balance. Present significant estimation uncertainty, control failures and unresolved exceptions rather than optimising the narrative for confidence.
Prepare for filing and tax changes accurately
Companies House guidance current at this cutoff still supports existing filing routes, subject to eligibility. The June 2026 announcement on software-only accounts filing from April 2028 is a future reform, not a current month-end requirement. Use the lead time to preserve structured data, but do not tell teams that paper or web filing has already ended for all accounts.
Making Tax Digital for Income Tax began in April 2026 for qualifying unincorporated businesses and landlords over the stated income threshold. HMRC’s digital record-keeping direction and record-creation guidance concern that population; they are not a universal company-close rule.
Keep statutory, tax and management calendars linked but distinct. A closed management period may still require tax adjustments or statutory events.
Secure finance data and integrations
Finance systems expose payroll, bank, customer, supplier and transaction data. Apply the NCSC’s secure AI system-development guidance: isolate environments, control models and connectors, log access, test recovery and manage suppliers.
Exercise:
- prompt injection in an invoice or spreadsheet comment;
- a duplicate bank feed;
- altered supplier master data;
- an out-of-period document;
- currency or unit mismatch;
- cross-entity retrieval;
- model outage on close day;
- compromised approver credentials; and
- a model upgrade that changes match rates.
Maintain a manual close checklist, read-only evidence export and tested ERP recovery. Do not let a supplier outage block statutory record access.
Run a 90-day close pilot
| Period | Work | Gate |
|---|---|---|
| Days 1–15 | Baseline tasks, exceptions and controls; select one reconciliation | Control owner and success measures approved |
| Days 16–35 | Build evidence links, test set, access and model-change process | Source-to-balance traceability passes |
| Days 36–55 | Shadow suggestions across two closes or equivalent cycles | Precision and review burden beat baseline |
| Days 56–75 | Limited live preparation with independent approval | No AI posting or control bypass |
| Days 76–90 | Compare close time, errors, evidence and full cost | Scale, revise or stop |
Pause for an unexplained balance difference, unsupported journal, changed source evidence, segregation-of-duties breach, cross-entity disclosure, high-severity cyber event or output that contradicts the ledger. Reopen the close when a material subsequent fact or control failure emerges; meeting a dashboard date is not a reason to suppress it.
Scale only when the control remains effective, net reviewer time falls, aged exceptions reduce and post-close adjustments do not rise. Related archive guides cover AI tax and accounting automation, AI in finance and fraud control-management-uk-2026) and UK AI cybersecurity.
Close faster by making evidence easier
The durable benefit is not prose generated at midnight. It is a close in which evidence arrives earlier, exceptions are visible, repetitive comparisons are automated and professional judgement receives more time.
If the system cannot reproduce a number, its source and its approvals after the period is locked, the month-end has not been automated. It has been made harder to audit.
