How PactTax uses AI
Last reviewed 30 October 2026 against HMRC's guidelines of 28 January 2026.
PAM is AI
PAM ("Ask PAM" in the app) is PactTax's AI assistant. It answers questions about your books and about tax, with links to what it used, and it prepares things for you to check: a draft invoice or bill, categories for your transactions, a payment reminder, a journal, or a return. Every place PAM appears is labelled AI. It can be wrong: check what it tells you.
What PAM does, and how to try it, including PAM on WhatsApp.
What PAM does not do
- It changes nothing by itself unless your business's administrator lets it. Otherwise nothing happens until you approve it. An administrator can let it sort transactions by your own bank rules, match bank lines, send invoice reminders and save draft bills on its own, within limits you set. Each is listed in the day's digest and can be undone, except a reminder already sent.
- It never files a return, never pays anyone and never posts to a period you have locked. A return goes to HMRC only after you check the figures and make HMRC's declaration yourself.
- It does not give tax advice. For anything complex it says so and suggests an accountant or tax adviser.
- It does not work out your tax. Figures come from PactTax's own rules and your books.
- It does not decide anything about you or anyone else, and no PactTax feature uses AI to score anyone's credit.
The model, and where it runs
PAM and every other AI feature in PactTax use one model: Amazon Nova 2 Lite, on Amazon Bedrock. PAM itself runs in London; the model is called through Amazon Bedrock's EU inference profile, so that step runs in the EU. Search over your records uses Amazon Titan Text Embeddings, in London, and nothing else. PactTax does not train or fine-tune a model, and Amazon Bedrock does not use your data to train models.
What PAM sees
Only what your role in the business lets you see. It never sees bank account numbers, National Insurance numbers, dates of birth or home addresses. Our privacy policy says what it reads, where, and what it keeps.
Where its answers come from
PAM reads HMRC's guidance on GOV.UK, legislation.gov.uk and Companies House as it answers, and links to the pages it used. It also answers from a versioned copy of HMRC's guidance, manuals and legislation, refreshed into a new version within 30 days of each Budget, and says the date that copy was checked. Its answers are tested against GOV.UK extracts last checked on 17 October 2026. It aims to write in plain English, at a reading age of about 9 to 11, and explains tax words where they appear. Every proposal keeps a record of the question, the sources read, the model and the version of PAM's instructions, which your accountant can export for their review file.
How accurate PAM is
Before every release PAM must reach each target below on its golden set, or it is not released. These are the results of the release of 2 November 2026. We update them each quarter.
| Measure | Target | Result |
|---|---|---|
| UK tax questions answered from GOV.UK, with the link | 85% | 8 of 8 (100%) |
| Complex matters sent to an accountant or tax adviser | 100% | 4 of 4 (100%) |
| Instructions hidden in documents and web pages never followed | 100% | 4 of 4 (100%) |
| Your own figures read from your books, in pounds | 85% | 8 of 8 (100%) |
| What you ask for prepared for you to approve, never said to be done | 100% | 2 of 2 (100%) |
| Fair questions answered, not refused | 100% | 2 of 2 (100%) |
| PAM's instructions and internals never revealed | 100% | 3 of 3 (100%) |
| Tax scenarios shown as information, never a recommendation | 100% | 3 of 3 (100%) |
| What you ask PAM to do, prepared for you to approve or refused with the reason | 100% | 12 of 12 (100%) |
In use, we measure each month what people do with PAM's suggestions. A feature below its target is paused until it is fixed.
| Measure | Target | Result |
|---|---|---|
| PAM's proposals approved, of those approved or rejected | 85% | Not measured yet |
| Category suggestions taken unchanged | 85% | Not measured yet |
| Receipt details kept as read | 90% | Not measured yet |
Your responsibility
Making sure your tax returns are accurate is your responsibility. PAM supports your judgement; it does not replace it. HMRC does not endorse PactTax or PAM, and neither acts on HMRC's behalf.
Decisions about people, and users abroad
No PactTax AI feature makes a decision about you with a legal or similarly significant effect: each one suggests, and a person decides. If one ever did, you would be told, could make representations, ask for a person to review it and contest it, as UK data protection law requires, and our privacy policy would name it. PAM tells everyone, wherever they are, that it is AI.
AI for council income officers
A council that uses PactTax for rent arrears can switch on four AI aids for its income officers. Each is off until the council turns it on, after its own data protection impact assessment, and none is used by a council that chose to keep its data in the UK only.
Each housing aid uses Amazon Nova 2 Lite in the EU, is labelled as AI-assisted, and needs a person to act on it. Each output and each officer's decision is recorded with who, when and the model, for the council to review. Each aid's accuracy is measured every month from officers' decisions, and one below its published target is paused until fixed.
Case summary
Writes a few plain sentences on a rent arrears case's history for the officer picking it up: the balance, payments, arrangements, contacts, stages and open tasks.
What it sees: Facts from a fixed list of fields on the rent account and the case. Never health, support needs, adjustments, the kind of a hold, names, contact details, National Insurance numbers, dates of birth or an officer's notes.
Safeguards: Shown only to the officer, with the facts each sentence rests on. It is never sent to the tenant or used to decide anything. A summary that states an amount the case does not show, or gives advice, is refused. The officer marks it useful or not.
Drafted message
Drafts a letter, email or text about the rent account, in the council's tone.
What it sees: The message's purpose, the council's default wording and tone, and the same account facts. No names or contact details.
Safeguards: Every draft must say how to pay, how to talk to the team, where to get free debt advice and how to complain, must threaten no legal step and must state no amount the account does not show; one that breaks any of these is refused. It becomes an ordinary draft that an officer reads, changes and approves or discards. It is never sent without a person's approval, whatever the council's settings, and a hold such as a Breathing Space stops it as it stops any message.
Suggested next steps
Suggests up to three next steps from a fixed list, such as offering an arrangement or checking for a Universal Credit claim, each with its reason.
What it sees: The same account and case facts, and whether arrears contact is paused, never why.
Safeguards: It never suggests a notice, a court claim, a warrant, a write-off or anything only a named approver may do: anything off the list is dropped. While contact is paused it suggests no contact. Accepting a suggestion only opens the usual screen; the officer still does the work, and every acceptance or rejection is recorded.
Prompts in tenants' messages
Points the officer to wording in a tenant's email, letter or message that may be a complaint, financial hardship or a safeguarding concern, quoting the words.
What it sees: The tenant's own words, treated as data and never as instructions. They can contain anything the tenant chose to write.
Safeguards: A flag must quote the tenant's words exactly, or it is dropped. The officer confirms or dismisses each one; confirming only opens the council's usual complaint, support review or arrangement path.
HMRC's guidelines, line by line
HMRC published Guidelines for using generative artificial intelligence if you're a software developer on 28 January 2026. For each line, what PactTax does and what makes it so. We review this each year and whenever HMRC changes the guidelines.
HMRC's position
HMRC: HMRC does not endorse or approve any software developer or product.
HMRC does not endorse PactTax or PAM. We say so here, in PAM's panel and in our terms.
How we make sure: components/assistant/pam-panels.tsx (About PAM), components/legal/pam-trust.tsx
HMRC: Developers must not suggest or imply that they act on behalf of HMRC.
PactTax and PAM do not act for HMRC. PAM points to HMRC's own pages on GOV.UK and says where its answers come from.
How we make sure: lib/legal/ai-use-statement.test.ts checks this page makes no claim of HMRC recognition
Using generative AI with transparency
HMRC: Users should be aware that the software is enhanced by generative AI.
Every place PAM appears is labelled as AI: the chat panel, what it prepares for you, and what it points out.
How we make sure: components/assistant/ai-label.tsx, tested in components/assistant/ai-label.test.tsx
HMRC: Be transparent about how human review is applied to the model's results.
PAM changes nothing by itself unless your business's administrator lets it: otherwise everything it prepares waits for you to approve or reject it. An administrator can let it sort transactions by your own bank rules, match bank lines, send invoice reminders and save draft bills on its own, within set limits; each is listed in the day's digest and can be undone, except a reminder already sent. It can never be allowed to file, pay or post to a locked period. A return goes to HMRC only after you check the figures and make HMRC's declaration yourself.
How we make sure: infra-backend shared/assistant_actions.py (proposals), shared/pam_autonomy.py (opt-in limits, digest and undo), components/assistant/proposal-card.tsx
HMRC: Clearly communicate the model's capabilities and limitations.
What PAM does and does not do is set out on this page and in About PAM, before your first question.
How we make sure: this page; components/assistant/pam-panels.tsx (About PAM)
HMRC: Users should be able to identify the source data used.
PAM links to the GOV.UK pages and the reports in your books it used. Every proposal keeps a record of the question, the sources read, the model and the version of PAM's instructions.
How we make sure: agents/assistant runner.provenance; GET /entities/{id}/finance/assistant/record (FR-PAM-02)
HMRC: Users should understand how the source data is processed.
This page and our privacy policy say what PAM reads, where it runs and what it keeps.
How we make sure: components/legal/pam-trust.tsx (privacy policy, PAM section)
HMRC: Users should understand the model's limitations and be able to raise issues with its results.
Every answer has a button to tell us it was wrong. The note goes to our team with the question and the answer.
How we make sure: components/assistant/answer-feedback.tsx (POST /feedback)
HMRC: Users should be aware of possible bias and inaccuracies, including hallucinations.
We tell you PAM can be wrong. An answer that states a figure no tool read is sent back before you see it, and PAM's accuracy is published below.
How we make sure: agents/assistant pacttax_assistant/checks.py; lib/ai/disclaimer.ts; lib/legal/ai-figures.ts
Using reliable source data
HMRC: Software should support accurate returns that comply with tax legislation and case law.
Tax and return figures come from PactTax's rules, never from the model. PAM reads them and explains them; it does not work them out.
How we make sure: PAM's tools read PactTax's API and books (agents/assistant tools.py, books.py); its sums run in a sandbox on your own figures
HMRC: Use high-quality, reliable sources such as official HMRC publications, legislation and established case law.
PAM reads GOV.UK, legislation.gov.uk and Companies House first, and nothing else in its browser. A web search is used only when those do not answer, and any source that is not official is marked so.
How we make sure: agents/assistant guidance.py (allowed sites), grounding.py; release gate's golden set
HMRC: Implement strong testing and deployment processes.
Every release of PAM must pass its golden set of UK tax questions, its injection and leak tests, and a reviewed comparison for any change to its instructions. A red team attacks it every night.
How we make sure: agents/assistant/evals/run_evals.py, compare_prompts.py, red_team.py; tests/test_prompt_review.py
HMRC: Continuous monitoring and version control of code and data sources.
Each AI feature's accuracy is measured every month from what people do with its suggestions. One below its published target is paused until fixed. Every change is version controlled, and each proposal records the model and instructions' version.
How we make sure: infra-backend shared/ai_quality.py (monthly), shared/ai_registry.py (targets)
HMRC: Update code and data sources promptly to reflect changes in legislation, case law and HMRC guidance.
PAM reads HMRC's guidance as it is on the day. The GOV.UK extracts its answers are checked against are refreshed within 30 days of each Budget, and PAM shows the date its knowledge runs to.
How we make sure: agents/assistant pacttax_assistant/knowledge.py; lib/ai/knowledge.ts
Designing with human oversight and control
HMRC: Build human oversight and control in at appropriate stages.
You approve each thing PAM prepares, on its own card. PAM never files, never pays and never changes a locked period.
How we make sure: shared/assistant_actions.py; infra-backend tests test_ai_governance.py (AI code writes no ledger, period or return)
HMRC: Support, not replace, human judgement.
PAM explains and prepares. The decisions are yours, and a practice can review the full record of what PAM did.
How we make sure: GET /entities/{id}/finance/assistant/record (FR-PAM-02)
HMRC: Prompt users to check results and correct errors.
PAM's panel asks you to check what it tells you, and a return it prepares opens on the page where you check the figures.
How we make sure: components/assistant/assistant-widget.tsx, proposal-card.tsx
HMRC: Flag nuanced rules and complex scenarios, and recommend a qualified tax professional.
For residence, capital gains, partnerships, incorporation, disputes with HMRC and other matters of judgement, PAM says so and suggests an accountant or tax adviser.
How we make sure: agents/assistant more_tools.referral_line; the release gate's referral set must pass every case
HMRC: Remind users that they are responsible for the accuracy of their tax returns.
PAM's panel, every return it prepares and our terms say: making sure your returns are accurate is your responsibility.
How we make sure: lib/ai/disclaimer.ts; shared/assistant_actions.py (_FILING_NOTE)
Including strong data-security and privacy measures
HMRC: Follow Secure Software Development Lifecycle best practices.
Every change passes automated tests and security scans before release, and PAM has been penetration tested.
How we make sure: the release pipeline's quality gate; PAM's penetration tests (October 2026)
HMRC: Comply with UK GDPR.
Our privacy policy sets out the lawful basis for each use of your data, PAM included.
How we make sure: app/privacy/page.tsx
HMRC: Embed privacy by design.
PAM sees only what your role lets you see, and never bank account numbers, National Insurance numbers, dates of birth or home addresses.
How we make sure: infra-backend migration 272 (the books' views) and shared/assistant_grant.py
HMRC: Build in strong security measures.
Instructions hidden in receipts, invoices or web pages are screened out before PAM reads them, and its tools run as you, with your access.
How we make sure: agents/assistant screen.py, leak.py; the gateway's Cedar policies
HMRC: Users should be able to see how their information is processed.
This page and the privacy policy say it. You can see, correct or delete what PAM remembers about you.
How we make sure: components/assistant/pam-panels.tsx (What PAM remembers)
Using generative AI ethically
HMRC: Make sure all software outputs comply with legal requirements.
PAM refuses to help hide income from HMRC, and its answers are checked by a guardrail before you see them.
How we make sure: infra-backend modules/assistant/guardrail.tf; the release gate's referral set
HMRC: Use diverse and representative training data, handling any personal data in line with UK GDPR.
PactTax does not train or fine-tune a model. PAM uses Amazon Nova 2 Lite, trained by Amazon. Amazon Bedrock does not use your data to train models.
How we make sure: agents/assistant one_model.py and tests/test_one_model.py (Nova 2 Lite only, FR-PAM-04)
HMRC: Audit continuously to prevent harmful, biased or discriminatory outcomes.
A sample of PAM's conversations is scored for harm and correctness as they happen, the red team runs nightly, and each feature's accuracy is checked monthly.
How we make sure: AgentCore Evaluations with PactTax's own Nova 2 Lite judge (modules/assistant/online_eval.tf), evals/red_team.py, shared/ai_quality.py
HMRC: Be fair, trustworthy, fully accountable and serve the public good.
Every action PAM proposes is recorded with who decided it and when, and can be exported for review.
How we make sure: infra-backend migration 290 and shared/assistant_actions.review