What problem does it solve?
Tax authorities face sharply seasonal demand. Around filing deadlines and after every policy change, phone lines and webchat fill with the same questions: where is my refund, how do I get a reference number, can I pay in instalments, what does this notice mean. Long waits push people to give up, file late or file wrongly, which creates more work later in compliance and correspondence.
The questions are repetitive but the stakes are not trivial: a wrong answer about a deadline or a relief can cost the taxpayer money. That is why the tax authority assistants on this page (HMRC's digital assistant and the IRS voice bots and chatbots) classify intent and return approved content, and add authenticated actions such as payment plans only behind identity checks.
How does it work?
- Recognise the intent. The assistant classifies the question (refund status, payment plan, notice, registration, how to file) on chat or on the phone.
- Answer from approved content. General questions get answers written or approved by the tax authority, with links to the guidance; unclear questions get a choice of likely meanings.
- Authenticate for account questions. For refund status, balance or a payment plan, the taxpayer verifies identity (shared secrets, PIN or national login) before the assistant reads the account.
- Act within rules. Within fixed limits the assistant can set up or change a payment plan or grant a payment extension, and confirms the result.
- Escalate. Complex, disputed or personal situations go to a human adviser, who sees the whole conversation and completes identity checks.
- Audience
- Customer facing
- Autonomy
- Supervised agent
- Adoption
- Mainstream
- Channels
- Web chat, Phone and voice, Mobile app, WhatsApp
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Interactions handled | Not pooled | 3 million to 5.5 million | 2 | 2 organization |
| Accuracy | Too few to pool | 83% | 1 | 1 organization |
| Contact deflection | Too few to pool | 20% | 1 | 1 organization |
| Users served | Not pooled | at least 200,000 | 1 | 1 vendor |
Value drivers: Lower cost to serve, Customer experience, Compliance quality, Inclusion and access.
Indicative value
A national tax authority with 3 million assisted contacts a year
USD 1.2 million to USD 7.2 million
Human handled contact cost avoided per year
How this is calculated
Formula: contacts * routineShare * containment * costPerContact. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Assisted phone and chat contacts per year contacts, contacts per year | 3,000,000 | 3,000,000 | The reference authority. |
| Share of contacts on routine topics (refund status, payments, forms) routineShare, fraction of contacts | 0.4 | 0.6 | Editorial assumption. Replace with your own contact reason data. |
| Share of routine contacts the assistant resolves containment, fraction of routine contacts | 0.2 | 0.4 | Editorial assumption. For context, HMRC reports that webchat escalations to advisers fell 20% while assistant interactions grew 18.8%. |
| Cost of a human handled contact costPerContact, USD per contact | 5 | 10 | Editorial assumption. Replace with your own fully loaded cost. |
What it leaves out: Gross contact cost only. It leaves out the effect on filing accuracy and late payment, the value of shorter queues at peak and the cost of building and running the assistant.
Who already uses it?
4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
HM Revenue and Customs
United Kingdom · Government and public sector · 2025
HMRC's digital assistant answers tax questions typed in plain language, matching them to intents with natural language understanding and replying with non personalised answers that link to GOV.UK guidance. It covers 60 of HMRC's more than 120 taxes, asks the user to choose between likely meanings when unsure, and needs no login. Complex questions escalate to webchat with a human adviser, who sees the whole prior conversation and only continues after identity and verification checks.
- Interactions handled: at least 5.5 million, tax year 2024/25 to 6 March 2025
"The digital assistant has had 5.48m interactions in the tax year 2024/25 (to date as at 6 March 2025)."
Claimed by: organization - Accuracy: 83%, NLU test set, March 2025, known intents
"83.03% on known intents"
Claimed by: organization - Contact deflection: 20%, webchat escalations to an adviser, tax year 2024/25 to 6 March 2025
"For webchat, 909,000 users have escalated to speak to an adviser in the tax year 2024/25 (to date as at 6 March 2025). That is a 20% decrease on the 2023/24 tax year."
Claimed by: organization
Internal Revenue Service
United States · Government and public sector · 2022
Since 2021 the IRS has put intent based voice bots on many toll free lines: payment plans and balance due (with authentication so taxpayers can set up or change a payment plan), Where's My Refund and amended return status, notice clarifications, Economic Impact Payments and the Advance Child Tax Credit. On IRS.gov, chatbots answer FAQs on refunds, identity theft, payments and relief. The inventory stresses that answers are not generated: the model classifies the question and returns content approved by the business owner, or routes the call to a live assistor.
- Interactions handled: at least 3 million, calls answered by voice bots, cumulative to June 2022
"To date, the voice bots have answered over 3 million calls."
Claimed by: organization
Internal Revenue Service
United States · Government and public sector · 2020
The IRS Linguistic Policy, Tools and Services team uses a cloud machine translation application on AWS, with the IRS Publication 850 glossary of English and Spanish tax terms, to translate text and files between English and Spanish, Chinese, Korean and Vietnamese and speed up responses to taxpayers. Separately, staff use SYSTRAN neural translation, augmented with a domain dictionary and translation memories, to triage non English documents for relevance to case work and as a starting point for manual translation. Both appear in the federal AI inventory as in operation.
No outcome disclosed.
ClearTax
India · Technology and software · 2024
ClearTax, an Indian online tax filing platform, built a generative AI agent on WhatsApp, using Azure OpenAI models, for low income and blue collar workers who have tax deducted at source from their income but rarely file a return, and so miss refunds and lack the income proof lenders ask for. The agent explains taxes and filing, walks users through the filing workflow and supports nine Indian vernacular languages. It is a private filing service, not a tax authority channel.
- Users served: at least 200,000, workers who filed their returns independently
"200,000+ blue-collar workers successfully filed ITRs independently"
Claimed by: vendor - Customer savings: about INR 300 million, tax refunds claimed by users, cumulative
"300 million INR unlocked in tax refunds, improving access to credit opportunities"
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Approved answers and guidance per tax and topic, with owners and effective dates
- Contact reason data by season, to pick intents and plan for peaks
- Rules for payment plans and extensions that can be applied without judgment
Systems to integrate
- Taxpayer account, refund and payment systems through APIs
- Identity verification and step up authentication
- Telephony IVR and webchat platforms with handover to advisers
- Notice and correspondence systems, to explain letters by reference
Complexity: Medium
Information answers are low complexity; authenticated actions such as payment plans need integration with taxpayer account systems, strong identity checks and strict rules on what the assistant may change.
- 1
Start where the queues are
Use contact reason data to pick the few intents that dominate peak season (refund status, payments, notices), as the IRS did when it put voice bots on its Economic Impact Payment, notice and payment lines.
- 2
Keep answers approved
Let the model classify and retrieve, and serve content the business owner approved. The IRS inventory stresses that its bots do not generate answers.
- 3
Add authenticated self service
Add account questions and payment plans behind identity checks, with limits (amount, term) written as rules, and confirm every change back to the taxpayer.
- 4
Design escalation with context
Pass the conversation to the adviser, as HMRC's webchat advisers see the assistant history, and complete identity checks before personal discussion.
- 5
Prepare for peaks and changes
Retest before each filing season and after each budget, and plan capacity for the spike.
Guardrails
- No assessment, penalty or dispute outcome is decided by the assistant
- Account data only after identity verification, at the level the action needs
- Actions such as payment plans only within written limits, with confirmation to the taxpayer
- Answers from approved content with effective dates; refusal when the topic is out of scope
- Personal data and tax identifiers masked in logs and model prompts
KPIs to instrument
- Containment per intent, counting repeat contacts within seven days as not contained
- Escalations to advisers, before and after launch
- Intent recognition accuracy on a labelled test set
- Payment plans set up through the assistant and their default rate
- Satisfaction on assistant and adviser conversations
Human in the loop
Advisers handle escalations, disputes, hardship and anything outside the rules. Content owners approve every answer and every change after a budget; a team samples conversations weekly and reviews failed intents and complaints.
Common failure modes
- Wrong deadline or relief answers
- A fluent but wrong answer costs the taxpayer money and trust. Serve approved content, show effective dates and refuse outside scope.
- Peak season collapse
- The assistant is tested in quiet months and fails at the deadline. Load test and retest before each season.
- Authentication friction
- Taxpayers fail identity checks and fall back to the phone. Offer several proportionate methods and measure drop off.
- Private helpers without oversight
- Third party filing assistants (such as ClearTax's WhatsApp agent) help people file but are not the authority; make official guidance easy for them to use and keep the authority's own channel authoritative.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
A taxpayer assistant must tell people they are interacting with an AI system (Article 50). It is not listed in Annex III as long as it only informs and applies fixed rules. It becomes high risk under Annex III point 5(a) if it evaluates eligibility for, or grants, reduces, revokes or reclaims, public assistance benefits (which can include benefits paid through the tax system). Recital 59 says systems used for administrative proceedings by tax and customs authorities are not high risk law enforcement systems; audit selection and risk scoring are covered on a separate page.
Rules that apply
Guidance
- Article 50, transparency obligations for providers and deployers of certain AI systems (European Union, Europe). Providers must design assistants so that taxpayers are informed they are interacting with an AI system, unless that is obvious from the context.
- Recital 59, AI systems used by tax and customs authorities (European Union, Europe). Systems for administrative proceedings by tax and customs authorities should not be classified as high risk law enforcement systems.
- Algorithmic Transparency Recording Standard Hub (Government Digital Service, Europe). HMRC publishes its assistant under this standard, including accuracy on known intents.
- M-25-21: Accelerating Federal Use of AI through Innovation, Governance, and Public Trust (US Office of Management and Budget, North America). Sets the rules for US federal AI use and requires agencies to inventory their AI use cases at least annually and publish the inventory, where the IRS lists its voice bots and chatbots.
Controls to put in place
- AI disclosure and a route to a human adviser on every channel
- Entry in the public AI inventory or transparency register
- Written limits for every action the assistant can take on an account
- Change control tied to the budget and filing season calendar
- Retention limits and access control on conversation logs containing tax data
Frequently asked questions
- How much volume do tax assistants handle?
- HMRC's digital assistant had 5.48 million interactions in the 2024/25 tax year up to 6 March 2025, and IRS voice bots had answered over 3 million calls by June 2022, about a year after the first one went live in May 2021.
- Do tax authorities use generative AI for these answers?
- Not in the deployments on this page. The IRS inventory states its chatbots and voice bots return content predetermined by content owners, and HMRC matches intents to approved answers. The IRS tested a generative AI chatbot for volunteer tax preparers as a proof of concept (listed as retired in 2025), while the private filing service ClearTax runs a WhatsApp agent on Azure OpenAI models.
- How accurate is intent recognition?
- HMRC reports 83.03% accuracy on known intents in its March 2025 test set. Plan for the rest with disambiguation questions and an easy route to a human.
How to cite this page
Blits.ai AI Use Case Library, "AI assistant for tax questions and filing support", last verified 26 September 2026, https://www.blits.ai/ai-use-cases/tax-questions-and-filing-assistant. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published