What problem does it solve?
This job is about the documents a bank produces on request, not the scheduled bulk run of periodic statements. The manual effort sits in the long tail: a balance confirmation for a customer's auditor, an interest certificate for a tax return, a closure letter for an estate, a statement reprint for a visa application, a certificate in another language. These requests arrive by email, letter, branch and portal, and each needs data from several systems and a specific layout. Where the process is not automated, staff assemble each document by hand and rekey the figures.
Auditor confirmation requests need extra care. Under the PCAOB standard on confirmation (AS 2310), the auditor seeks evidence from a knowledgeable external source, such as the bank that holds the company's cash, and should consider asking that source about other financial relationships with the company, such as lines of credit, other indebtedness and guarantees. The standard puts the duty to keep control of the process on the auditor, so that requests and responses are not intercepted or altered.
For the bank this is a design consequence rather than a requirement of the standard: its confirmation flow should verify where a request comes from and answer only through channels the auditor controls. A wrong or incomplete answer weakens the audit evidence it was meant to provide.
How does it work?
- Take in the request. Requests from email, portals, branch systems and confirmation platforms are classified by document type, and the requester's authority is checked (customer consent for an auditor, executor documents for an estate).
- Gather the data. The workflow retrieves balances, interest, tax withheld, facilities, guarantees and other relationships from the systems of record for the requested date.
- Assemble from the template. Figures go into the approved template deterministically; the AI only drafts free text parts (for example a note on a closed account) and translations, which are checked.
- Validate against source. Automated checks reconcile every figure in the document with the source data, and document AI compares the output with what the request asked for.
- Approve, send and retain. Legal and regulated documents, such as audit confirmations, get a human check on a sample or in full; documents go out through secure channels and are retained with their data for the statutory period.
- Audience
- Back office
- Autonomy
- Supervised agent
- Adoption
- Emerging
- Channels
- Email, API and system to system, Internal tools
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 |
|---|---|---|---|---|
| Automation rate | Too few to pool | 62% | 1 | 1 vendor |
Value drivers: Lower cost to serve, Speed and cycle time, Risk and loss reduction, Customer experience.
Indicative value
A bank that produces 100,000 on request statements, certificates and confirmations a year
USD 200,000 to USD 1.9 million
Document assembly effort avoided per year
How this is calculated
Formula: documents * minutesPerDocument / 60 * automatedShare * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| On request documents per year documents, documents per year | 100,000 | 100,000 | The reference bank. Count only documents assembled by staff, not bulk statements. |
| Staff minutes per document today minutesPerDocument, minutes per document | 10 | 30 | Editorial assumption; audit confirmations take longer than statement reprints. Replace with your own time study. |
| Share of that effort removed automatedShare, fraction of effort | 0.4 | 0.7 | Editorial assumption. A vendor reports 62 percent of certificate applications at one bank handled solely by an AI agent (see the evidence records), but no public source measures effort removed per document. |
| Fully loaded operations cost per hour costPerHour, USD per hour | 30 | 55 | Editorial assumption, replace with your own. |
What it leaves out: Labour only. It leaves out the value of fewer errors in legal documents, faster turnaround for customers and auditors, and the cost of the platform and template work.
Who already uses it?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Federal Bank
India · Banking · 2026
Federal Bank, an Indian private sector bank, describes Feddy as its AI powered virtual assistant. On WhatsApp, customers can ask Feddy in chat for an account statement or a mini statement, and can generate and download a loan certificate, a deposit interest certificate, a balance confirmation certificate and a TDS certificate. The bank says these functions need no prior registration on the WhatsApp channel, and that Feddy recognises natural requests such as "I need my account statement" before authorization and authentication run and the document is returned in the chat. The bank's page gives no volumes, accuracy or time figures, and does not say how the documents are rendered.
No outcome disclosed.
Crédit Agricole Bank Polska
Poland · Banking · 2024
Crédit Agricole Bank Polska runs an AI agent, built by the integrator Deviniti, inside its after sales service system. The agent reads incoming letters and forms, classifies them, registers them and routes them to the right team, and carries out simple standard requests on its own. Certificate applications are one of the request types it handles. Deviniti reports that 62 percent of certificate applications are handled solely by the AI agent. The case study does not say which certificates these are (for example interest, tax or balance certificates), how the certificate itself is produced, or how many applications there are.
- Automation rate: 62%
"62% of certificate applications are handled solely by AI Agent"
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 templates per document type, language and jurisdiction
- Access to balances, interest, tax and relationship data by date
- Rules for who may request which document, with the evidence required
Systems to integrate
- Core banking, card, loan, deposit and custody systems
- Customer communications management platform
- Email, customer portal and audit confirmation platforms
- Document archive for retention
Complexity: Medium
Filling approved templates from structured data is a well understood pattern. Teams spend their time reaching every system of record for a given date, the authority checks, and keeping many templates and languages approved and current.
- 1
Inventory the long tail
List the on request document types, volumes, turnaround times and error rates. Pick the three with the most manual effort, often audit confirmations, interest certificates and balance letters.
- 2
Separate figures from words
Design each template so that every number comes from a system query and only clearly marked free text is drafted. This is the core safety property.
- 3
Automate intake and authority checks
Classify incoming requests and check authority and consent before any data is gathered, so documents are never produced for the wrong requester.
- 4
Validate every document against source
Add an automated reconciliation of each document's figures with the source data, and block sending when it fails.
- 5
Sample, then widen
Start with full human review, move to sampling for document types with a clean record, and keep full review for legal documents.
Guardrails
- Every figure comes from the system of record; the model never generates or edits numbers
- Authority and consent checks before any document is produced for a third party
- Automated reconciliation of each document with source data before sending
- Audit confirmations are sent only to verified auditor addresses or confirmation platforms
KPIs to instrument
- Turnaround time per document type
- Documents produced without manual intervention
- Errors found in sampling and reported by customers or auditors
- Requests rejected for failed authority checks
- Staff minutes per document
Human in the loop
Operations staff review exceptions, failed validations and a sample of documents. Legal and audit confirmation documents keep a human check, and the template owner approves every change to templates and wording.
Common failure modes
- A figure that does not match the books
- A document shows a balance for the wrong date or omits a facility. Reconcile every figure and check completeness of relationships.
- Documents sent to impostors
- A fraudster poses as an auditor or executor to obtain account information. Verify requester authority and use trusted confirmation channels.
- Template sprawl
- Local copies of templates drift from approved wording. Keep one managed template library with owners.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
Generating statements and confirmations from system data is not listed in Annex III and takes no decision about credit, insurance or access to essential services. If a customer facing agent takes the requests, Article 50(1) requires telling people they are interacting with an AI system. Under Article 50(2), providers of systems that generate text must mark the output as AI generated, unless the system only performs an assistive function for standard editing, or does not substantially alter the input data provided by the deployer or the semantics thereof; the free text the model drafts here is checked against source, but marking still applies where those exemptions do not.
Rules that apply
Guidance
- SYSC 9 Record keeping (Financial Conduct Authority, Europe). SYSC 9 requires firms to keep records of their services and transactions that let the FCA check compliance, including obligations to clients; common platform firms must retain records of their MiFID business for at least five years. Generated statements and confirmations, with the data behind them, belong in that record.
- AS 2310: The Auditor's Use of Confirmation (Public Company Accounting Oversight Board, North America). Applies to audits of fiscal years ending on or after 15 June 2025. Auditors must direct requests to knowledgeable confirming parties and keep control of the process so that requests and responses are not intercepted or altered, which sets what a bank's confirmation flow has to support.
Controls to put in place
- Template library under change control with owners and approval dates
- Reconciliation evidence stored with every generated document
- Access logging for every document produced for a third party
- Retention of each document and its data for the statutory period
Frequently asked questions
- Should a language model generate the figures on a statement or certificate?
No. Every figure must come from the system of record and be placed into an approved template deterministically. AI helps with request intake, authority checks, free text parts, translation and validating the finished document against source.
- Why are audit confirmations a good first candidate?
They ask about a known set of items (cash balances and, where the auditor chooses to ask, other financial relationships such as lines of credit, other indebtedness and guarantees) and draw on data the bank already holds, so much of the assembly can be automated. They also need strict control: PCAOB AS 2310 makes the auditor keep control of the process so that requests and responses are not intercepted or altered, and a bank flow that verifies requesters and answers through trusted channels supports that.
- Is any bank using AI on these documents in production?
Two named banks use AI on the request side. Federal Bank in India describes its Feddy assistant as AI powered and lists account statements, loan certificates, deposit interest certificates, balance confirmation certificates and TDS certificates among the documents customers can request in a WhatsApp chat. Deviniti, the integrator for Crédit Agricole Bank Polska, reports that 62 percent of certificate applications there are handled solely by an AI agent. Neither source says the model writes the figures, and we found no public report of AI producing bank audit confirmations. AI written spending reviews that explain a customer's year are a different job, covered on the financial wellbeing coach page.
How to cite this page
Blits.ai AI Use Case Library, "AI for statement, tax certificate and confirmation letter generation", last verified 13 October 2026, https://www.blits.ai/ai-use-cases/statement-and-tax-document-generation. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 13 October 2026: First published