What problem does it solve?
Banks, insurers and public bodies send large volumes of letters. The high volume ones come from fixed templates, but a long tail does not fit a template cleanly: a complaint response that has to address the customer's specific points, an arrears letter that must reflect an agreed payment plan, a decline letter with the right reasons, a notice in the customer's own language. Staff write these by hand, copying clauses from a library and data from several systems, which is slow and produces uneven quality.
The quality matters because many of these letters are regulated. In the UK, a final response to a complaint must meet content and timing rules, in the US an adverse action notice on a credit application must state the specific reasons, or tell the applicant they can get them within 30 days, and the FCA Consumer Duty expects communications that customers are likely to understand. A free writing model is not acceptable here; the safe gain comes from drafting inside approved wording.
How does it work?
- Start from the case. The trigger is a case event (complaint investigated, arrears stage reached, application declined, product changed) with its structured data.
- Select the approved template. Rules, not the model, choose the template and the mandatory clauses for the notice type and jurisdiction.
- Draft the variable parts. The AI writes the case specific paragraphs (the summary of the complaint and findings, the payment plan terms, the plain language explanation) using only facts from the case and wording retrieved from the approved clause library.
- Check before a human sees it. Automated checks compare every figure and date in the draft with the case data, confirm mandatory clauses are present, and score readability.
- Approve, send and store. A person approves regulated notice types; approved letters are sent through the customer's preferred channel and stored with their version and data.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Early adopters
- Channels
- Email, 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 |
|---|---|---|---|---|
| Hours saved | Not pooled | 276 hours to 11,000 hours | 2 | 1 organization, 1 vendor |
| Cycle time reduction | Too few to pool | 25% | 1 | 1 vendor |
| Handling time reduction | Too few to pool | about 50% | 1 | 1 vendor |
Value drivers: Employee productivity, Compliance quality, Customer experience, Speed and cycle time.
Indicative value
A bank or insurer whose staff write 150,000 non standard letters a year
USD 262,500 to USD 2.3 million
Drafting effort avoided per year
How this is calculated
Formula: letters * minutesSaved / 60 * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Letters written or heavily edited by hand per year letters, letters per year | 150,000 | 150,000 | The reference organization. Count only letters that do not go out from a fixed template. |
| Minutes saved per letter minutesSaved, minutes per letter | 3 | 15 | The low end is Acentra Health's reported fall from about six to three minutes per appeal letter; the high end is an editorial assumption for complaint responses written from scratch. Replace with a time study. Source |
| Fully loaded cost per hour of the writing staff costPerHour, USD per hour | 35 | 60 | Editorial assumption, replace with your own. |
What it leaves out: Drafting time only. It leaves out the review time that remains, the value of fewer complaint escalations and ombudsman referrals, and the cost of the platform and template work.
Who already uses it?
5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Health Resources and Services Administration
United States · Government and public sector · 2024
HRSA's auditors faced a sharp rise in Single Audits linked to COVID era Provider Relief Fund payments. Its AI Audit Resolution Assistant (AIARA) puts the Single Audit documents assigned to HRSA in a vector database and uses retrieval augmented generation with a large language model to summarise findings and recommendations, answer auditors' questions and draft the Management Decision Letters that close findings out, while robotic process automation pulls data from the Federal Audit Clearinghouse into letter templates. HRSA reports in the 2025 federal inventory that the pilot, operational since July 2024, has processed and resolved 73 audits and saved an estimated 276 hours.
- Interactions handled: 73, since launch in July 2024, as reported in the 2025 inventory
"Since its launch, the AIARA has successfully processed and resolved 73 audits."
Claimed by: organization - Hours saved: about 276 hours, since launch in July 2024, as reported in the 2025 inventory
"Automation has resulted in an estimated total of 276 hours of work saved."
Claimed by: organization
SS&C GIDS and RS
United Kingdom · Wealth and asset management · 2026
SS&C Global Investor and Distribution Solutions and Retirement Solutions, which runs customer service for asset managers, insurers and wealth managers under FCA rules, rebuilt its complaints process with AI agents. After a human investigator records the findings, the agents use the case notes to draft the closing letter that summarises them; an employee checks the letter before the agents send it. SS&C Blue Prism, a business in the same SS&C group, reports that complaint cycle times fell by 25%.
- Cycle time reduction: 25%, complaint cycle time, whole process
"Cycle times have been reduced by 25%, so customers get a follow-up letter more quickly."
Claimed by: vendor
Hiscox
United Kingdom · Insurance · 2025
Hiscox is rolling out Microsoft 365 Copilot to its more than 3,000 employees after a trial. A senior technical claims underwriter in the UK claims team uses it to identify and record the key information of a new claim, to summarise long expert medical evidence and legal advice, and to pull the progress of a claim from several emails and compose an update to a broker or customer. He says that recording a new claim now takes him as little as 10 minutes instead of up to an hour. This is one user's experience, not a measured program result.
No outcome disclosed.
Acentra Health
United States · Healthcare · 2024
Acentra Health, which reviews Medicare appeals, built MedScribe on Azure OpenAI Service to turn a physician's clinical rationale into a plain language, empathetic appeal determination letter for the beneficiary and the provider, a task specially trained nurses used to do by hand in an appeals process where a decision can be due within 24 hours. It was tested with 10 nurses who rated every draft, then rolled out to all nurses who write these letters. Microsoft reports that time per letter fell by about 50%, from six to three minutes, saving 11,000 nursing hours and nearly USD 800,000 since deployment, and that nurses gave the generated letters a 99% approval rating.
- Handling time reduction: about 50%, Nurse time per appeal determination letter
"With MedScribe, Acentra Health reduced the time that its specially trained nursing staff spent on each appeal determination letter by approximately 50%."
Claimed by: vendor - Hours saved: 11,000 hours, Nursing hours saved by mid 2024
"By mid-2024, the company had saved 11,000 nursing hours and begun preparations to expand MedScribe to other areas to drive organizational efficiency."
Claimed by: vendor - Cost savings: about USD 800,000, Saved since deployment
"This adds up to 11,000 nursing hours and nearly $800,000 that have been saved since deploying MedScribe."
Claimed by: vendor
SS&C GIDS
Global · Wealth and asset management · 2024
SS&C GIDS handles customer communications for asset managers and other financial institutions, such as instructions on selling assets and responses to changes of address, broker or customer ID. After an employee investigates a request and records comments in a template, a digital worker validates the case and prompts an in house large language model, which generates a personalised letter; quality control reviews and adjusts it before it is sent. SS&C Blue Prism, a business in the same group, reports that these communications are now produced three times faster than with the manual process.
No outcome disclosed.
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 and clause library per notice type, language and jurisdiction
- Structured case data for each trigger (complaint findings, arrears status, decision reasons)
- A sample of good historical letters per type as a quality reference
Systems to integrate
- Case management and complaint systems
- Collections, lending and servicing systems for case data
- Customer communications management platform for layout and dispatch
- Document archive for versioned storage
Complexity: Medium
The model work is modest. Most effort goes into a clean, approved clause library, the mapping of case data into drafts, and agreeing with compliance which notice types need human approval.
- 1
Pick the notice types that are hand written today
Fully templated letters need no AI. Start where staff write free text, such as complaint responses and bespoke arrears letters.
- 2
Clean up the clause library
Give every clause an owner, a version and a legal approval date, and remove duplicates. The model can only be as safe as the wording it is allowed to use.
- 3
Draft next to people first
Let staff start from the AI draft and measure edit distance and time per letter by type before changing any approval rule.
- 4
Automate the checks
Build deterministic checks for figures, dates, names and mandatory clauses so reviewers focus on tone and judgment instead of proofreading.
- 5
Decide approval by notice type
Keep human approval on regulated and adverse notices, and consider sampling instead of full review only for low risk confirmations with a stable quality record.
Guardrails
- Template and mandatory clauses are chosen by rules, never by the model
- Every figure, date and name in the draft is checked against the case data before review
- Regulated and adverse notices always have a named human approver
- Every sent letter is stored with its version, template and source data
KPIs to instrument
- Minutes per letter from draft to approval, by notice type
- Share of drafts approved without material edits
- Factual errors caught by the automated checks and by reviewers
- Readability score of sent letters
- Complaints about letters and repeat contacts after a notice
Human in the loop
Case handlers approve every regulated or adverse notice and edit drafts where needed. Compliance owns the clause library and samples approved letters monthly, including those in other languages.
Common failure modes
- A plausible but wrong fact
- The draft states an amount or date that is not in the case. Block sending until the automated fact check passes.
- Reviewers stop reading
- After weeks of good drafts, approval turns into a click. Track review time and seed test drafts with known errors.
- Tone that fails vulnerable customers
- A correct letter that is cold or confusing. Include vulnerability flags in the case data and test drafts with plain language checks.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
Drafting letters for human approval is not listed in Annex III. The decision the letter communicates may come from a separate high risk system, such as credit scoring (Annex III point 5(b)) or a public body's eligibility decision on benefits (point 5(a)); the drafting tool does not make that decision. Article 50(2) requires the provider of an AI system that generates text to mark the output as artificially generated, which puts this on the limited risk (transparency) tier; this includes an organization that builds its own drafting tool. Article 50(2) does not apply where the AI has only an assistive function for standard editing and does not substantially alter the input data or the semantics of the output.
Guidance
- PRIN 2A.5 Consumer Duty, consumer understanding outcome (Financial Conduct Authority, Europe). Communications must be likely to be understood by the customers they are aimed at, which applies to AI drafted letters too.
- DISP 1.6 Complaints time limit rules (Financial Conduct Authority, Europe). Sets the deadline for a final response (eight weeks for most complaints, 15 business days for payment services and e money complaints) and what it must contain, including the Financial Ombudsman Service referral rights. A drafted complaint response must meet these rules.
- Regulation B, 12 CFR 1002.9 Notifications (Consumer Financial Protection Bureau, North America). Adverse action notices on credit applications must state the specific reasons, or tell the applicant they can get them within 30 days; a drafting tool must use the reasons the credit decision produced.
- Article 50, transparency obligations for providers and deployers of certain AI systems (European Union, Europe). Article 50(2) requires providers of systems that generate text to mark the output as artificially generated.
Controls to put in place
- Clause library under change control with legal approval dates
- Automated fact and clause checks logged for every draft
- Named approver recorded for every regulated notice
- Versioned storage of every sent letter for the statutory retention period
Frequently asked questions
- Is it safe to let generative AI write regulated customer letters?
- Only inside approved wording and with a human approver for regulated and adverse notices. The template and mandatory clauses are chosen by rules, the model drafts only the case specific paragraphs from case data, and automated checks compare every figure and date with the case before a person reviews it.
- Where does AI drafting save the most time?
- In the letters staff write by hand today, such as complaint outcome letters and bespoke customer letters. SS&C Blue Prism reports that SS&C GIDS produces customer letters with an in house language model three times faster than with the manual process, and that complaint cycle times fell by 25% after AI agents took over steps including drafting the closing letter for an employee to check.
- Does this work for insurance claims letters?
- Yes, claim updates and decision letters follow the same pattern. Microsoft reports that a Hiscox claims underwriter uses Microsoft 365 Copilot to pull the progress of a claim from several emails and compose an update to a broker or customer. In Medicare appeals, Microsoft reports that Acentra Health cut nurse time per appeal determination letter by approximately 50% with its MedScribe drafting tool.
- Does the EU AI Act make letter drafting high risk?
- No, it is limited risk. Drafting letters is not listed in Annex III. The decision the letter communicates, such as a credit decline, may come from a high risk system, which is governed separately. The provider of the drafting system must mark generated text under Article 50(2), unless the AI only performs an assistive function for standard editing.
How to cite this page
Blits.ai AI Use Case Library, "AI for drafting customer letters and outbound notices", last verified 26 September 2026, https://www.blits.ai/ai-use-cases/outbound-notice-drafting. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published