What problem does it solve?
Banks, insurers and public bodies still receive a large share of their work as unstructured correspondence: scanned letters, emails with attachments, portal uploads and secure messages. Someone has to open each item, decide what it is (a complaint, a power of attorney, a bereavement notice, a change of address, a payment instruction), find the customer and send it to the right queue. That sorting step adds delay before anyone starts the real work.
Manual sorting is also where risk hides. A complaint filed as a general enquiry can miss its regulatory deadline, a bereavement letter can sit in the wrong queue, and a fraud warning can be read days late. Rule based keyword routing helps with obvious cases but struggles with free text and mixed documents.
How does it work?
- One intake. Post is scanned; emails, uploads and secure messages land in the same queue with their attachments.
- Classify. The AI identifies the document or request type, the language and any urgency signal (complaint, vulnerability, fraud, legal deadline), with a confidence score.
- Extract and link. It extracts the key fields (names, account numbers, dates, amounts, reference numbers) and matches the item to the customer and account in the core systems.
- Route or trigger. It sets the priority and service level, routes the item to the right team, or starts the downstream workflow directly (for example an address change or a bereavement case), with a short summary for the receiving officer.
- Fall back to people. Low confidence items, unmatched customers and anything sensitive go to a human review queue, and every correction becomes training data.
- Audience
- Back office
- Autonomy
- Supervised agent
- Adoption
- Mainstream
- Channels
- Email, Internal tools, API and system to system
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 | 25,000 to 100,000 | 2 | 1 organization, 1 vendor |
| Accuracy | Too few to pool | 91% | 1 | 1 vendor |
| Automation rate | Too few to pool | 99% | 1 | 1 organization |
Value drivers: Lower cost to serve, Speed and cycle time, Compliance quality, Employee productivity.
Indicative value
A bank or insurer that receives 1 million inbound letters and emails a year
USD 500,000 to USD 2.7 million
Manual sorting effort avoided per year
How this is calculated
Formula: items * minutesPerItem / 60 * automatedShare * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Inbound items per year items, items per year | 1,000,000 | 1,000,000 | The reference organization. Replace with your own mailroom and mailbox volume. |
| Minutes to read, classify, index and route one item manually minutesPerItem, minutes per item | 2 | 4 | Editorial assumption, replace with your own time study. |
| Share of items routed without human touch automatedShare, fraction of items | 0.5 | 0.8 | Editorial assumption, replace with your own. None of the evidence on this page reports a share of correspondence routed with no human touch; measure it on your own labelled sample before relying on it. |
| Fully loaded cost per hour of intake staff costPerHour, USD per hour | 30 | 50 | Editorial assumption, replace with your own. |
What it leaves out: Sorting labour only. It leaves out the value of faster downstream handling, fewer missed complaint deadlines and the cost of scanning, the platform and integration.
Who already uses it?
6 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
U.S. Department of Veterans Affairs
United States · Government and public sector · 2020
The Veterans Benefits Administration runs Mail Automation Services, an intake platform for claims material and other submissions from veterans and their representatives, mostly arriving through the Centralized Mail Portal. It combines form recognition, OCR, handwriting recognition and natural language processing to extract an average of 95 fields from more than 1,500 form layouts and to establish the claim and send each submission to the right business line. The agency reports that it ingests 25,000 to 40,000 packets a day; it has been in operation since 2020.
- Interactions handled: at least 25,000, packets per day (25,000 to 40,000)
"The platform ingests and reviews 25k to 40k packets of information per day, primarily derived from the Centralized Mail Portal."
Claimed by: organization
Loadsure
United Kingdom · Insurance · 2024
Loadsure, a London based insurtech for freight insurance, automated the intake of claim documents such as bills of lading, invoices and shipping documents with Google Cloud Document AI. Each incoming document is first classified and then sent to an extractor built for its type, which feeds the claims verification process; Gemini was later used for a similar extraction workflow elsewhere in the business. The blog post, written by Google Cloud and Loadsure staff, says work that took 30 to 60 minutes per claim now happens in near real time.
No outcome disclosed.
Encova Insurance
United States · Insurance · 2023
Encova Insurance, a US mutual insurer, replaced traditional OCR with UiPath Document Understanding in its claims invoice process, so incoming invoices are read, their data extracted and passed to automated processing, with exceptions fixed by staff in UiPath Action Center. Its solution architect says that traditional OCR got 40% of documents through without issues and that the new process has a 99% success rate, documents processed through without issues, which the vendor headline separately calls 99% accuracy. The vendor also reports that manual data entry time in the wider policy intake automation programme was cut by over 99% over the year.
- Automation rate: 99%, document understanding success rate on claims invoices, processed through without issues
"With this new [UiPath] process, the success rate is 99%."
Claimed by: organization
The Master Trust Bank of Japan
Japan · Banking · 2023
The Master Trust Bank of Japan, a trust bank specialising in asset servicing, uses Rossum's intelligent document processing to read inbound Japanese financial documents such as trade instructions, dividend notices and tax returns, and pass the extracted data to its own robotic process automation. The deployment grew to more than 90 document types used by 10 to 15 departments and 100,000 documents a year. The vendor reports that manual workload fell by 75% and that full processing now takes under 10 minutes instead of up to 1.5 hours per document.
- Productivity gain: 75%, manual data capture workload
"The Rossum IDP platform has reduced the manual workload by 75% and improved the time spent verifying and validating documents."
Claimed by: vendor - Interactions handled: 100,000, documents per year
"To date, Rossum has scaled to handle 100,000 documents per year for MTBJ."
Claimed by: vendor
Travelers
United States · Insurance · 2023
Travelers receives millions of emails a year from agents and customers asking for policy service. With AWS it built a classifier that reads each email and its attachments (Amazon Textract turns PDF attachments into text) and assigns one of 13 service categories with a prompted foundation model (Anthropic's Claude) on Amazon Bedrock; the post says the classifier powers an automation system for these requests but does not confirm that it runs in production. Initial testing without prompt engineering gave 68% accuracy; prompt engineering, condensed categories and better instructions raised it to 91%. The ground truth set held over 4,000 labelled emails.
- Accuracy: 91%, evaluation accuracy after prompt engineering (testing, not production)
"After using a variety of techniques with Anthropic’s Claude v2, such as prompt engineering, condensing categories, adjusting document processing process, and improving instructions, accuracy increased to 91%."
Claimed by: vendor
Ecclesia Group
Germany · Insurance · 2022
Ecclesia Group, a German insurance broker, uses ABBYY to process incoming claims correspondence. The platform extracts key data such as case numbers and licence plates from scanned documents, matches each document to the right record in the customer database and routes it to the responsible claims manager, replacing manual sorting. No outcome figures could be verified.
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
- A labelled sample of historical correspondence per request type
- An agreed taxonomy of request types with owner teams and service levels
- Customer and account data reachable for matching
Systems to integrate
- Scanning and mailroom capture
- Shared mailboxes and secure messaging
- Core banking or policy administration for customer matching
- Case management and workflow tools for routing
- Complaint management system
Complexity: Medium
Classification and extraction are mature. The effort is in the taxonomy of request types, the customer matching against core data and the connections to every downstream queue.
- 1
Build the request taxonomy with the receiving teams
List the request types, their owner team, priority and service level. Keep it short at first; a catch all class routed to people is better than twenty rare classes.
- 2
Label a real sample
Label a few thousand recent items, including the messy ones (handwritten, multi topic, forwarded chains), and use them as the test set for every model change.
- 3
Start with classification and summaries only
Let the AI propose the class and a summary while people still route. Measure accuracy per class before switching routing on.
- 4
Route automatically per class above a confidence threshold
Switch on automatic routing class by class, with a review queue below the threshold and mandatory human review for complaints, vulnerability and legal documents.
- 5
Trigger downstream workflows
For simple requests, start the fulfilment workflow directly from the extracted fields instead of dropping a task in a queue.
Guardrails
- Complaints, vulnerability signals and fraud warnings are always flagged and never auto closed
- Items below the confidence threshold go to a human review queue
- Customer matching requires at least two strong identifiers before an item is linked
- Personal data in documents is masked in logs and model prompts
KPIs to instrument
- Classification accuracy per class on a weekly sample
- Share of items routed without human touch
- Time from receipt to arrival in the right queue
- Misrouted items reported by receiving teams
- Complaints identified at intake versus later
Human in the loop
People review low confidence items and every item flagged as a complaint, vulnerability or legal matter. A quality team samples automatically routed items weekly and feeds corrections back into the labelled set.
Common failure modes
- A complaint routed as an enquiry
- The regulatory clock runs while the item sits in the wrong queue. Treat complaint detection as its own high recall check.
- Linked to the wrong customer
- A document attached to the wrong account is a data breach. Require strong identifiers and route ambiguous matches to people.
- Taxonomy drift
- New products and campaigns create request types the model never saw. Review the catch all class monthly.
What are the risks and rules?
EU AI Act
Depends on design
It depends on where the system runs. Classifying and routing a bank's or insurer's correspondence is not a use listed in Annex III, so it is minimal risk: the AI literacy duty of Article 4 applies, and the Article 50 duty to tell people they are dealing with AI does not, because the system does not interact with the sender. Used by or for a public authority in a benefits process covered by Annex III point 5(a), the provider can treat it as not high risk only while it performs a narrow procedural or preparatory task under Article 6(3); the provider must then document that assessment before it goes live (Article 6(4)) and register the system in the EU database (Article 49(2)). If the system evaluates eligibility for benefits or profiles the people who write in, it is high risk, so those judgements stay with people.
Rules that apply
Guidance
- DISP 1.6 Complaints time limit rules (Financial Conduct Authority, Europe). The response deadlines (eight weeks for most complaints, 15 business days for payment services and electronic money complaints) run from the firm's receipt of the complaint, so intake must recognise complaints wherever they arrive.
Controls to put in place
- Misclassification rate tracked as a model health metric, per class
- Log of every classification, extracted field and routing decision
- Separate high recall check for complaints and vulnerability
- Retention and access controls on scanned documents in line with the records policy
Frequently asked questions
- How accurate is AI at classifying inbound correspondence?
- Good enough to route most items, not all. In a post written by AWS and Travelers staff, the Travelers policy service email classifier reached 68% accuracy in initial testing and 91% after prompt engineering and condensed categories; the ground truth set had over 4,000 labelled emails in 13 classes. These are test results, not production figures, which is why the design keeps a human review queue below a confidence threshold.
- Does this work for scanned paper as well as email?
- Yes, once the capture step has turned the scan into text. The US Department of Veterans Affairs reports that its Mail Automation Services platform combines OCR, handwriting recognition and language processing, ingests 25,000 to 40,000 packets a day, mostly from its Centralized Mail Portal, and sends each submission to the right business line.
- What should never be routed automatically?
- Complaints, signs of customer vulnerability, fraud warnings and legal documents such as court orders and powers of attorney should always be flagged for a person, even when the classifier is confident.
How to cite this page
Blits.ai AI Use Case Library, "AI for inbound correspondence triage and routing", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/correspondence-triage-and-routing. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published