What problem does it solve?
Written service channels are slow in a way phone is not. An email or ticket waits in a queue, an agent opens it, reads the history, looks up the account, searches the knowledge base, writes a reply and often has to ask a colleague. Much of that effort is repeated for questions that have been answered many times before, and the reply quality depends on who happens to pick the ticket up. Backlogs build after product changes and incidents, and response time targets slip.
Fully automated replies are tempting but risky for anything beyond simple, well understood requests: a confident wrong answer in writing is a record the customer can forward. A common middle ground is a drafted reply that the agent owns. The AI does the reading, the lookup and the first draft; the agent brings judgment, corrects and sends. Real time assist during calls and chats is covered on its own page; this page is about the asynchronous queue.
How does it work?
- Read and summarize. When a message or ticket arrives, the AI summarizes the request and the thread, detects the intent, language and urgency, and labels it for routing.
- Gather context. It retrieves the customer's relevant data (orders, account status, open cases) through approved system calls and the relevant passages from the approved knowledge base.
- Draft the reply. It writes a response in the house tone and the customer's language that addresses every point raised, cites the knowledge it used and leaves placeholders where it lacks information.
- Agent reviews and sends. The agent edits, completes and sends. Drafts for regulated topics, complaints or vulnerable customers are marked for extra care.
- Automate only the safe tail. For a narrow set of simple, low risk intents, the organization may send replies automatically, with disclosure and sampling.
- Learn from edits. The difference between draft and sent reply is logged to improve prompts and to find missing knowledge.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Mainstream
- Channels
- Email, Agent desktop, Social messaging, 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 |
|---|---|---|---|---|
| Cycle time reduction | Too few to pool | 33% to 50% | 2 | 2 vendor |
Value drivers: Employee productivity, Lower cost to serve, Customer experience, Speed and cycle time.
Indicative value
A service team that answers 400,000 emails and tickets a year
USD 288,000 to USD 1.2 million
Agent handling cost released per year
How this is calculated
Formula: tickets * minutesPerTicket * reduction * costPerMinute. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Emails and tickets answered per year tickets, tickets per year | 400,000 | 400,000 | The reference organization. |
| Agent handling time per ticket today minutesPerTicket, minutes per ticket | 6 | 10 | Editorial assumption for written service. Replace with your own handling time. |
| Reduction in handling time with drafted replies reduction, fraction of handling time | 0.2 | 0.33 | Editorial assumption, replace with your own. No source on this page measures agent handling time per email or ticket: Google Cloud reports Turing's HR ticket processing time down by 33%, without defining what that processing time covers, and Microsoft reports HYPE's agents resolving WhatsApp chat conversations in half the time. The range is set at or below both. |
| Fully loaded agent cost per minute costPerMinute, USD per minute | 0.6 | 0.9 | Editorial assumption, replace with your own. |
What it leaves out: Counts only agent time. It leaves out faster responses and more consistent quality, the cost of the platform, and any effect of automating simple replies end to end. Savings usually show as backlog cleared and capacity redeployed rather than immediately lower cost.
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.
Centers for Disease Control and Prevention
United States · Government and public sector · 2024
CDC's National Center for Immunization and Respiratory Diseases runs SmartFind, an internal knowledge bot with a SharePoint component that helps program staff manage partner emails and lets mailbox managers use a shared knowledge base. The agency lists partner mailbox email management and knowledge base maintenance as the purpose. The bot matches free text questions to agency cleared answers and flags complex or unanswerable queries for manual review. Earlier public facing versions gave agency cleared answers to public and partner questions during the COVID-19 pandemic. No outcome figures are published.
No outcome disclosed.
Transportation Security Administration
United States · Government and public sector · 2024
AskTSA answers traveller questions in writing through X, Facebook Messenger, Apple Messages and text message. Since December 2024 TSA has used generative AI in the AskTSA customer service process to summarize why a traveller is getting in touch, recommend replies tailored to that summary, categorize incoming inquiries and report on where the virtual assistant falls short. The recommended replies assist the human agents; the entry does not describe automatic sending. The department classifies it as not high impact; no outcome figures are published.
No outcome disclosed.
HYPE
Italy · Banking · 2025
Italian neobank HYPE handled nearly 45% of its customer contacts by email and 40% by chat and WhatsApp when it moved to a new CRM. It built an agent that reads customer emails, segments them by subject and triggers an automatic reply, plus a voice agent that handles calls about common issues. Its human agents use Copilot in Dynamics 365 Customer Service for case and conversation summaries and email assistance. Microsoft reports that agents resolve WhatsApp conversations in half the time and that the custom agents cut human customer service intervention by 70% over a year.
- Cycle time reduction: about 50%, WhatsApp chat conversations resolved by human agents with Copilot summaries and email assistance (chat, not email or tickets)
"Copilot in Customer Service provides automated, AI-powered case and conversation summaries and email assistance, helping human agents resolve WhatsApp conversations in half the time."
Claimed by: vendor
Turing
United States · Technology and software · 2025
Turing (listed by Google Cloud as Turing Enterprises), an AI company headquartered in San Francisco, built a custom AI model trained on its internal knowledge that drafts replies to HR support tickets. Google Cloud reports a one third cut in ticket processing time after two days of development. A plan to automate 60% of its 52,000 annual HR tickets with Gemini Gems is a target, not a result.
- Cycle time reduction: 33%
"Turing also built a custom AI model trained on internal knowledge to draft replies to HR support tickets, reducing ticket processing time by 33% after two days of development."
Claimed by: vendor
Nomad eSIM
United States · Telecommunications · 2024
Nomad eSIM, a LotusFlare brand used by international travellers, gives its customer support agents Gemini in Google Workspace to respond to trouble tickets more efficiently. Google Cloud reports higher customer satisfaction from faster support responses, but gives no figure.
No outcome disclosed.
First National Bank
South Africa · Banking · 2023
First National Bank, a division of FirstRand, rolled out Microsoft Copilot for Sales to its bankers in December 2023, on top of a Dynamics 365 CRM that nearly all of its more than 3,500 bankers already used. The customer story focuses on drafting replies to commercial clients in Outlook that address each point in the client's message, which the banker reviews, edits and sends. Because bankers worried the drafts would sound robotic, the bank paired the rollout with a training programme with modules for each type of task.
- Employee adoption: at least 94%, commercial bankers
"More than 94% of commercial bankers now use Copilot to help them craft richer communications with their customers."
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, current knowledge articles and reply templates with owners
- A contact reason taxonomy with volumes for email and ticket channels
- Historical tickets with the replies that resolved them, for testing
- Customer and order data reachable through APIs
Systems to integrate
- Ticketing or case management platform (for example Zendesk, Salesforce, Freshdesk, ServiceNow)
- Email and messaging channels
- CRM, order and account systems for context
- Knowledge base
- Quality assurance tooling for sampling sent replies
Complexity: Medium
Drafting from a knowledge base is quick to prototype. The work is in reading customer data through proper integrations, keeping knowledge current, handling many languages and designing which topics may never be answered without extra review.
- 1
Pick intents by volume and risk
From the contact reason report, choose high volume topics where the answer is in approved knowledge. Mark complaints, legal and regulated topics for extra review from day one.
- 2
Ground every draft
Draft only from approved knowledge and data from system calls, show the agent the sources, and leave explicit gaps instead of guessing when information is missing.
- 3
Put the draft where agents work
Embed summaries and drafts in the existing ticket view. A separate tool that needs copy and paste loses most of the gain.
- 4
Measure edits, not only speed
Track how much agents change each draft and why. Heavy edits point at missing knowledge or a bad prompt; no edits on complex topics may point at over reliance.
- 5
Test in every language you serve
Build a test set of real tickets per language and intent, and run it on every prompt, model or knowledge change.
- 6
Automate the safe tail last
Only after months of low edit rates on an intent consider sending automatically, with AI disclosure, sampling and an easy route to a person.
Guardrails
- The agent sends; no reply leaves without human action unless the intent is on an approved automation list
- Drafts use only approved knowledge and data from authorized system calls, with sources shown
- Complaints, legal threats and signs of vulnerability are flagged and routed, not just answered
- Personal and payment data masked in prompts and logs
- Regulated statements (fees, rights, deadlines) come from approved templates, not free generation
KPIs to instrument
- Handling time per ticket by intent, before and after, on the same case mix
- Draft acceptance rate and edit distance per intent
- Reopen and repeat contact rate within seven days
- Quality assurance score of sent replies
- Time to first response and backlog age
Human in the loop
Agents review and own every reply they send. Team leads sample sent replies weekly for accuracy and tone, knowledge owners fix the gaps that heavy edits reveal, and any move to automatic sending for an intent needs sign off from the service owner and compliance.
Common failure modes
- Rubber stamping
- Agents send drafts unread under time pressure. Sample sent replies and watch for near zero edit rates on complex topics.
- Fluent but wrong
- A plausible reply built on outdated knowledge. Show sources, refuse when retrieval finds nothing and keep knowledge owned and current.
- Missed complaint
- A complaint is answered as a routine question and never logged. Detect complaint language and route it to the complaints process.
- Data from the wrong customer
- Context pulled for a similar name or a shared email address. Match on verified identifiers only and show the agent what was used.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
A drafting copilot whose output an agent reviews and sends falls under the transparency tier at most. When replies are sent without human review, customers interact with the AI system directly, and Article 50(1) requires that they are informed unless this is obvious from the context. Article 50(2) separately requires the provider of a system that generates text to mark its output in a machine readable format as artificially generated, whether or not a person reviews the draft. It becomes high risk only if it is used for a purpose listed in Annex III, such as evaluating eligibility for public benefits or creditworthiness (point 5), or evaluating the performance of the agents who use it (point 4).
Rules that apply
Guidance
- Regulation (EU) 2024/1689 (AI Act), Article 50, transparency obligations for providers and deployers of certain AI systems (European Union, Europe). Official text on EUR-Lex. Article 50(1) is relevant when replies are sent automatically, so the customer interacts with the AI system directly; Article 50(2) covers machine readable marking of generated text by the provider.
Controls to put in place
- Documented list of intents eligible for drafting and, separately, for automatic sending
- Logging of draft, edits and sent version for quality review and disputes
- Weekly quality sampling with feedback into knowledge and prompts
- Personal data masking and retention limits on drafts and logs
- Complaint detection linked to the complaints handling process
- For UK retail financial services firms, drafted replies checked against the consumer understanding outcome of the FCA Consumer Duty (PRIN 2A.5) before they are sent
Frequently asked questions
- How much time does AI reply drafting save?
- The figures on this page come from vendors and measure different things. Google Cloud reports that Turing cut its HR ticket processing time by 33% with a model that drafts replies, and Microsoft reports that HYPE's human agents resolve WhatsApp conversations in half the time with Copilot case and conversation summaries and email assistance. It is unclear whether either figure measures agent handling time per email or ticket, so measure your own baseline before relying on either.
- Should AI replies be sent automatically?
- Only for a narrow set of simple, low risk intents after a period of drafts that agents barely change, and with AI disclosure. Everything else is better drafted by the AI and sent by an agent. TSA's AskTSA service, for example, uses AI to summarize inquiries and recommend replies to the human agents who answer them.
- How is this different from live agent assist?
- Live agent assist works in real time during a call or chat. Reply drafting works on the asynchronous queue of emails, tickets and messages, where the agent has time to review and the draft is the main output.
How to cite this page
Blits.ai AI Use Case Library, "AI reply drafting for customer email and support tickets", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/email-and-ticket-reply-drafting. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published