Saved replies and AI reply drafts at a glance
A saved reply (support tools also say canned response or macro) is an answer written in advance that the agent drops into the conversation and completes. The tutorial on building a saved reply library works from the business's own contact reasonsThe main reason a customer contacted support, tagged by the agent from a fixed, defined list so volume and other service metrics can be read by cause.: a single reply for each question, the answer up front, the parts an agent must change marked in brackets, the detail left in a linked help article, and an owner and a review date against every reply. Its strength is that the facts are decided once by the person accountable for them, so everyone who asks the same thing hears the same facts. Its limit is that it answers only the question it was written for.
An AI reply draft is produced for each incoming message. The guide to AI in customer service files drafting under assist: the draft is built from approved content, names its sources and keeps to the service tone, and the agent verifies each fact against those sources, adjusts the tone, strips out any promise the business cannot keep and sends the reply under their own name. Because nothing reaches a customer unread, the guide rates drafting as the safest and most useful assist task. Its strength is that it fits the message in front of it, including questions nobody wrote a reply for. Its limit is that its facts are only as good as the content it retrieves and the attention of the agent who checks it.
So the choice is not between a manual tool and an automated one. In both, a person sends the reply. The difference is where the accuracy is secured: in a library of owned answers written before the customer arrives, or in the content a draft is built from and the human reviewThe step where a named person checks AI-assisted work for accuracy, sources, originality, privacy and disclosure before it is published or sent. each draft receives on its way out.
Compare table
Saved replies vs AI-drafted replies
| Row label | Saved replies | AI-drafted replies |
|---|---|---|
| Who writes the words | A person writes each reply once, in advance; the agent fills the marked gaps and adds a line for this customer | A model writes a new draft for each message; the agent checks, edits and sends it as their own |
| Where the facts come from | The reply itself, checked by the owner of the policy or product, linking to the help article where the detail lives | The approved help content, policies and, through defined read-only integrations, account details retrieved for the question |
| How answers stay current | Each reply carries an owner and a review date and is listed against the changes that affect it | Only as current as the content it retrieves; a stale article or reply in that content becomes a stale draft |
| Consistency | The same wording of the facts for every customer who asks that question | Wording varies from draft to draft; the facts hold only while the source is right and the agent checks it |
| Fit to the question | Fits the one question it was written for; used for a nearby question, it answers the wrong thing | Written to the message, including unusual questions and messages that bundle several questions |
| Agent effort | Find the reply by name, fill the gaps, respond to what the customer actually said | Read the whole draft, check each fact against the cited content, rewrite any commitment |
| Failure mode | Pasted in unaltered for a customer who asked something else, or still stating a policy that has changed | Sent unread, or drifting into a polite, generic register that says little |
| Review | A monthly check of use and edits, and the quality review's questions on understanding and correctness | Sampled sent replies compared with the drafts behind them, plus the same quality review |
| Set-up and upkeep | Writing, naming and owning the library from contact reasons, then retiring what no longer fits | Choosing a drafting tool whose data terms fit, keeping the source content owned and current, writing the tone down so drafts follow it |
| Measurement | Repeat contacts after the reply, how heavily agents edit it, replies never used | Repeat contacts, the accuracy of sampled drafts, edit rates and handling time, judged alongside quality |
The rows that decide most cases are where the facts come from and fit to the question. When a question recurs and its answer is a rule the business has decided, such as a returns window or a delivery promise, the facts belong in one owned answer that every agent sends, and a saved reply holds them most directly. When questions vary so much that no library could anticipate them, but a complete and current body of help content already answers them between its articles, a draft built from that content fits each message better than a reply written for a different question.
When to choose saved replies
Choose saved replies when a question arrives often, its answer is stable, and the answer belongs to someone outside support: the person who sets the delivery promise, the returns policy or the subscription terms. The library tutorial gives each reply to that owner and ties it to the business's policy, price and product changes, so a change elsewhere reaches the reply on the day it happens. That ownership is the point of the method. A saved reply turns a decision someone accountable has made into the words every customer receives.
They are also the stronger choice when the help content is thin or unowned. A library can be written from a month of tagged contacts and a handful of help articles, and the work of writing it shows which policies are unclear and which articles are missing. If contacts are not yet tagged, the tutorial on tagging contacts by reason produces the reason list both methods need. And saved replies carry the agreed voiceThe consistent personality a brand expresses in its words across channels, with tone adjusting to the situation and the audience. into every reply without asking each agent to reinvent it, provided they are written in the tone set out by the customer service tone guidelines tutorial and finished for each case before sending.
Saved replies are the weaker choice for the long tail. A question asked a few times a month in many different forms cannot be covered without a library too large to search, and agents who cannot find the right reply reach for the nearest one. The library tutorial warns that an unedited reply sent to someone whose question differed even slightly leaves them having to write in again.
When to choose AI reply drafts
Choose drafts when the questions vary widely but the answers already exist in approved content: product details, compatibility, ingredients, how a feature works, combinations of facts that sit in different articles. The AI guide sets the conditions: the draft starts from the same approved content an assistant would use, cites the articles it relied on, and follows the service tone, and the agent checks the cited facts before sending. How grounding and retrieval work explains why that GroundingMaking a language model answer from specified material, such as documents retrieved for the question, rather than from what it absorbed in training. narrows invented answers without removing them, and why the condition of the documents usually decides the quality of what comes back. If the help centre is not already run like a product, with named owners and dates for review, as the support channels and self-service guide recommends, drafting will expose that before it saves anyone time.
Drafts also need a decision about data. A draft that reads the customer's message, and perhaps their order or account, is handling personal information the business collected for service; the comparison of first-party and third-party data notes that, with data of that kind, the business knows where it came from and what people were told and agreed to. The guide to AI disclosure and data privacy sets the rule for which tools may receive it: customer data goes only into tools whose contract rules out training on inputs, caps how long data is kept and fixes the region it is stored in, and only when the job needs it. Where the drafting runs matters for that rule. A feature built into the support tool works inside an agreement the business already has, though its AI terms still need checking, while agents pasting messages into a personal assistant account work under whatever plan they happen to use, which is the choice the comparison of standalone assistants and built-in AI works through.
Drafts are the weaker choice when the content behind them is incomplete or stale, when agents are under so much time pressure that they will send drafts unread, and for anything that commits the business. The AI guide's rule is that anything that promises a date, a refund or an exception is the agent's own wording, never the draft's, and that complaints, refunds and customers who may be vulnerable remain with people. Neither method writes those replies.
Decision table
| Situation | Choose | Why |
|---|---|---|
| A question arrives every day and its answer is a policy every customer must hear the same way | Saved reply | One owned, dated answer, changed when the policy changes |
| Many different product questions, each asked a few times, and the help articles answer them | AI drafts | No library could anticipate them, and the content already holds the answers |
| The help centre is thin, unowned or out of date | Saved replies first, then fix the content | Drafts repeat whatever the content says, confidently |
| A message bundles several questions | AI draft, checked against each cited article | Saved replies answer one question each, and stitching them invites mistakes |
| The reply commits to a refund, a date or an exception | The agent writes that part | A commitment needs a person who can make it |
| A complaint, or a customer who may be in a vulnerable situation | The agent writes the reply | A template or a draft reads as indifference when it matters most |
| Agents are measured mainly on speed | Change the measure before adding drafts | Speed targets reward sending drafts unread |
| Both are in use | Drafts that draw on the library | The agreed wording reaches drafted replies too |
Letting drafts draw on the reply library
Most teams that add drafting keep a library, and the clean arrangement is to make the library one of the sources the drafts are built from. The saved replies tutorial already warns that outdated replies matter twice over once a team drafts with AI, because a system that draws on out-of-date content passes the errors on. Turn that warning into a design: a reply written to the tutorial's pattern, one question, answer first, owner and review date, is also the kind of self-contained, dated source the grounding explainer asks for. When a policy changes, the owner updates the reply and the help article on the same day, and the drafts change with them.
The arrangement fails when the library and the help centre disagree. Duplicate and superseded documents are, in the grounding explainer's words, the commonest cause of wrong answers, and a draft may quote either version. Retire superseded replies rather than leaving them in the library, and keep the detail in the article with the reply linking to it, as the tutorial already recommends for agents. The library's record of where a reply must not be used, such as a complaint or a neighbouring question it does not quite answer, belongs in the drafting set-up too, so a complaint does not receive a draft assembled from a delivery reply.
Review both in one routine. The tutorial on quality reviews of support conversations reads the whole conversation against a scorecard whose lines ask whether the reply answered what the customer asked and whether the information was current, which catches a saved reply sent for a question it does not answer and a draft built from a stale article alike. The AI guide adds a check that samples sent replies against the drafts behind them, which shows how much agents change. The customer service audit scores saved replies and any AI assistant in separate rows; a team running both should expect to answer for each. When customers ask whether a person wrote their reply, the advice on answering customer questions about AI explains how to reply from the business's record of where AI is used, rather than improvising.
Judge both methods on what the customer received. A Repeat ContactA customer contacting the business again about the same problem, the clearest sign that the first answer did not end the customer's work. about the same problem within a set window shows a reply that did not answer; sampled accuracy shows a reply that answered wrongly. The Goodhart's law framework explains why speed alone is the wrong target: once handling time is rewarded, the quickest route is to send the draft or the template unread. Read every measure by contact reason, and use a control chart to tell whether repeat contacts for a reason moved beyond their usual week-to-week range after a change. Where contacts for one reason can be split at random between the two methods, a test shows which caused the difference; the control charts and A/B testing comparison explains when each fits. The academy checkpoint on AI drafts, owned replies and assistant tests tests the two methods at a ski and snowboard hire shop, from which answers stay in owned replies to an agent sending drafts unchanged under a speed target and a team lead judging both methods on handling time.
A pet supplies shop splits its questions between a library and drafts
Pryddock Pet Supplies, an invented online pet food and supplies shop, has 3 people answering email and chat. In one month they tag 1,240 contacts: 380 about delivery, 210 about subscriptions, 150 about returns, 330 product and allergy questions and 170 others. Of the 740 delivery, subscription and returns contacts, 96 customers wrote again about the same problem within 7 days (13%); of the 330 product and allergy questions, 59 did (18%).
The delivery, subscription and returns contacts reduce to 18 recurring questions with fixed answers, so the team writes 18 saved replies, each owned by the person who sets that policy and linked to its help article. The product and allergy questions are different: they cover about 140 distinct questions across 600 products, and the answers sit in ingredient sheets and product articles. The team trials the drafting feature built into its support tool on those questions, with the 18 saved replies and the product content as its sources. Two rules are written first: any question about a pet that is unwell is answered by the agent and points the customer to a vet, and any promised date or refund is written by the agent.
In the second week, a sample of 40 sent replies read against their drafts finds 5 drafts quoting a recipe the shop stopped selling, taken from an old article that had never been retired, and 3 replies sent with no edits at all. The article is retired, and the team agrees that every draft is read against its cited sheet before sending. Over the next 2 months, 104 of 1,480 saved-reply contacts lead to a repeat contact within 7 days (7%), and 72 of 650 product and allergy contacts do (11%). The shop keeps both methods, each on the questions it suits, and reports repeat contacts by reason rather than replies sent per hour.
Common mistakes
- Choosing one method for every question. Policy answers belong in owned replies; varied questions suit drafts built from complete content.
- Adding drafts on top of a stale help centre. The drafts repeat the stale content in a confident new form.
- Sending drafts or templates unread. Either way the customer receives an answer nobody checked against their message.
- Two sources for one fact. A saved reply and an article that disagree give drafts and agents a choice of answers.
- Accepting commitments from a draft. Dates, refunds and exceptions are written by the agent who can make them.
- Templates or drafts for complaints. Someone who complains deserves a reply a person wrote for them.
- Pasting customer messages into personal assistant accounts. Customer data goes only into tools whose terms the business has checked.
- Rewarding speed or usage. Handling time and template use rise fastest when nobody reads what is sent.
The comparison library lists the other pairs that are easy to confuse.
