Scoping: automate, assist or keep human
Start by listing the contacts the service team actually handles, by reason, and sort each into one of three categories. Automate means an AI system answers or acts without a person reviewing each instance. Assist means AI prepares something, a draft, a summary, a suggested tag, and a person decides. Keep human means AI plays no part in the decision, though it may still help with admin around it.
Three questions place each contact reason. How much harm can a wrong answer do to the customer or the business? Is there approved content or a written rule that fully answers it? And does it need judgement, discretion or empathy that a policy cannot capture? Low harm, full coverage and no judgement points to automate. Any doubt on the first two points to assist. A need for discretion or empathy keeps it human.
| Task | Default | Guardrail | Why |
|---|---|---|---|
| Routine questions covered by help content | Automate | Answer only from approved content; hand over when it does not cover the question | Low harm, full coverage |
| Order, booking or account status | Automate | Read-only access through defined integrations; no changes without confirmation | Factual lookup with a clear source |
| Reply drafts for agents | Assist | Agent edits and sends; draft cites the content it used | Agent owns the words and the commitment |
| Triage, tagging and routing | Assist, then automate for proven tags | Sample tags weekly; urgent and sensitive tags always route to people | Errors delay rather than decide |
| Thread and call summaries | Assist | Agent checks before the summary is saved | Summaries become the record |
| Complaints | Keep human | AI may summarise the history for the handler | Judgement, recovery and accountability |
| Refunds, credits and exceptions | Keep human, or a narrow rule | Pre-approved rule only; anything outside it goes to a person | Commits money and sets precedent |
| Customers in vulnerable situations | Keep human | Signals route immediately to trained staff | Harm from a wrong or cold answer is high |
| Legal, safety or regulatory matters | Keep human | Route on keywords and topics; never answer automatically | Consequences outlast the conversation |
The AI for marketing automation guide covers classification, routing and agents in general terms; this guide applies them to service work, where the customer is usually waiting for the answer.
Answering from approved content
An assistant that answers customers should answer from a defined body of content: the help centre, product documentation, policies and current service notices, retrieved for each question. It should not answer from the model's general knowledge, which may be out of date, may describe another company's policy and has no reference point to check against. The instruction to give is simple: answer only from the supplied content, cite the article used, and say plainly when the content does not cover the question, then offer a person. The grounding and retrieval explainer shows how a system finds the passages it answers from, and why 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. That one instruction is the core of a longer set, and the tutorial on writing the standing instructions for an AI assistant writes the rest: scope, hand-over topics, disclosure and handover lines, tone and a test conversation for every rule.
That makes the help centre the assistant's limit. The support channels guide treats the help centre as a product with an owner, and that matters more once an assistant reads it: an out-of-date article becomes an out-of-date answer delivered confidently in chat. Every policy change, price change or service notice needs an owner who updates the content before the assistant starts giving the new answer. If the assistant is already quoting old prices or wrong policies, the diagnosis for an AI assistant giving customers wrong answers traces a sample of conversations to the library, retrieval, the instructions, the test set and the handover.
Keep personal data out of the content the assistant draws on, and give it access to account information only through defined, read-only integrations, so that one customer's details cannot leak into another customer's answer. Account and order data is First-Party DataData a business collects directly from its own customers, visitors and users., collected for service, and the disclosure and data privacy guide sets out which tools may handle it and what their terms must say.
Drafting replies for agents
Reply drafting is the assist use with the clearest benefit and the lowest risk, because a person reads every word before it is sent. The draft should start from the same approved content, cite what it used and follow the brand's service tone. The agent checks the facts against the cited content, adjusts tone to the customer and the situation, removes anything the business cannot commit to, and sends it as their own reply. If the service tone has never been written down, the tutorial on writing customer service tone guidelines sets it out with approved examples that drafts and agents can both follow.
Two things go wrong. Agents under time pressure start sending drafts unread, which turns an assist use into an automated one without anyone deciding it. And drafts drift towards a generic register that sounds polite and says little, which customers notice. The guardrails are a sampling check of sent replies against their drafts, so the team can see how much agents edit, and a rule that any reply making a commitment, a date, a refund, an exception, is written or rewritten by the agent rather than accepted from the draft. Where the team already keeps a library of saved replies, the comparison of saved replies and AI reply drafts sets out which questions suit each method and how drafts can draw on the library.
Triage, tagging and summaries
Most service queues spend a lot of effort reading and sorting before anyone answers. A model can read incoming contacts, assign a reason tag, flag urgency and sentiment, detect the language and route to the right queue. Errors here usually delay a contact rather than decide its outcome, which makes triage a good candidate to move from assist to automate once a tag has proven reliable in sampling. Language detection feeds the wider plan in the guide to supporting customers in several languages and time zones, which decides which languages each channel answers in and who reviews translated replies.
Some tags must never be left to chance. Signals that a customer may be in a vulnerable situation, threats to safety, legal language and complaints should route to people immediately, with conservative rules that accept false alarms rather than missing a real case. Keep a person reviewing a sample of routing decisions each week, and treat a rising share of contacts in an "other" tag as a sign that the categories need revisiting.
Summaries save the next agent from re-reading a long thread or call. They also become the record, so a summary that drops the customer's main complaint or invents a promise does lasting damage. The agent who handled the contact checks the summary before it is saved, and summaries never replace the original thread, which stays available for anyone who needs it.
Handover to a person
The handover is where an automated assistant is judged. The support channels guide sets the rule: the customer can reach a person easily at any point, without having to defeat the bot first, and the person receives the full conversation so the customer does not repeat themselves. Design the handover before deciding what the assistant answers.
Hand over automatically when the content does not cover the question, when the customer asks for a person, when the conversation loops, when sentiment turns sharply negative, and when any keep-human topic appears. When no person is available, say so honestly and offer a realistic alternative, a callback or an email with a stated response time, rather than keeping the customer in a conversation that cannot help. When the logs show customers asking for a person instead of getting an answer, the diagnosis for an AI assistant not resolving customer questions traces the pattern to the scope, the library, the instructions, the route to a person or the measure.
Complaints deserve particular care. The service recovery and complaints guide describes the recovery sequence: acknowledge, apologise where appropriate, fix, follow up. That sequence depends on a person who can take responsibility and use discretion, and an assistant that tries to handle a complaint usually makes it worse. Map the service journey with a customer journey map to see where frustrated customers arrive, and make sure those points lead straight to people.
Disclosure and sampling review
Tell customers when they are talking to an automated assistant. The European approach to AI regulation, as the Commission's overview sets it out, includes transparency expectations so that people know when they are interacting with an AI system. The US Federal Trade Commission's business guidance holds that deception standards do not change with the technology, so an assistant presented as a person, or answers that mislead, are treated like any other deceptive practice. The disclosure and data privacy guide gives the practical rule: say it is an automated assistant at the start and whenever asked, and offer a route to a person. This is general direction, not legal advice.
Disclosure does not make answers right, so sampling reviewA person reading a random sample of AI answers or generated copy on a fixed rhythm, plus every case that ended badly, and routing findings to the owners. is the other half. Each week, a person reads a random sample of automated conversations and AI-assisted replies, plus every conversation that ended in a handover, a complaint or a negative rating. For each one they check whether the answer matched the cited content, whether the content itself was correct and current, whether the handover happened when it should have, and whether the tone fitted the situation. Findings go to the owners who can fix them: content gaps to the help centre owner, routing errors to whoever maintains the rules, recurring failures to the decision about scope. Conversations handled entirely by people deserve the same reading, and the tutorial on reviewing support conversations for quality sets out the scorecard, sampling and reviewer calibration for them. Before an assistant first answers customers, the AI customer service launch checklist gathers scope, content, handover, disclosure, data and testing into one gate to tick.
Scoping an assistant for a regional broadband provider
Brinemoor Broadband, an invented regional internet provider, lists its 14 most common contact reasons. It automates 4, including outage status and router setup, all covered by help articles; assists on 6 with agent drafts; and keeps 4 human, including complaints, bereavement requests and disputed bills. In the first month the assistant handles 2,300 conversations. The weekly sample of 50 finds 6 answers taken from an article describing an old router model, which the content owner rewrites. Handover analysis shows 180 customers typed "agent" within their first 2 messages; most were asking about bills, which the team had wrongly assumed the help content covered, so billing questions move to assist. Repeat contact within 7 days for automated conversations falls from 22% in the first fortnight to 15% after the fixes, and the team reports that figure, not the number of conversations the assistant closed.
Measuring service quality, not deflection
The tempting KPIA metric tied directly to an objective. for an assistant is deflectionThe share of customer contacts handled without ever reaching a person, usually by an assistant or help content; read beside resolution, never as the goal.: the share of contacts that never reached a person. It rewards the wrong behaviour. An assistant that makes it hard to reach a person, or answers wrongly and leaves the customer to give up, scores well on deflection while service gets worse. Customers who give up do not show as contacts; they show later as complaints, bad reviews and a rising Churn RateShare of customers at the start of a period who left during it..
Measure what the customer experiences instead. Resolution: did the issue get fixed, judged by the absence of 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. on the same subject within a set window. Accuracy: the share of sampled answers that matched current approved content. Handover quality: how quickly a customer who wanted a person reached one, and whether they had to repeat themselves. Satisfaction and effort scores, read by contact reason rather than overall. And for assist uses, agent edit rates and handling time, read together with quality so that speed is never bought with errors.
Read every measure by contact reason and compare automated, assisted and human-handled contacts for the same reason. That comparison, rather than a single headline number, shows which scoping decisions are working and which contact reasons should move between categories. To find out what letting an assistant answer first changes for sales, the AI chat assistant test designs a randomised comparison of assistant-first and staff-only chat on product pages.
Common mistakes
- Scoping by volume instead of risk. The most common contact reason is not automatically safe to automate.
- Answering from the model's general knowledge. Without approved content as the limit, answers have no reference point to check.
- Hiding the route to a person. A handover that customers have to fight for turns a convenience into a complaint.
- Letting an assistant handle complaints or refunds. Discretion and accountability belong to people.
- Agents sending drafts unread. That silently turns assist into automate.
- Not disclosing the assistant. Customers should know when they are not talking to a person.
- Out-of-date help content. The assistant repeats whatever the content says, confidently.
- Reporting deflection as success. Customers who give up look like savings until they leave.
