The best agentic AI use cases in customer service to start with are high-volume intents with written rules, low action risk and data the agent can reach: order status, login help and simple account questions. Intents that look just as simple, like cancellations, address changes, returns and promo codes, often hide undocumented business rules and belong in a second wave, after you write those rules down.
In short
- Pick the first intents for an AI agent by score, not gut feel: volume, resolution complexity, policy clarity, action risk, data access and customer emotion, each rated 1 to 5 and weighted.
- Add one gate before any intent goes live: if you cannot tell from the transcript and the action log whether the agent got it right, it does not go first.
- Cancellations, address changes, returns and promo codes look like one action but hide decisions experienced agents make from memory, which an AI agent will not know.
- Find hidden rules before launch: read 50 resolved tickets per intent, interview your best agents, list every exception, then score policy clarity.
- Splitting an intent (standard return inside the window versus every exception) often turns a poor first candidate into a good second-wave one.
This guide covers the first scoping decision in the transition to agentic customer service.
Which customer service intents should an AI agent handle first?
An AI agent should first handle intents that many customers raise, that follow written rules with few exceptions, and where a wrong action is cheap to undo. Typical first candidates are order status, password and login help, opening hours and plan information. Anything that moves money, ends a contract or depends on judgment waits.
An intent is a contact reason: the thing the customer wants done, such as “where is my order” or “cancel my subscription”. Your helpdesk’s contact reason list, however messy, is the starting inventory.
Choosing by volume alone is the common mistake: volume shows where the savings are, not whether the AI agent can get it right, what a mistake costs or whether anyone would notice. Many vendor guides put returns and account management in the safe group, yet that is where undocumented rules live. The rest of this guide turns the heuristic into a score you can defend.
What are typical agentic AI use cases in customer service?
Typical agentic AI use cases in customer service are order and delivery status, account and login help, returns and refunds, address and order changes, subscription changes and cancellations, billing questions, appointment booking and ticket intake and routing. They differ less in technology than in risk: some only read data, others move money.
An AI agent takes the action in your systems, so the action matters as much as the question. The table groups common use cases by the autonomy tier the action usually needs, using the tiers from our guide to AI agent autonomy levels.
| Use case | What the agent does | Usual autonomy tier | First wave? |
|---|---|---|---|
| Order and delivery status | Reads order and carrier data, explains the status | 0. Inform | Usually yes |
| Password and login help | Sends a reset link to the verified address | 1. Act alone | Usually yes |
| Appointment booking | Books inside open slots | 1. Act alone | Often |
| Plan change or upgrade | Changes the plan, explains the new price | 2. Act within a limit | Second wave |
| Address change | Edits the order before dispatch | 1. Act alone | After its rules are written |
| Return request | Checks eligibility, issues a label | 2. Act within a limit | After its rules are written |
| Cancellation | Stops renewal or cancels now with a refund | 1 or 3, depends on the action | After its rules are written |
| Promo code problems | Applies or replaces a discount | 2. Act within a limit | Later (hidden rules) |
| Billing disputes and double charges | Investigates, refunds or escalates | 3. Act with approval | Later |
| Ticket intake and routing | Classifies and routes to the right queue | 1. Act alone | Often, behind the scenes |
The last column is a typical outcome, not a rule; your own score decides.
How do you score support intents for automation?
Score each intent from 1 to 5 on six criteria (volume, resolution complexity, policy clarity, action risk, data access and customer emotion), multiply by the weights below, and divide the total by 5 to get a priority score from 20 to 100. Then apply the measurability gate and the knockout rules before anything goes live.
On every criterion, 5 means “a better first candidate”, so higher is always safer and more valuable to automate early.
The intent prioritization scorecard
| Criterion | Weight | 5 (good first candidate) | 3 | 1 (poor first candidate) |
|---|---|---|---|---|
| Volume | 20 | 15% or more of monthly contacts | 4% to 7% | Under 2% |
| Resolution complexity | 15 | One step, one system | 2 or 3 steps, or 2 systems | Many steps, judgment at each |
| Policy clarity | 20 | Written rule, no unwritten exceptions found in 50 tickets | 3 to 5 unwritten exceptions in 50 tickets | More than 10, or rules contradict |
| Action risk | 20 | Tier 0: inform only | Tier 2: acts within a money limit | Tier 4: must never act |
| Data access | 10 | Agent can read and write everything it needs through an API | Reads all, writes some | Key data is in emails, PDFs or people’s heads |
| Customer emotion | 15 | Calm, transactional | Mildly annoyed | Distressed, vulnerable or angry |
Fill in the in-between values the obvious way: volume 4 is 8% to 14% and volume 2 is 2% to 3%; policy clarity 4 is 1 or 2 unwritten exceptions and 2 is 6 to 10. Action risk follows the autonomy tier: tier 0 scores 5, tier 1 scores 4, tier 2 scores 3, tier 3 scores 2 and tier 4 scores 1. If an intent covers actions in two tiers (cancel at end of term versus cancel now with a refund), score the riskier one, or split the intent.
Priority score = (volume × 20 + complexity × 15 + clarity × 20 + risk × 20 + data × 10 + emotion × 15) ÷ 5.
Bands: 80 or more goes in the first wave. 65 to 79 goes in the second wave, once the weakest criterion is fixed. Under 65 stays with people, or the AI agent assists a person, until you rescore.
The quality-measurability gate
Before an intent goes live, ask: could a reviewer, or an automated scorer, tell from the transcript and the action log whether the AI agent got this right? You must be able to write the correct outcome as scorecard criteria, such as “gave the carrier status that matches the tracking data” or “stopped renewal and did not cancel immediately unless asked”. If not, the intent fails whatever its score, because its errors stay invisible until customers complain. See how to QA AI agents for writing those criteria.
Knockout rules
Three results stop an intent from going first, even with a high total:
- Policy clarity 1: the rules contradict or most cases use unwritten ones, so the agent will improvise.
- Action risk 1: tier 4 actions (deleting personal data, changing bank details) stay with a person.
- A failed measurability gate.
Worked example: ranking 8 support intents
The worked example below is an illustration, not customer data: a subscription e-commerce team with 40,000 contacts a month, scoring eight intents. Order status and login help land in the first wave, plan change in the second, and the four “looks simple” intents score under 65 until their hidden rules are written down.
| Intent | Share (contacts) | Vol | Cplx | Clarity | Risk | Data | Emotion | Score | Gate | Wave |
|---|---|---|---|---|---|---|---|---|---|---|
| Order status | 22% (8,800) | 5 | 5 | 4 | 5 | 5 | 4 | 93 | Pass | 1 |
| Password and login help | 9% (3,600) | 4 | 4 | 5 | 4 | 4 | 4 | 84 | Pass | 1 |
| Plan change | 5% (2,000) | 3 | 4 | 4 | 3 | 4 | 5 | 75 | Pass | 2 |
| Address change | 6% (2,400) | 3 | 3 | 2 | 4 | 3 | 4 | 63 | Pass | Later |
| Return request | 14% (5,600) | 4 | 3 | 2 | 3 | 4 | 3 | 62 | After rules | Later |
| Cancel subscription | 10% (4,000) | 4 | 3 | 2 | 2 | 4 | 2 | 55 | Pass | Later |
| Promo code not working | 7% (2,800) | 3 | 3 | 1 | 3 | 3 | 3 | 52 | Fail | Knockout |
| Billing dispute | 4% (1,600) | 3 | 2 | 3 | 2 | 3 | 1 | 47 | Pass | Later |
To check one row: order status is (5 × 20 + 5 × 15 + 4 × 20 + 5 × 20 + 5 × 10 + 4 × 15) ÷ 5 = 465 ÷ 5 = 93.
What the ranking tells this team:
- The first wave covers 31% of contacts (12,400 a month) with two intents: a smaller start than “everything under five minutes”, and one you can measure.
- Returns have the second-highest volume but score 62. The problem is policy clarity: in 50 resolved return tickets the team found 7 decisions no help article described.
- Promo codes are knocked out. Nobody could write down when an agent honors an expired code, so no scorer could judge the AI agent’s answer either.
- Cancellations score on their riskier action. Cancel now with a refund needs approval (tier 3), so action risk is 2, not the 4 of a plain renewal stop.
Now split the return intent. “Standard return inside the window, eligible item” is 10% of contacts. With its rules written down, complexity and clarity rise to 4: (4 × 20 + 4 × 15 + 4 × 20 + 3 × 20 + 4 × 10 + 3 × 15) ÷ 5 = 365 ÷ 5 = 73. That moves it into the second wave. The other 4% (damaged, personalized, late, goodwill) stays with people and is handed over with context.
Which “simple” intents hide business rules?
The intents that most often hide business rules are cancellations, address changes, returns, promo codes, refunds for late deliveries and payment method updates. Each one looks like a single action, but your best agents make several decisions along the way from experience, and none of those decisions is written where an AI agent can read it.
| Intent | Questions your agents answer from memory |
|---|---|
| Cancel subscription | Stop renewal or cancel now? Prorated refund or not? Is a retention offer allowed, and how often? Is the customer inside a legal withdrawal period? |
| Address change | Has the order been dispatched? Can the carrier redirect it? Is a new address plus a new card a fraud signal? Does it change tax or shipping cost? |
| Return request | Inside the window? Which items are excluded? Who pays return shipping? What if the item is damaged, or the customer wants a discount to keep it? |
| Promo code not working | Can codes stack? Which products are excluded? Is it a partner code with its own rules? When do agents honor an expired code anyway? |
| Late delivery refund | Automatic or on request? Does it depend on the carrier, the delay or the customer’s tier? |
| Update payment method | Is there a failed payment to retry? Does it restart a billing cycle? What identity check is required? |
Some hidden rules are legal ones. In the EU, a consumer generally has 14 days to withdraw from a distance contract “without giving any reason” (Directive 2011/83/EU, Article 9), and Article 16 lists 13 exceptions, such as goods made to the consumer’s specifications, goods that spoil quickly and sealed hygiene goods unsealed after delivery. An agent that says “all sales are final” or “you can return anything” is wrong either way. Check your markets with counsel and write the result into the rule.
A quick test: if two experienced agents would handle the same ticket differently, the intent has an unwritten rule, and the AI agent will apply one version to everyone.
How do you uncover hidden rules before launch?
Uncover hidden rules in four steps: read 50 resolved tickets per intent and note every decision made, interview two or three of your best agents about the cases where they leave the standard answer, write each exception as a rule with an owner, and turn each one into a test case. Score policy clarity only after this work.
1. Read 50 resolved tickets per intent. Not the 50 most recent, but a mix: about 20 ordinary ones, 10 reopened, 10 escalated and 10 with a low satisfaction score or the longest handle time. For each, note the decision, the data used and whether the rule applied is written anywhere. The count of unwritten decisions sets the policy clarity score.
2. Interview two or three top agents, thirty minutes each, with questions that pull out exceptions:
- “When do you not do what the help article says?”
- “What do you check before you do this?”
- “Which customers get treated differently, and why?”
3. List the exceptions and give each an owner. Write every exception as a rule with a trigger, an action and a limit, and decide whether the AI agent applies it or hands over. The policy card template in AI agent policy management keeps the human macro and the AI agent instruction pointing to the same rule.
4. Turn the exceptions into test cases: a customer message, the expected action and the expected answer, run before launch and after every change. See chatbot testing for scripted conversations and AI agent testing for actions and tool use.
Plan about half a day per intent for the tickets, plus the interviews: cheaper than finding the rules from complaints.
Should an intent run as a deterministic or a generative flow?
Use a deterministic flow, with fixed steps and rules in code, when the decision follows clear criteria such as eligibility windows, limits and identity checks. Let the generative model handle what code cannot: understanding the customer’s request, asking for missing details and explaining the outcome. Most good intents mix the two.
Anthropic describes workflows as “systems where LLMs and tools are orchestrated through predefined code paths”, and agents as systems where models “dynamically direct their own processes and tool usage”, and recommends “finding the simplest solution possible, and only increasing complexity when needed” (Anthropic, Building effective agents).
A practical split for support intents: identity checks, order lookups, eligibility (windows, exclusions, limits), the API call with its enforced limits, legally required wording and fixed handover triggers belong in code or the tool. Understanding the request, asking for missing details, choosing which approved action fits, tone and spotting frustration belong to the model. Eligibility for money or contracts is never the model’s call alone.
There is a tension. OpenAI’s guide to building agents names refund approval in customer service as an example of “complex decision-making” where agents fit, and says otherwise “a deterministic solution may suffice” (OpenAI, A practical guide to building agents). The same guide says “canceling user orders, authorizing large refunds, or making payments” should trigger human oversight until confidence in the agent grows. Both hold: judgment-heavy intents are where agents add the most value and need the most evidence first. They come later, not never.
How do you know when to add the next intent?
Add the next intent when the current ones hold steady for several weeks: correct resolutions stay at your target, escalations happen at the right moments, and the reviewed failures have known causes with fixes in place. Rescore the waiting intents at the same time, since written rules and new data access change their scores.
Use evidence, not calendar dates:
- Correct resolution, not containment. A closed conversation is not a solved one; see correct resolution rate for how to measure the difference.
- Handover quality. The agent hands over at the right trigger, with context, so the customer does not repeat themselves (human to AI handoff QA).
- Failure causes. Each failed conversation gets a cause: missing rule, wrong data, wrong action, tone. Fix a recurring cause before you widen.
- An owner who reviews each live intent weekly; see the AI agent manager role.
Your level in the agentic support maturity model matters too: early on, add one intent at a time; a team that scores every AI conversation and has a working change process can add a small group at once.
When is this scorecard not for you?
This scorecard is more than you need if you handle a few hundred contacts a month, have only three or four contact reasons, or your AI agent only answers questions from approved content and takes no actions. In those cases, read a sample of tickets, write the rules and test the answers by hand.
It also does not replace a clean contact reason list (unreliable tags give wrong volume scores) or legal review: the scorecard tells you an intent has legal rules, not what they are.
Where Kaizo fits
Kaizo checks every support conversation, human or AI, against your own rubric, and shows you what to fix. AutoPilot scores each conversation when it resolves, so you can see where the AI agent and your team get it right and where quality breaks: the evidence to decide when to add the next intent. See how QA for AI agents works, or book a demo to see it on your own conversations.
How to start this week
You can score your first intents in one week with one analyst, one team lead, a helpdesk export and a spreadsheet. By Friday you have a ranked list, an exception list for the top candidates and an agreed first wave of two or three intents.
- Monday: export last month’s contacts by contact reason and pick the top 10 by volume.
- Tuesday: score volume, complexity, action risk (from the autonomy tiers), data access and emotion.
- Wednesday and Thursday: for the top 4, read 50 resolved tickets each, count the unwritten decisions to set policy clarity, and interview two top agents.
- Friday: run the gate and knockout rules, compute the scores, agree the first wave with your AI agent owner and put the exceptions into policy cards.
Then check the controls you need before go-live in the agentic customer service guide.
Frequently asked questions
What are real examples of AI agents in customer service?
AI agents that check an order’s delivery status and explain a delay, reset a password, book or move an appointment, process a standard return and issue a label, or change a subscription plan. They are agents rather than chatbots because they take the action in your systems, not just describe it.
What is the difference between a customer service chatbot and an AI agent?
A rule-based chatbot follows fixed intents and scripted answers, usually with buttons or keywords. An AI agent understands free text, decides which steps to take and acts through your tools, such as refunding an order. That freedom is why the choice of intents and their rules matters more.
How narrow should the first AI agent scope be?
Narrow enough that you can check every result. Two or three intents with written rules and low action risk suit most enterprise teams; in the illustration above, two intents covered 31% of contacts.
Can an AI agent handle refunds from day one?
Rarely. Small refunds inside a written policy can run within a limit once the rules and test cases exist, but refunds combine money, exceptions and emotion, so they usually score in the second wave or later. Larger refunds and disputes need a person’s approval.
How can AI agents improve customer service without hurting quality?
When they resolve simple intents fully and hand everything else to a person with context. Quality suffers when teams automate intents with unwritten rules or measure containment instead of correct resolution; scoring AI conversations against your team’s standard keeps that visible.
