How to Give Customers a Case Reference Number and SLA From an AI Voice Call

TL;DR

• Create the case once the caller confirms the details.
• Read the reference back slowly, then text it.
• State an SLA for the case type and notify a named owner.

Decide what counts as a case

Treat anything that needs work after the call as a case. A question the agent can answer on the spot is not one. A disputed charge, a missing refund or a repair request is. Buyers raise this themselves: they ask what happens to sensitive issues or a disputed amount.

Next, fix the minimum fields the agent must capture before it creates a case:

• Verified identity
• Case type
• A one-line issue summary
• Channel
• Consent
• Language
• Callback number

On an inbound line, the agent can look the caller up from the number they ring from. That gives it a head start on identity.

Then give each case type an SLA tier and an owner queue. A billing dispute might go to the billing team with a one-day first response. An outage report might go to operations with a one-hour response. Your support lead should sign off on this map, because the agent can only promise what a queue can deliver.

Last, design the reference format:

• Keep it short, about six to eight characters.
• Leave out look-alike characters such as 0 and O, or 1 and I.
• Start with a type prefix, such as BIL for billing.
• End with a check character calculated from the others, so your system can spot a mistyped reference.

BIL-7K3M4 is easy to say and type. The final 4 is the check character.

Create the case before the call ends

Create the case only after the caller confirms a short summary. Have the agent read it back: "You are reporting a double charge on your March bill, and you want a refund. Is that right?" A "yes" is the trigger. Without it, your queue fills with cases built on misheard details.

Next, make the agent call your case system as a tool in the middle of the call. That system might be a helpdesk, a CRM or a case API. Buyers ask early whether a voice agent replaces their stack or plugs into it. Let the case system generate the reference, so the AI model only reads out what the API returns.

Build in these safeguards:

• Idempotency key. Send the call ID with every create request. A retry or a dropped line then returns the same case.
• Attachments. Add the transcript, the recording link and the summary, so the owner never asks the customer to repeat themselves.
• Failure path. If the API times out, the agent says so, logs a callback task and promises a message. It never offers a made-up reference.

Tell callers they are talking to an AI at the start of the call. The rules differ by region:

• EU: Article 50 of the AI Act has applied since 2 August 2026. Any AI that talks to people, including automated phone systems, must say it is AI from the first interaction.
• US: AI-voice calls already need consent under the TCPA, following the FCC's 2024 ruling. A rule to say "this is AI" at the start was still not final in mid-2026.
• US states: Disclosure laws are spreading state by state, and Colorado is one example.

Check the rules in each market you call.

Read the reference back, then send it

Read the reference in short chunks with a pause between them. For BIL-7K3M4, the agent says "B, I, L", pauses, says "seven, K, three", pauses, then says "M, four". Chunks of two or three characters fit what a listener can hold in their head. If callers mishear letters, add a spelling word such as "K as in Kilo". That stops "M" and "N" from blurring on a bad line.

Then ask the caller to say it back. If one chunk is wrong, the agent repeats only that chunk and asks again. Loop until the caller's version matches the one the system returned.

Say the reference twice: once at confirmation, and again in the closing line, right before the SLA.

Send the written copy during the call, by SMS or email. Trigger the message from the same API response that returned the reference, so the text can never differ from what was spoken.

Written copies matter because voice is linear. A caller cannot re-read a string they missed. Text-to-speech engines also tend to mangle long strings, running digits together or reading "7K3M" as a word. Short references with a fixed pattern reduce that risk.

Test before launch with real numbers, in every language you support. Buyers ask for local accents and languages, and a reference that sounds clear in one language can fall apart in another. Try dozens of references, and have someone with a different accent write down what they hear.

Set an SLA the team can keep

Pick the SLA from the case type and its severity, then let the agent state it. Keep one small table that your support lead owns. These times are examples to adjust.

Promise first response, not resolution. The agent can't know how long an investigation will take, but it can promise that a named team will be in touch. Promise a resolution time only for routine types, where the work is known.

Say the SLA as a promise with a clock: "You will hear from the billing team by 4 pm tomorrow." A fixed time is easier to hold than "within one day".

Start the clock when the case is created. Pause it only for defined reasons, such as waiting on the customer, and log each pause.

Check your regulator's complaint rules before you set any deadline, because regulators enforce them. The CBN fined institutions N430M in 2025 for slow complaint handling. Keep your internal SLA tighter than the legal limit.

Offer a human route in the script: "If you would rather speak to a person, I can book a callback." Gartner's 2026 survey found that 87% of customers say a company using GenAI for service must offer a way to reach a human.

Notify the owner and chase pickup

Route the case by type to a named person or queue, and notify them the moment it exists. Use email, Slack or your ticket tool. Each message should carry four things:

• The one-line summary
• The transcript link
• The SLA deadline
• The exact words the agent promised the customer

That last item matters most. The owner then knows the promise, not just the problem.

Next, set an escalation rule. If the owner has not acknowledged the case in a set number of minutes, alert the supervisor. Pick the minutes by tier. An outage might get ten. A routine request might get an hour.

Callback requests need a home too. Give that queue an owner and a clock.

Then close the loop with the customer. Send a short message when the owner picks the case up. Send another if the SLA is at risk. Silence after the call is where trust is lost.

Last, check the agent's work. A mid-sized insurer's QA team cross-checks each logged interaction against the customer's policy number. You can do the same with the reference and the SLA.

Insight7 can help here. Every call can be scored against a weighted scorecard your team owns, so add two criteria: "reference read back" and "SLA stated". Alerts can flag low scores by email or Slack, so check that feature in the app first. In beta, inbound voice lines can look the caller up and send callback requests to a queue. You can also test a campaign's agent by talking to it in the browser, with no phone number needed.

Related reading: How to Design Tool Calls That Let AI Voice Agents Look Up Accounts, Create Cases and Book Callbacks, How to Prepare Your Knowledge Base for an AI Voice Agent to Answer Customer Questions and How to Find Knowledge Base Gaps From Questions Your AI Agent Couldn’t Answer.

Questions people also ask

Is a ticket number the same as a case number?

Often, yes. Many systems use the words for the same thing. A ticket is usually one unit of work, and a case can group several tickets about one issue. Give the caller one reference.

What should a case reference number look like?

It should be short, with a type prefix and a fixed pattern. End it with a check character, so your system can spot a mistyped reference when a caller dictates it back. Avoid look-alike characters.

Can an AI voice agent create a case in my CRM or helpdesk?

Yes, if your system has an API that creates a record and returns its ID. The agent calls it as a tool mid-call. Check the required fields and rate limits first.

What happens if the case system is down during the call?

The agent tells the caller the system is unavailable. It logs a callback task in a fallback queue and promises a message by a stated time. A background job retries the create, and an alert tells operations.

How long should an SLA be for a customer complaint?

Promise a first response within hours or one business day, and a final answer inside your regulator's limit. Make the first-response promise tighter for complaints, because an unanswered complaint escalates.

Does an AI voice agent have to say it is AI?

In a growing list of places, yes. Put the disclosure in the agent's first sentence, before it asks for any detail. Keep the same line in every language. Your legal team should confirm the wording for each market.

How do I check the AI gave the right reference and SLA?

Compare the reference in each transcript with the one in the case system, and flag any mismatch. Then check that the stated SLA matches your table for that case type. Scoring every call makes this routine.

Close

Your agent can only promise what your queues can deliver. Before you start, confirm that your case system creates a record through an API and returns the reference in the response.

Book a demo of Insight7