How to Implement AI in Your Call Centre

Research date: 22 September 2026. Last updated: 22 September 2026.

Implementing AI on a call centre line takes anywhere from ten minutes to six weeks, depending entirely on which product you buy and how deeply it has to reach into your existing systems. The work that decides success is not the setup. It is choosing three call types, mapping your integrations honestly, and defining escalation before a single call is answered.

This guide is written from vendor documentation and published platform behaviour, read on 22 September 2026. It does not include screenshots of any admin console, and it does not claim a hands-on walkthrough of any specific product.

What you need before you start

• A phone number you can forward, or SIP access to the line you want the AI to answer.

• Admin access to the system the AI must write into — Salesforce, HubSpot, Zendesk, Microsoft Dynamics 365, your PMS or your scheduling tool. Read-only access is not enough if you want bookings created.

• Two weeks of call data, or failing that, a list of the reasons people actually ring you, ranked by volume.

• A named human escalation destination — a queue, a rota, or a mobile number. Not "the team".

• Agreement on two numbers before launch: a containment target and an abandonment target.

• Sign-off from whoever owns compliance for call recording and data processing.

If you cannot produce the third and fourth items, stop here. Those are the two that make the difference between a deployment that works and one that gets switched off in month two.

Step 1 — Pick three call types, not thirty

Choose the three reasons people ring you most often, and automate only those. The most common failure in this category is launching with a long list of intents, each half-configured, none reliable.

Good first candidates share three traits: the caller wants one specific thing, the answer exists in a system you can reach, and getting it wrong is recoverable. Opening hours, appointment booking and rescheduling, order and delivery status, and simple account queries all qualify. Complaints, billing disputes and anything requiring clinical, legal or financial judgement do not.

Write the three down. Everything in the rest of this guide applies to those three and nothing else. You widen later, once each one runs clean for a fortnight.

Step 2 — Map the integrations you actually need

This is the step that decides whether the project delivers anything, and it is the one most often skipped.

For each of your three call types, write down the system that holds the answer and whether the AI needs to read it or write to it. Checking an order status is a read. Booking an appointment is a write. The two are very different asks, and a vendor that says "we integrate with Salesforce" may only mean the first.

Ask every vendor the same question in the same words: does the agent write back into this system during the call, in real time, or does it send a notification somebody re-keys afterwards? A notification that a human re-types is not an integration. It is a message-taking service with extra steps, and it will not reduce your handling time.

Brilo publishes integrations with Salesforce, HubSpot, Zoho, Zendesk, Microsoft Dynamics 365 and cal.com, among others, and states it connects to over 6,000 apps. Source: brilo.ai, read 22 September 2026.

Step 3 — Decide where the AI sits in your stack

You have three architectural options, and they cost very different amounts of effort.

In front of your existing platform. The AI answers first, handles what it can, and passes the rest through to your current queue. Usually done with number forwarding or SIP. This is the fastest route and the one that requires no change to the platform you already run.

Inside your existing platform. Your CCaaS vendor's own AI, bought as a tier or an add-on. Genesys meters this with AI Experience tokens; Five9 sells Intelligent Virtual Agents and Agent Assist as add-ons on a platform with a 50-seat minimum; Talkdesk includes its AI components in the standard plans. Sources: genesys.com/pricing, five9.com/pricing, talkdesk.com/pricing, all read 22 September 2026.

Built from infrastructure. Amazon Connect at $0.038 per voice minute plus telephony, assembled by your own engineers. The lowest unit cost and the highest total cost of ownership. Source: aws.amazon.com/connect/pricing, read 22 September 2026.

Most teams implementing AI for the first time should take the first option. It is reversible.

Step 4 — Define the escalation path before go-live

Decide, in writing, what sends a call to a human, where that call lands, and whether the context travels with it.

Three triggers are non-negotiable: the caller asks for a person, the caller is distressed, and the AI has failed to understand twice. Anything clinical, legal or financial goes straight through with no attempt at resolution.

Then decide what the human receives. A transfer that arrives with no context forces the caller to repeat everything, which is worse than not automating at all. Ask the vendor whether a warm transfer carries the transcript and any captured details.

Finally, decide what happens when nobody is available. Voicemail with a callback commitment is acceptable. Silence is not. Test this path before launch, not after, because it is the one that generates complaints.

Step 5 — Launch on the calls you are already losing

Point the AI at after-hours and overflow first, not your main line.

The logic is that those calls currently go unanswered. Anything the AI captures is a gain, and nothing can get worse, because the alternative outcome was a voicemail or a dropped call. It also gives you two weeks of real transcripts from real callers before the AI touches a single daytime conversation.

Tell your staff before you switch it on. Teams that skip this find agents and regular callers hanging up on the agent because nobody explained what it was. Put the escalation number in the staff handbook and brief the people who take the transfers.

Set the go-live for a Tuesday. Not a Friday, and not the day before anyone's leave.

Step 6 — Read every transcript for two weeks

This is not optional, and it is where the real configuration happens.

Read them daily for the first week and every other day for the second. You are looking for four things: calls where the AI answered correctly but awkwardly, calls where it misunderstood, calls it escalated that it should have handled, and calls it handled that it should have escalated. The last category is the one that matters most.

Fix the phrasing as you go. Most early failures are wording problems rather than capability problems: the agent asks a question in a way that invites an answer it cannot parse.

Only widen to a fourth call type when all three of the originals run clean for a full week.

How to verify it worked

Check four things at the two-week mark, against the numbers you agreed before launch.

Containment. What proportion of calls the AI resolved without a human. For routine call types, production rates in the 70–80% range are a reasonable expectation. Treat claims of near-total automation across all call types with suspicion.

Abandonment. Whether callers are hanging up mid-conversation. A rise here cancels out any containment gain.

Escalation quality. Whether the humans receiving transfers say the context arrived intact. Ask them directly; the dashboard will not tell you.

Cost per resolved call. Not cost per minute. A platform that resolves 40% of calls is more expensive than a dearer one resolving 75%, because every unresolved call still costs you an agent.

If containment is below 50% on call types you chose for being simple, the problem is configuration, not the product.

Troubleshooting: six failures that actually happen

Callers talk over the agent. Latency. Under roughly 700 milliseconds feels like a conversation; above a second, people start interrupting. Ask the vendor for their measured round-trip latency, not their marketing number.

The AI books appointments into the wrong slots. Scheduling rules have not been configured, or the integration is reading availability rather than respecting constraints. Expect this in week one and check every booking daily for the first month.

Staff hang up on the agent. Nobody briefed them. Fixable in an afternoon, and entirely preventable.

It handles the easy calls and the queue gets harder. This is working as designed, not a fault. What reaches your humans is now the difficult residue. Adjust your handling-time expectations and your quality scoring accordingly, or you will penalise agents for the AI's success.

Containment looks great, satisfaction drops. The AI is closing calls it should escalate. Tighten the escalation triggers and re-read the transcripts for calls that ended without resolution.

The integration works in the demo and not in production. Almost always a permissions or sandbox-versus-live issue. Confirm during evaluation which environment the demo ran against.

The integration questions to ask every vendor

Ask these in writing, before the demo, and compare the answers side by side.

1. Does the agent write into our system during the call, or notify someone afterwards?

2. Which exact version or edition of that system do you support?

3. What is your concurrency limit at peak, and what happens at the limit?

4. Does a transfer carry the transcript and captured fields to the human?

5. Where are recordings and transcripts processed and stored, and which sub-processors touch them?

6. What is your measured latency, and how was it measured?

Question one eliminates more vendors than the other five combined.

What "integrates with Salesforce" actually means

Three different things get called an integration, and the word does not distinguish them. Establish which one you are being sold.

CRM integration means the agent reads a customer record and writes the call outcome back to it. This is what most buyers picture. The questions that matter are whether it writes during the call or after, which objects it can create, and whether your custom fields are reachable. A vendor supporting Salesforce Sales Cloud does not necessarily support Salesforce Financial Services Cloud, and the same caveat applies to Microsoft Dynamics 365 editions.

Telephony integration means the agent can answer on your existing numbers without you porting anything. This is usually SIP or number forwarding, and it is the integration that determines whether this is a two-week project or a platform migration. It is also the one most often assumed rather than confirmed.

Scheduling integration means the agent reads live availability and creates the booking itself. Reading a calendar is not the same as writing to it, and only the write version removes work from your team. Brilo publishes connections to cal.com and to Salesforce, HubSpot, Zoho, Zendesk and Microsoft Dynamics 365 among over 6,000 apps. Source: brilo.ai, read 22 September 2026.

A vendor may do one of these three well and the other two not at all. Ask about each separately, and ask which specific edition and version.

What to do next

Once your three call types run clean, the sequence is: add a fourth, then move from after-hours to overflow, then to the main line during peak hours only. Full-time main-line coverage is the last step, not the first.

If you are still choosing a platform, the comparison of AI agents for call centres and the enterprise contact centre platforms roundup both cover the shortlist. For the surrounding work that is not the call itself, see call centre automation. Teams evaluating a cloud deployment specifically should read cloud-based call centres, and smaller operations should start with the smaller teams guide.

Brilo is an AI voice agent that answers and places phone calls, and advertises deployment within ten minutes. It is not a contact centre platform: if you need omnichannel routing, workforce management or a full reporting stack, the platforms above are the right category. Source: brilo.ai, read 22 September 2026.