Most “AI for call centres” pitches lead with the wrong number. Average handle time looks like the metric to chase, but it is downstream of three things that AI can move directly: how calls reach the right agent, how much post-call work an agent has to do alone, and how quickly a new agent gets to competent. When teams report disappointing results from contact-centre AI, the cause is almost always that the deployment touched the surface (a chatbot on the website) without touching the routing logic, the wrap-up workflow, or the agent-assist surface. Call centres in 2024 are a mix of voice, chat, email, and social-media tickets handled through software that pretends they are one queue. The technology stack underneath — automatic call distribution, predictive dialers, CRM lookups, knowledge bases, quality monitoring — has been around for two decades. What changed in the last few years is that the routing decisions, the post-call summaries, and the real-time prompts shown to agents can now lean on language models that read the conversation as it unfolds. That is the part worth examining carefully. What does AI actually change inside a call centre? The honest answer is: four workflows, and the gains are mostly in the seams between them, not in any single one. The first is intent classification at the front of the call. Older systems route on IVR menu selections (“press 1 for billing”). Modern systems transcribe the opening seconds of speech, classify intent against a taxonomy specific to the business, and skip the menu entirely when confidence is high enough. The win is measurable — first-contact resolution rises when the routing decision uses what the caller said rather than what they pressed — but only if the intent taxonomy is maintained. A stale taxonomy is worse than a menu. The second is real-time agent assist. As the conversation proceeds, a model surfaces the relevant knowledge-base article, the customer’s recent ticket history, or a suggested response. This is the surface that most directly compresses handle time for complex calls, and it is also the one most likely to be ignored by experienced agents who already know the answer. The structural fix is to make the assist surface auxiliary — visible but not blocking — and to instrument whether agents actually consulted it. The third is automated post-call work. Summarisation, disposition coding, and CRM updates have historically eaten three to six minutes per call. A language model that listens to the call and drafts the summary, then lets the agent edit and submit, removes most of that. We see this pattern regularly in contact-centre modernisation work: the automation that earns its keep is rarely the customer-facing one — it is the after-call wrap that nobody outside the building notices. The fourth is sentiment and quality monitoring at scale. Traditional QA samples one or two per cent of calls. A pipeline that transcribes everything and scores against rubric criteria covers the full population, which changes what a QA team can actually do — from spot-checking to identifying systemic issues across thousands of conversations. Where the routing decision lives Call routing has become the test case for whether a contact centre will see real efficiency gains. The decision is not “use AI for routing” but “what data does the routing system have access to at the moment of decision?” A useful way to think about it: Routing input What it enables What it requires Caller phone number + CRM lookup Match to existing customer record, previous agent, open tickets CRM integration, deduplicated contact records Speech-to-text of opening 5–10 seconds Skip IVR menu, classify intent directly Real-time STT (Whisper, Google Speech-to-Text, AWS Transcribe), maintained intent taxonomy Conversation context mid-call Re-route to specialist or supervisor without restarting Bidirectional transfer with context, agent-side handoff UI Predicted call complexity Send to senior agent for likely-difficult issues Historical training data, retraining cadence The point of the table is not that all four are required. The point is that each new input creates a routing decision that the simpler tier could not make, and each input has an operational cost — integration, retraining, monitoring — that has to be honest in the planning. Predictive dialers in outbound centres face a similar layering. Naive dialers call from a list. Better dialers model when each customer is most likely to answer based on past contact attempts. Modern ones also predict which agents are most likely to close which leads, and pair them. Each layer requires more data than the previous one, and the data quality determines whether the system helps or hurts. Why does agent assist underperform in deployments? A pattern we see across engagements: the agent-assist surface is technically working — it surfaces relevant articles in under a second, it suggests reasonable next steps — and the metrics do not move. Three reasons recur. The first is that experienced agents are faster than the system. If the model surfaces an answer 800 milliseconds after the customer’s question, and the agent already typed the response, the assist did nothing. The deployments that work measure assist consultation as a separate metric and concentrate on the agents who actually benefit — new hires and agents handling unfamiliar call types. The second is that the suggestions arrive in a UI that competes with the CRM, the dialer, the chat window, and the knowledge base. Visual real estate matters. An assist panel that requires the agent to click away from the customer record to read it will not be used during a live call. The fix is layout, not modelling. The third is that the knowledge base the assist draws from is out of date. The language model is only as accurate as the corpus it retrieves from. If the knowledge base has not been pruned in eighteen months, the assist will confidently surface stale procedures. This is a content-operations problem dressed as an AI problem. The cloud-versus-on-premise question For most contact centres above a hundred seats, the architecture decision is largely settled: cloud platforms (Genesys Cloud, Amazon Connect, Five9, Twilio Flex, NICE CXone) have become the default because they integrate with the speech and language services those same providers offer. The remaining on-premise deployments are usually constrained by data residency, regulatory environment, or a legacy telephony asset that has not amortised. The trade-off worth naming honestly: cloud platforms make experimentation cheap. Spinning up a new routing flow, a new dispatch rule, or a new sentiment threshold is a configuration change rather than an infrastructure project. That accelerates the feedback loop that makes AI deployments work. On-premise environments can match the capability, but the iteration speed is structurally slower, and most of the contact-centre AI literature assumes the cloud cadence. For virtual and hybrid centres — where agents work from home offices and the “centre” is logical rather than physical — cloud is effectively a requirement. The orchestration of presence, queue assignment, and supervisor oversight across distributed agents does not have a serious on-premise solution at scale. What to measure before declaring success The efficiency gains from contact-centre AI are real but uneven. Before scaling a pilot, the metrics worth pinning down — with baselines from before the deployment — are: First-contact resolution rate by intent category. If routing improved, this rises for the categories where intent classification is reliable. Average wrap-up time per agent. If post-call automation works, this falls by one to three minutes per call without quality degradation. Time-to-competence for new agents. If real-time assist helps the population it should help, new hires reach target handle times faster than the pre-deployment cohort. QA coverage. If full-call transcription and rubric scoring are in place, coverage moves from a sampling percentage to one hundred per cent, which changes which problems are visible. Customer effort score on the post-call survey. This is the one that catches deployments where efficiency went up and customer experience went down. A deployment that improves handle time while degrading customer effort score is not a successful efficiency gain — it is a cost reduction the customer noticed. The metrics have to move together to count. Engineering reality The pieces that make contact-centre AI work in production are unglamorous: a maintained intent taxonomy, a curated knowledge base, real CRM integration, an agent-assist UI that respects screen real estate, and instrumentation that distinguishes which agents benefit from which surfaces. The models themselves — speech-to-text, summarisation, sentiment — are largely commoditised through cloud APIs at this point. The differentiation lives in the integration work and in the operational discipline around the data those models consume. Our work in this space focuses on the integration layer: connecting routing decisions to CRM history, building the agent-assist surfaces that fit alongside existing tooling, and instrumenting the metrics that reveal whether each component is earning its keep. The failure mode we see most often is that a contact centre buys a vendor’s AI module, switches it on, and waits for numbers to move — without addressing the routing taxonomy, the knowledge-base hygiene, or the agent workflow that determines whether the module gets used at all. For background on the customer-facing layer, see our broader treatment of how AI chatbots are transforming customer service across industries. The automation patterns that work in call centres rhyme with those in other operationally dense environments — there is a parallel worth drawing with the future of automation in construction, where the gains also live in the seams between specialised systems rather than in any single tool. Frequently Asked Questions Does AI in a call centre replace human agents? In practice, no — the deployments that produce durable efficiency gains keep human agents on the complex calls and use AI to absorb routine queries, draft post-call summaries, and surface relevant context during conversations. Replacing agents wholesale tends to surface in customer effort scores within weeks. Which call-centre AI improvement gives the fastest payback? Automated post-call summarisation is usually the quickest win, because it removes one to three minutes of wrap-up work per call without touching the customer-facing experience. Routing improvements take longer because they depend on a maintained intent taxonomy. How does AI handle multilingual customer support? Modern speech-to-text and translation pipelines (Whisper, Google Speech-to-Text, AWS Translate) cover dozens of languages well enough for routing and summarisation. Real-time agent-to-customer translation during a live call is harder and still imperfect — the deployments that work usually route the call to a native-speaking agent rather than translating mid-conversation. Is cloud contact-centre software necessary for AI features? Not strictly, but cloud platforms (Genesys Cloud, Amazon Connect, Five9, NICE CXone) make AI integration substantially faster because the speech, language, and analytics services come from the same providers. On-premise environments can match the capability but iterate more slowly. What is the biggest risk when deploying AI in a call centre? Degrading customer experience while improving internal efficiency metrics. A deployment that cuts handle time but raises customer effort score has shifted cost onto the customer — the metrics have to move in the same direction for the gain to be real.
