Voice AI in Teams, without the marketing fog.

1. What Voice AI in Teams actually means

"Voice AI" collapses three quite different capabilities into a single phrase, and most procurement conversations are confused before they start because nobody insists on separating them. Pulled apart, voice AI in a Microsoft Teams estate covers:

• Productivity AI on internal calls and meetings — transcription, summarisation, action capture and search. Microsoft Copilot for Teams is the obvious example, working over the Teams meeting and call experience.
• Voicebots and Intelligent Virtual Agents (IVAs) — automated voice answering PSTN calls, replacing or fronting an IVR. The caller talks; the bot listens, decides and acts; complex calls are handed to a human with full context.
• Agent-assist in the contact centre — real-time transcription, knowledge retrieval, next-best-action suggestions, automated wrap-up and quality monitoring across every call, not the 2% sampled today.

These three need different architectures, different data flows, different licensing and different governance. Treating them as one programme is the first mistake. Treating each as its own initiative, but on a shared voice foundation, is the pattern that works.

2. Where the AI sits in the call path

Voice AI is fundamentally a question of where the audio goes and what is allowed to listen to it. Three architectures matter in a Teams estate.

In-cloud, inside Teams

Copilot transcription and summarisation operate inside the Teams service itself. There is no separate media path: the audio is already in Microsoft 365, the AI runs against it there, and outputs land in Teams chat, OneNote or Outlook depending on the feature. From an architecture perspective this is the simplest pattern — no SBC work, no integration — but it is also the most constrained: you get the capabilities Microsoft ships, on the data Microsoft already holds.

In-stream, via the SBC

For voicebots and agent-assist on PSTN calls, the AI needs access to the live media stream. The cleanest way to do this is to fork the call at the SBC using SIPREC, sending a duplicate stream to a Voice AI gateway. The original call continues on Teams unaffected; the AI runs in parallel and either talks back to the caller (voicebot) or whispers to the agent (assist). Because the fork happens at the SBC, the estate keeps a single point of policy enforcement: media encryption, recording, fraud controls, emergency routing.

Bot-as-endpoint, via the Teams calling APIs

A third pattern is to register the bot as a Teams calling endpoint using Microsoft’s Bot Framework and Graph calling APIs. The bot effectively becomes a user that can answer, transfer and host calls. This is powerful for internal scenarios but slower to evolve when the bot needs to interoperate with non-Microsoft platforms — which most enterprise IVAs do.

3. Native Microsoft capabilities

Before adding any third-party voice AI, it is worth knowing what Microsoft already provides on a properly licensed Teams estate, because for many use cases it is enough.

Copilot for Microsoft 365 in Teams

Copilot generates meeting transcription, summaries and follow-up actions; surfaces what a missed attendee needs to know; and answers natural-language questions across meetings and chats the user has access to. It requires Copilot licensing in addition to the Microsoft 365 entitlement and depends on a properly configured Teams environment with transcription enabled.

Azure AI Speech and Azure OpenAI

Underneath Copilot, the broader Azure AI stack — speech recognition, neural text-to-speech, translation and the OpenAI models on Azure — is the toolkit most enterprise Voice AI projects end up building against. Hosted in UK regions, integrable with Teams via the patterns above, and commercially predictable through Azure consumption.

Copilot Studio

Microsoft’s low-code agent builder. Good for internal bots and simpler customer journeys, particularly where the data and systems already sit in Microsoft 365 or Dynamics. The honest assessment for enterprise voice work is that it handles a useful slice of use cases, but complex contact-centre IVAs still typically use a specialist platform talking to voice via a gateway like VoiceAI Connect.

4. AudioCodes VoiceAI Connect

AudioCodes is one of the two SBC vendors most commonly deployed in UK Teams estates (alongside Ribbon and, less often at this scale, Oracle). Its Voice AI offering, VoiceAI Connect, sits next to the SBC and acts as a broker between the voice channel and a chosen bot or speech platform. It exists because telephony and conversational AI speak different protocols, and the translation is non-trivial.

What VoiceAI Connect does

• Terminates the SIP/RTP voice channel from Teams or the SBC, and presents it to the bot platform as a clean event stream.
• Brokers speech-to-text and text-to-speech to a choice of engines — Azure AI Speech, Google, AWS, Nuance, or alternatives — so the speech vendor is not locked to the bot vendor.
• Talks to the bot platform of choice — Microsoft Copilot Studio and Azure Bot Service, Google Dialogflow CX, Cognigy, Kore.ai, IBM watsonx, AWS Lex and others — over standard webhooks. The bot author writes business logic once; the channel layer handles voice.
• Handles the things voice cares about that text channels do not: barge-in, DTMF, transfer to a queue or human, music on hold, call recording control, secure capture for PCI.

Deployment options

VoiceAI Connect ships in two shapes: VoiceAI Connect Cloud, a managed service consumed by the minute, and VoiceAI Connect Enterprise, a self-hosted appliance or VM for organisations that need the gateway inside their own UK tenancy. The choice is driven by data residency, the desired commercial model and how much the customer wants to own the integration plumbing.

5. Use cases that pay back

The voice AI deployments that justify themselves in a 12-month review tend to fall into a small number of patterns.

IVR replacement and call deflection

The classic entry point: replace a tree of touch-tone menus with a voicebot that asks "how can I help?" and routes — or resolves — based on the answer. Handled well it cuts average handling time, lifts containment for routine queries (balance checks, opening hours, appointment moves, password resets) and frees human agents for the calls that actually need them. Handled badly it just adds a layer of frustration before the caller eventually mashes 0 for an agent.

After-hours and overflow cover

For services that don’t justify 24/7 staffing but can’t leave callers hanging, an IVA can take ownership of basic enquiries, log requests into the CRM or service management tool, and promise a callback the next working day. NHS Trusts, universities during clearing and councils during severe weather all see disproportionate value here.

Agent-assist on every call

Real-time transcription of the live call is fed into a sidebar in the agent’s desktop, with knowledge-base answers surfaced as the caller talks and a draft wrap-up note populated by the time the call ends. Agents spend less time searching and writing; supervisors gain a 100% sample of calls for quality monitoring rather than the conventional 1–3%.

Outbound voice notifications

Less glamorous but high-value: appointment reminders, collection notices, service updates, with the option to press 1 to confirm or 2 to reschedule into a live agent queue. Voicebots do this at a fraction of the cost of agent time and at a scale humans cannot match during peaks.

Internal Copilot adoption

Quietly, the biggest hours-saved metric in many estates is simply staff using Copilot transcription and summarisation in their day-to-day Teams meetings. It is the easiest voice AI to turn on, and the only one where the licensing model is straightforward.

6. Voice AI in the contact centre

Voice AI and contact centre are not the same thing, but they collide constantly. The pragmatic position for a Teams estate with a Teams-native CCaaS layer (Anywhere365, Luware, Landis, Tendfor, NICE CXone for Teams, Genesys Cloud for Teams) is:

• Let the CCaaS platform own routing, agent state, recording and reporting. That is what it is for.
• Use a Voice AI gateway — VoiceAI Connect is the obvious one — to insert voicebots in front of the queue and agent-assist alongside the agent. The CCaaS and the AI layer integrate at well-defined points: data dip from the bot into the CRM, hand-off from bot to queue with full context, transcript and intent passed into the agent desktop.
• Resist the temptation to let any one vendor own the whole stack — voice, AI, CCaaS, CRM. The platforms that win today are rarely the platforms that win in five years, and swappability matters.

7. Data, privacy and UK compliance

Voice calls are personal data the moment they contain a caller’s voice, name, account reference or health information — which is most of them. Voice AI processes that data in flight, often through third-party engines, and the governance work is non-trivial. The questions to settle before scaling are:

• Data residency. Where does the speech recognition run, where do transcripts and recordings sit, and which Microsoft Azure regions (or equivalent) is the bot platform hosted in. For UK public sector and most regulated industries the answer needs to be UK or, at minimum, EU with documented controls.
• Retention and deletion. Transcripts created by Copilot, by the agent-assist platform and by the bot are all separate stores. Each needs a retention policy aligned to the records-management schedule, and a deletion path that survives a subject access request.
• Consent and notification. Callers need to be told that AI is processing the call, in language that would survive a complaint to the ICO. "Calls are recorded for training and quality purposes" was borderline already; it is not enough for a fully AI-processed call.
• PCI and sensitive capture. Where card details or other sensitive data are spoken, the bot or gateway must support secure capture — muting recording and transcription during the sensitive segment, tokenising the captured value — or the call must transfer to a payment IVR designed for it. AudioCodes VoiceAI Connect, and competitive gateways, support this pattern; not every direct bot integration does.
• Model training. What, if anything, is the speech vendor or LLM allowed to learn from the audio? The answer in regulated environments is almost always "nothing", and the contract needs to say so explicitly.

8. Measuring success

Voice AI programmes get into trouble when nobody agrees what success looks like before procurement closes. The metrics that hold up over time are unglamorous and measurable.

• Containment rate — the percentage of bot calls fully resolved without human transfer, by intent. Track per intent, not as a single number; the average hides everything.
• Handle time delta — average handle time on calls that did transfer, compared with a control group not using agent-assist. Time-to-first-knowledge-suggestion is a useful sub-metric.
• First-call resolution on bot-handled versus agent-handled enquiries. Bots that contain calls but leave the customer calling back the next day are not a win.
• CSAT or equivalent, sampled per intent, for bot-only journeys. Many estates see CSAT improve for simple intents and dip for complex ones — the right response is tighter routing, not a panicked rollback.
• Agent sentiment. Agent-assist that helps gets used and improves; agent-assist that distracts gets switched off as soon as supervisors stop looking.

9. Common mistakes

• Treating voice AI as a single procurement when it is three programmes (Copilot productivity, voicebots, agent-assist) with different sponsors, governance and timelines.
• Choosing the bot platform before the voice architecture. By the time the integration team picks it up, the channel decisions have already constrained what is achievable.
• Skipping the gateway layer. A direct bot-to-SBC integration looks elegant in a PoC and becomes a millstone the first time the bot vendor changes its API or the speech vendor raises prices.
• Underestimating the conversation design effort. The technology is the smaller half of the project; getting the prompts, fallbacks, hand-offs and tone right is the larger half and is often handed to whoever happens to be free.
• Skipping UK data-residency due diligence on the speech and LLM tier. Most public-sector projects have to redo this under audit pressure if it wasn’t front-loaded.
• Measuring containment without measuring CSAT and repeat calls. A bot that successfully refuses to transfer is not the same as a bot that successfully resolves the query.
• Letting the contact-centre vendor own the AI roadmap. CCaaS vendors have a legitimate interest in being the centre of gravity; the customer’s interest is usually in keeping options open.

10. Sector scenarios

Three illustrative deployments, drawn from the kinds of work UK organisations of 800+ and 1,000+ users typically bring to us. Anonymised; sectors and shapes are representative rather than specific.

Sector scenario: NHS Trust patient access line

Context. An acute NHS Trust’s outpatient booking line handling hundreds of thousands of calls a year, with predictable peaks and persistent abandon rates the Trust had been unable to staff out of.

Approach. A voicebot fronted the patient access line, taking appointment confirmations, cancellations and reschedule requests with a clear and immediate "transfer me to a person" path. Architecture: SBC forks the call to AudioCodes VoiceAI Connect, which brokered to a healthcare-specialist IVA platform and Azure AI Speech in a UK region. Hand-off to a Teams-native CCaaS queue with full transcript and intent for the cases that needed a human. Recording governance unchanged from the existing Trust policy; consent script rewritten with information governance.

Outcome. A material share of routine interactions handled end-to-end by the bot during business hours, with abandon rates falling sharply in the first peak after go-live. Patient feedback on the bot itself was mixed and improved over the first six months as the conversation design was iterated.

Sector scenario: Russell Group university clearing

Context. A university running an annual clearing operation that triples inbound call volume for 72 hours, with an established Teams voice estate and a small permanent contact-centre team.

Approach. An outbound and inbound voicebot for course-availability enquiries, with hand-off to human advisers for live offers. Built on Copilot Studio for the simpler journeys and on a specialist IVA for the offer conversation, both fronted by VoiceAI Connect. Agent-assist deployed on the human queue so that temporary clearing staff, hired weeks before the event, had a knowledge layer instead of a printed crib sheet.

Outcome. Clearing run completed without the usual recourse to outsourced overflow, with hold times shorter than the previous year and a recording of every offer conversation available for compliance and dispute resolution.

Sector scenario: County Council customer services

Context. A County Council with general customer services receiving repetitive enquiries on bins, council tax, parking and pothole reporting, with a small specialist team handling adult social care behind a separate queue.

Approach. A general-purpose voicebot at the front door, integrated to the council’s case management system, designed to take a complete fix-my-thing request end-to-end where possible (a pothole report, a missed-collection log) and to triage everything else to the right queue with context. Adult social care explicitly out of scope for the bot; every call into that queue continued to be answered by a human, with agent-assist transcribing and supporting the call for quality and supervision purposes.

Outcome. Demand into the general queue reduced as the bot took ownership of fully-automatable requests; specialist teams freed up to focus on the calls that genuinely needed them. Critically, the human-only path for vulnerable callers was preserved and visibly so — the council was able to explain its design to scrutiny committee without difficulty.