Connect Microsoft Teams Phone to an AI Voice Agent

Connect Microsoft Teams Phone to an AI Voice Agent

Yes, you can connect an AI voice agent to your Microsoft Teams Phone system without changing Teams or porting your numbers. The path is Direct Routing: the AI agent attaches as its own SIP endpoint to the same Session Border Controller (SBC) or carrier trunk that Teams already uses. Nothing changes for your staff — they keep working in the familiar Teams client.

Here is the point most projects discover too late: there is no Teams plugin and no Microsoft Store app that drops an AI phone assistant into Teams. The integration happens one layer down, on the telephony infrastructure. Knowing that saves weeks of dead ends. This guide covers which of the four Teams PSTN models actually works, what the architecture looks like in practice, the three SIP details that routinely break deployments, and how to go live in about a week.

Why Teams Phone and AI agents belong together

Microsoft Teams Phone has become the phone system in a large share of mid-market companies. The legacy PBX is gone, auto attendants and call queues run in the Microsoft cloud. That solves internal communication well. What it does not solve: the call that arrives at 6:40 p.m., when the call queue has nothing left to offer but voicemail.

That is where the gap opens. An auto attendant can read a menu and transfer. It cannot write an appointment into your calendar, look up a customer record in your CRM, answer a question about a delivery status, or qualify a number for a callback the next morning. An AI voice agent can — it understands free speech, reaches into your systems, and completes actions.

The obvious question follows: do I have to replace Teams? No. Teams remains the phone system for your people. The AI agent takes over defined numbers or time windows. Both share the same telephony infrastructure.

The four Teams PSTN models — and which one works

Microsoft offers four ways to connect Teams Phone to the public telephone network. Only one of them supports attaching an external AI voice agent. This distinction is the single most consequential decision in the whole project.

The reason is structural. With Calling Plan, Operator Connect, and Teams Phone Mobile, trunk control sits with Microsoft or the certified partner. You cannot register an additional endpoint there. With Direct Routing you operate the SBC yourself — or use a Direct-Routing-as-a-Service provider — and therefore decide which endpoints are allowed to take calls.

One practical note: Microsoft explicitly supports running multiple PSTN models in parallel within the same tenant. Your 40 inside-sales users can stay on Operator Connect while you stand up a Direct Routing trunk for the service hotline where the AI agent lives. You do not need to migrate your entire telephony estate to get started.

What the architecture actually looks like

There are three clean topologies. Which one fits depends on who should see the call first.

Option A — The AI agent as its own number

You dedicate an extension or service number to the agent. The SBC routes that number straight to Famulor instead of Teams. Calls to every other number continue to Teams unchanged. This is the lowest-risk option and the usual entry point: you test on a number nobody considers business-critical, then expand.

Option B — The AI agent as overflow

The call hits Teams first. If nobody answers within 20 seconds, or the call queue is empty, the SBC hands off to the AI agent instead of voicemail. This option rescues the calls that go unanswered today — for most companies the economically significant case. A trades business with six field technicians loses calls between 7 a.m. and 7 p.m. simply because one person staffs the office.

Option C — The AI agent as a front filter

Every call reaches the agent first. It establishes intent, resolves standard questions on its own, and passes only the conversations that need a human into Teams — with context prepared. This option delivers the largest relief and demands the most configuration care. It suits teams that have already validated the agent through Option A or B over several weeks.

The three SIP details that break deployments

Microsoft's Direct Routing documentation is precise and unforgiving. Three requirements are hard, and all three are routinely missed.

1. FQDN instead of an IP address in the Contact header

Microsoft's SIP proxy inspects the Contact header on every inbound message — both OPTIONS and INVITE. If it carries an IP address instead of a fully qualified domain name, the connection is refused with 403 Forbidden. The FQDN must also match the Common Name or a Subject Alternative Name on the SBC certificate. Wildcard certificates are supported, but only one level deep: *.company.com matches sbc1.company.com, not sbc.dc1.company.com.

Two further details from the same specification: Direct Routing does not support SIPS URIs, and it rejects SIP requests that carry a Replaces header. Anyone migrating from a legacy PBX often has such headers configured for historical reasons.

2. REFER support determines transfer quality

Direct Routing implements two transfer methods. If your SBC advertises REFER in its ALLOW header, Microsoft's SIP proxy acts as the transferor and lets the SBC handle the transfer — Microsoft's preferred behavior. Without REFER, the proxy terminates the transfer itself and issues a new INVITE.

In practice: if your AI agent should warm-transfer a call to a Teams user — announcing the caller and handing over context rather than blind-transferring — you need REFER on the SBC. Without it transfers still work, but less cleanly. Combined with media bypass they do not work at all: Microsoft supports call transfer in media bypass mode only when the SBC handles REFER. For defining the handover point properly on the business side, see Famulor's call transfer capability.

3. Media bypass and SDP size

Media bypass shortens the media path by letting audio flow directly between the SBC and the Teams client instead of traversing Microsoft's media processors. That reduces latency — a tangible factor with AI agents, because network latency and model latency stack. The cost is the stricter transfer requirement described above.

There is also a quiet trap: Direct Routing messages can exceed 1,500 bytes, usually because of large SDP payloads carrying ICE candidates. If a UDP trunk sits behind your SBC, it may drop the message. Microsoft recommends stripping unused codecs and ICE candidates from the SDP on the SBC before forwarding to the UDP trunk. If you are still weighing which transport suits your setup, the comparison WebRTC vs SIP for AI voice agents covers the trade-offs.

Step-by-step implementation

A realistic timeline for a mid-market company runs five to ten working days — spread out, not consecutive.

1. Take inventory. Establish three facts: which PSTN model runs today, who operates the SBC (you, your integrator, or a DRaaS provider), and which numbers sit on which trunk. Without those answers, everything downstream is guesswork.
2. Choose the target architecture. Pick Option A, B, or C. When in doubt start with A or B — reverting is a routing rule, not a project.
3. Stand up the trunk to Famulor. Famulor supports SIP trunking and BYOC for any VoIP or PBX provider. You keep your carrier, your numbers, and your rates. If the fundamentals are unclear, the SIP trunking glossary entry clarifies the terms.
4. Write the SBC rules. Define which numbers or conditions route to the agent. While you are there, verify the Contact header, certificate match, and REFER support.
5. Configure the agent's substance. Only now does content matter: greeting, conversation design, handover rules. Load your domain knowledge into a knowledge base so the agent answers from your documents instead of improvising.
6. Connect your systems. Calendar, CRM, ticketing, ERP. Through 300+ integrations you can wire Microsoft 365, Teams, HubSpot, Pipedrive, Google Calendar, Salesforce, and Zendesk without writing code. For Teams-specific notifications and workflows there is a dedicated Microsoft Teams integration.
7. Run a live test. One week, one number, daily transcript review. Expect to tune — that is normal, not a failure signal.
8. Scale. More numbers, more time windows, more languages. Going from one number to twenty is configuration, not a new integration.

Best practices and common mistakes

Start small, but real. The most common mistake is an oversized first step: every number, every intent, immediately. The second most common is an undersized one — a test with three internal calls that tells you nothing about daily reality. Take a real number with real traffic, but only one.

Define handover criteria up front. Write down when the agent transfers to a human — as a list, not a feeling. Complaints, cancellations, amounts above a threshold, an explicit request for a person. Anything not on the list, the agent handles itself.

Measure latency, do not estimate it. Perceived call quality is decided in milliseconds. Evaluate whether media bypass makes sense for your topology, and measure real response time under load rather than idle.

Test the fallback path before you need it. What happens if the trunk fails or the agent does not answer? Configure a fallback rule on the SBC to return calls to Teams — and test it actively instead of trusting it.

Set caller ID deliberately. Outbound calls from the agent should display your recognized main number, not a technical extension. Otherwise answer rates drop noticeably and callbacks go nowhere.

Do not assume one language. If you serve international customers or multilingual sites, the agent detects the language at the start of the call and switches. Famulor supports 40+ languages — a factor frequently underestimated in Teams environments spanning several countries.

Industry examples from practice

IT services provider, 45 staff, Teams Phone via Direct Routing. The support hotline forwards to the AI agent after 5 p.m. The agent captures incidents, establishes customer, system, and severity, opens a ticket, and calls the on-call engineer for priority-one cases. Everything below that is waiting the next morning as structured tickets instead of six voicemails.

Accounting firm with 12 partners. Between February and May the front desk is overwhelmed by call volume. The agent runs as a front filter: it answers questions about deadlines and required documents from the knowledge base, books appointments in the Outlook calendar, and transfers only substantive questions to the partners — with client name and topic already posted in Teams chat.

Machinery manufacturer with three sites in Central Europe and one in Poland. The central service number sits on a Direct Routing trunk. The agent detects the language, captures service cases in German, English, and Polish, and routes by machine type to the correct Teams call queue at the responsible site.

Trades business with 20 employees, 14 of them in the field. The office administrator cannot catch every call during the day. The agent runs as overflow after 20 seconds, qualifies requests by trade and urgency, escalates emergencies straight to the on-call mobile, and writes everything else into the CRM as a callback list.

Putting the cost in perspective

With Direct Routing you continue paying your existing telephony costs to your carrier — the AI agent changes nothing there. On top come the voice AI platform costs, typically usage-based per conversation minute. Because BYOC lets you bring your own trunk, you are not paying for telephony twice through the platform — a difference that grows meaningful at higher volumes.

The honest calculation weighs three items: recurring platform cost, one-time setup effort, and the current value of the calls that go unanswered. The third item is almost always underestimated, because it appears in no report. Companies quantifying it for the first time usually land on a number that puts the other two in perspective. The pricing page shows the available models.

Conclusion

Microsoft Teams Phone and an AI voice agent are not alternatives — they are two layers of the same telephony estate. Teams remains your employees' workplace. The AI agent takes the calls nobody there is free to take: after hours, on overflow, in pre-qualification.

The technical prerequisite is narrow and unambiguous: Direct Routing. If you already run it, connecting an agent is a matter of SBC rules and configuration, not migration. If you sit on Operator Connect or Calling Plan, add a separate Direct Routing trunk for your service numbers and leave everything else untouched.

Famulor is built for exactly this scenario: SIP trunking and BYOC for any carrier, no-code flows instead of a development project, 300+ integrations for calendar, CRM, and ticketing, and 40+ languages for international sites. You keep your numbers, your carrier, and your Teams environment.

Your next step: spend ten minutes confirming which PSTN model your tenant runs today — in the Teams Admin Center under "Voice". If it says Direct Routing, you can start this week with a test number. If it says something else, talk to your carrier about an additional trunk. Both routes reach the same destination; only the timeline differs. If you are migrating from an existing solution, the voice AI migration cutover plan lays out the sequence. For larger environments, see the requirements under contact center.

FAQ

Can I install an AI voice agent directly inside Microsoft Teams?

Not as an app or plugin. The connection runs through Direct Routing at the SIP layer: the agent attaches as its own endpoint to the same SBC as Teams. Users never see the difference.

Which Teams PSTN model do I need?

Direct Routing. Microsoft Calling Plan, Operator Connect, and Teams Phone Mobile do not allow custom SIP endpoints, because Microsoft or the carrier controls the trunk.

Do I have to port my phone numbers?

No. With BYOC you keep your carrier and your numbers. You simply define on the SBC which numbers route to the AI agent instead of Teams.

Can the AI agent transfer to a Teams user?

Yes, via SIP transfer back into a Teams call queue or to an extension. Clean warm transfers require your SBC to support the REFER method — in media bypass mode it is mandatory.

Can Teams and the AI agent run in parallel?

Yes. Microsoft supports multiple PSTN models in one tenant. Your staff can stay on Operator Connect while the service number runs over a Direct Routing trunk.

Why is my SBC connection refused with 403 Forbidden?

Usually the Contact header carries an IP address instead of an FQDN, or the FQDN does not match the Common Name or Subject Alternative Name on the SBC certificate. Direct Routing requires both.

How long does setup take?

With Direct Routing already in place, typically five to ten working days — SBC rules, agent configuration, system connections, and one test week with real traffic.

Does this also work with 3CX, NFON, or other PBX systems?

Yes. BYOC and SIP trunking are provider-agnostic. For common systems such as 3CX the principle is identical: your own trunk, your own routing rule.

What happens if the connection fails?

You configure a fallback rule on the SBC that returns calls to Teams or voicemail when no answer arrives. Test that rule actively before going live.

Can the agent also make outbound calls?

Yes. The same trunk carries callbacks, appointment reminders, and campaigns outbound — displaying your familiar main number as caller ID.