A medical clinic in Brossard rolled out its first AI voice agent in December 2025. One single agent, trained to do it all: reception, triage, booking, emergency transfer, no-show billing. Six weeks later, the director calls me. "We pulled the plug. Patients give up mid-call." The issue? The agent did everything — but never really well. Fuzzy transfers, missed follow-ups, a billing flow that kept triggering the greeting script on loop.
That's 2026 in one anecdote. Seven out of ten Quebec SMBs unplug their AI voice agent within 90 days, and the architecture chosen on day one is a big part of the reason. Since ElevenLabs' April 2026 update (multi-agent tracking, workflow tool overrides) and OpenAI's general availability of gpt-realtime with native SIP, the question has shifted. It's no longer "should I deploy an AI voice agent?" It's "should I deploy one, or several that talk to each other?"
The two architectures, no jargon
A single-agent setup is one conversational brain that takes the call, understands, responds, runs tools, and hangs up. One personality, one system prompt, one tool catalog. When a patient calls to cancel an appointment, it's the same agent that would have handled a booking thirty seconds earlier.
A multi-agent architecture is several specialized agents that hand calls off to each other via explicit rules. An intake agent that qualifies the intent, a "booking" agent talking to the calendar, a "billing" agent talking to Stripe or your accounting system, an "escalation" agent that holds the patient's hand while the human joins. Each has its own prompt, its own tools, its own limits.
Since April 2026, the ElevenLabs platform exposes a visited_agents field in the conversation response: for the first time, you can see which sub-agents handled a call and in what order. This isn't cosmetic. It changes the whole Bill 25 audit story — I'll come back to that.
What the research says (and what it doesn't)
Google Research published benchmarks in 2026 on agentic architectures that should temper the "multi-agents for everything" crowd: multi-agent coordination delivers a +81% measured gain on parallelizable tasks, but degradation of up to -70% on sequential ones. Translation for an SMB: a phone call is, by definition, sequential. A human speaks, the agent listens, responds, waits. Every handoff between sub-agents adds latency and a context-break risk.
Microsoft, in its Cloud Adoption Framework, proposes a simple threshold: a single agent works up to 3 to 5 distinct functions. Beyond that — or when functions cross a security boundary (e.g., an agent that touches medical data should not be the same one that reads the public calendar) — move to multi-agents.
That threshold lands well for a Quebec SMB. In practice, a clinic has 4 to 7 intents: booking, cancellation, address change, urgency request, billing, results request, administrative call. You're sitting just above the threshold.
Head-to-head with numbers
Here's how the two stack up on the dimensions SMB leaders actually care about:
- Time to deploy. Single-agent: 2 to 4 weeks of configuration and testing. Multi-agent: 5 to 9 weeks, mainly due to handoff design and transfer logic.
- Perceived latency. Single-agent on gpt-realtime: roughly 600 to 900 ms between the end of the caller's sentence and the start of the response. Multi-agent: each handoff costs an extra 200 to 500 ms depending on platform.
- Successful barge-in rate (when the caller interrupts). Single-agent: ~92% in Quebec French based on our internal measurements. Multi-agent: drops to ~84% during transfer windows.
- Monthly cost for 1500 calls/month. Single-agent: CA$450 to $800. Multi-agent: CA$650 to $1100, because of extra processing time and additional LLM calls at each handoff.
- Bill 25 audit complexity. Single-agent: one log, one retention policy. Multi-agent: one log per agent, plus a matrix showing who accessed what data. More rigorous — but also more costly to maintain.
- First-contact resolution rate. Single-agent: 61% on average across our 2026 Quebec deployments. Well-designed multi-agent: 78%. Badly-designed: 52%. The variance is huge, and that's the central point.
That last number deserves a pause. A single agent gives you a predictable result — mediocre, but predictable. A multi-agent architecture gives you either the best or the worst, depending on design quality. You're signing up for volatility.
When single-agent is the right call
Three cases where we advise TECHMA clients against multi-agent:
Restaurants and neighborhood shops. Intents are few: reservations, hours, menu, takeout. Three or four functions, well below the Microsoft threshold. A well-prompted single agent connected to OpenTable or Square resolves 75% of calls. Adding a second agent reduces fluidity without measurable gain.
SMBs in pilot phase. You're launching your first agent. You don't yet know your real call patterns. Starting multi-agent is like building a house on a theoretical blueprint before you've lived on the land. Start single-agent, listen to 200 to 500 real calls, then decide whether you need specialization.
Limited technical teams. Multi-agent demands engineering discipline: prompt versioning for each sub-agent, regression tests on transfers, monitoring of call paths. If your IT team is a manager and an intern, you'll lose your footing fast. We've watched multi-agent architectures abandoned because nobody remembered anymore why the "billing" agent was talking to the "emergency" agent.
When multi-agent earns its keep
Conversely, three scenarios where it's a game-changer:
Medical clinics and dental offices. The need to separate flows is real. An intake agent that never touches the medical record, a "booking" agent that only reads the calendar, an "emergency" agent that only has the on-call line. That separation isn't a luxury — it's a Bill 25 access-minimization requirement. ElevenLabs' new workflow tool overrides let you restrict tools and knowledge bases per sub-agent, which is exactly what this enforces.
Law firms and accounting practices. As we covered in our law-firm analysis, conflict-of-interest triage and confidentiality demand strict boundaries. The first-contact agent must have no client database access. Only a "validation" agent — gated by a conflict check — can enrich the file. Mono-agent is technically possible, but the audit becomes a nightmare.
Call volumes above 3000/month with 3+ cleanly distinct categories. Past that threshold, the benefit curve flips. The added multi-agent cost (30 to 40%) is absorbed by the first-pass resolution gain (+17 points on average). Our home-services clients (plumbers, multi-tech HVAC companies) cross that line by the second quarter of deployment.
What April 2026 actually changed
Three things make the debate more interesting than it was in 2025:
1. Native sub-agent tracking. ElevenLabs' visited_agents field gives you, for each call, the ordered list of agents touched plus their branch_id. Before, you had to stitch logs manually. Now, your compliance team pulls a Bill 25 report in 3 minutes.
2. Scoped evaluation. Evaluation criteria can now be scoped to a specific sub-agent. Concretely: you measure the "booking" agent quality independently from the "intake" agent. Weak links become visible. In 2025, you evaluated the whole conversation, which diluted the problems.
3. gpt-realtime with native SIP. OpenAI's general availability announcement includes SIP, meaning you can now route phone calls directly to the API without Twilio as a middleman. The 20% price cut ($32 / 1M input audio tokens) makes multi-agent architectures economically viable for SMBs that would have said no in 2025.
The Bill 25 angle nobody talks about
Bill 25 imposes data-collection minimization and fine-grained access logging for personal information. In a mono-agent setup, your minimization policy is theoretical: the agent has access to everything and "chooses" not to use it. The CAI (Quebec's privacy commission) doesn't love that argument.
In a well-split multi-agent setup, minimization is structural. The intake agent can't read the medical record — not because it chooses not to, but because the tool is simply not configured for it. That's the difference between "I promise I won't look" and "I don't have the key." In an audit, only the second holds up.
For Quebec SMBs in healthcare, finance, and legal services, that argument alone often justifies the multi-agent premium.
The pragmatic rule for deciding
We'll close with a simple rule that works for 90% of the SMBs who call us:
- Fewer than 3 clear intents, fewer than 800 calls/month, no sensitive data: single-agent.
- Between 3 and 5 intents, moderate volume, non-sensitive personal data: single-agent, well-structured, with conditional tools.
- More than 5 intents OR sensitive data (medical, financial, legal) OR volume above 2000/month: multi-agent architecture, but with a 4-to-6-week single-agent pilot first.
Above all: never start multi-agent without having operated a mono-agent first. You'll be optimizing handoffs for call patterns you haven't seen yet, and you'll get it wrong.
What to do this week
If you're evaluating your first deployment, list your real intents (not theoretical — real, based on your phone logs from the last 90 days). Count them. If the total is under 5, stop reading multi-agent literature — you don't need it.
If you already have an agent deployed and your resolution rate is stuck under 65%, look at where it fails. If failures cluster around 1 or 2 specific intents (typically: billing or emergency transfer), isolating those intents in dedicated sub-agents gives you maximum gain for minimum investment.
At TECHMA, we deploy an average of two multi-specialized agents per month, always after a single-agent pilot phase. If you want us to look at your specific case — volumes, intents, Bill 25 constraints — book a free diagnostic call. We'll tell you honestly whether you need one or five agents. Most of the time, the answer is "one done well, not five done poorly."