Introduction
Open your SMB's ElevenLabs dashboard on a Monday morning. 832 conversations from the past month. They're sitting there, lined up by date, each one with a transcript attached. But how many were appointment bookings? How many emergencies? How many complaints that should have been transferred to a human and weren't?
You don't know. Nobody knows. Because unless you open every call one by one, it's a tangled mess.
On May 7, 2026, ElevenLabs shipped a feature in SDK v2.46.0 that changes that: Conversation Tags. You label every call automatically (booking, emergency, complaint, pricing-question, etc.), then you filter. The mess turns into a dashboard.
This tutorial shows you how to set this up for your Quebec SMB in 15 minutes, and how to use it afterwards for a weekly audit that takes 12 to 15 minutes instead of hours.
What you need before starting
Three things, that's it.
One: an ElevenLabs account with a voice agent already deployed and taking real calls in production. If you're not in production yet, run your tests first — tag value only shows up once you have a few dozen calls.
Two: admin access to the ElevenLabs dashboard. Creating tags is a protected action.
Three: 15 minutes of quiet time. No more. If you ever tried setting up manual call tagging in the past and spent two afternoons on it, forget that memory — the automatic version is radically faster to deploy.
Step 1 — Pick the 6 categories that actually matter for your SMB
Before you touch the dashboard, take a piece of paper. List the call types your AI voice agent actually handles. The temptation is to create 15 or 20 categories to cover everything. Don't. Six is the right number — enough to discriminate, few enough that a glance is all you need.
For a Quebec dental clinic, it looks like: new-booking, reschedule, dental-emergency, pricing-question, complaint, other. For a plumber, it's more like: leak-emergency, renovation-quote, new-booking, invoice-follow-up, complaint, other.
The rule: five tags specific to your trade, plus one catch-all 'other'. If 'other' goes above 20% of your calls after a month, that's the signal that one of your specific categories is missing. You'll add it then, not before.
Step 2 — Create the tags in the ElevenLabs dashboard
Head to Conversational AI → Conversations → Tags button in the top right. Click Create tag. For each category, give it a short kebab-case name (leak-emergency, not 'Leak Emergency'), pick a color (you'll use this for visual scanning later), and save.
For an SMB's six categories, budget 90 seconds. It's the fastest part of the tutorial.
One detail that saves time later: use the same vocabulary your human staff uses. If your receptionist says 'bookings' and not 'appointments', name the tag new-booking. When you filter, your brain will move faster.
Step 3 — Get the agent to assign tags on its own
Here's where the magic happens. You are NOT going to manually tag calls. You're going to configure the agent to do it at the end of every conversation.
In your agent's config, go to Analysis → Post-call analysis. Add a call_category field of type single-select, with possible values that match your six tag names exactly. In the field's description, write in plain language: 'Categorize this call based on the customer's primary intent. Pick new-booking if the customer wants a new appointment, leak-emergency if they're reporting an active leak, etc.'
Be precise in the description — two or three sentences per category. The LLM uses this text as its guide. This is where tag quality is won or lost.
Then — and this is the critical step — wire this analysis field to a webhook that calls ElevenLabs' Conversations API to assign the matching tag. The endpoint is POST /v1/convai/conversations/{conversation_id}/tags, documented on the ElevenLabs Conversations API page. If your SMB is a TECHMA client, this is the plumbing our team handles for you — you don't touch the code.
Step 4 — Test with 3 calls before letting it run
Don't deploy silently. Make three test calls from your cell phone, with three different intents. One as 'I want a booking', one as 'I have a complaint', one as 'how much does it cost'.
Wait 30 seconds after the last call (time for post-call analysis to run), then open the dashboard. Filter by tag. Your three conversations should be tagged correctly.
If one is off — often it's the 'pricing' call that ends up in 'new-booking' because the customer eventually said 'ok I'll book at that price' — go back to step 3 and refine the tag's description. Two or three iterations are usually enough. Same principle as with Keyterms, which we covered in our article on the parameter that actually saves your agent — precision comes from definition, not from the model.
Step 5 — Use the tags for a 15-minute weekly audit
Here's what your Monday morning looks like once tags are live.
You open the dashboard, filter on the last 7 days. You see: 192 calls — 68 new-booking, 41 emergency, 33 pricing-question, 18 complaint, 23 reschedule, 9 other. Readable. Counted.
You click on the complaint tag. 18 conversations. You skim the 18 transcripts — that takes 6 minutes. This is where you find the real problems: a customer who called three times about the same thing, another one the agent misunderstood.
You click on emergency. 41 conversations. You verify all of them triggered the human transfer correctly. If one didn't, you know exactly what to fix in the prompt this week.
Total: 12 to 15 minutes. You walk out of the audit with two or three concrete fixes to ship to your agent, based on real calls instead of impressions. That's exactly the kind of light ritual we describe in our 12-minute Monday ritual, except with tags you now have a category view on top of the four dashboards.
The 3 mistakes to avoid
Too many tags. Twenty categories paralyzes you. You stare at the dashboard and don't know where to start. Six, max. If you discover a missing one after a month, add it — but start small.
Ambiguous names. Avoid 'miscellaneous' next to 'other', or 'booking' next to 'new-booking'. Each tag should have a unique definition, readable in four words or fewer.
Never revisiting the tag description. The description you wrote in step 3 is what the LLM uses to decide. If after two weeks a tag is mis-assigned 15% of the time, it's the description you need to rework — not the model. According to Gartner research on customer interaction classification, the quality of descriptive categories is the number-one factor in a taxonomy's reliability.
What you'll see in 30 days
First effect, immediate: your Monday audit drops from 90 minutes (on the weeks you used to do it seriously) to 15 minutes. That's recovered time, straight up.
Second effect, around week 3: you spot a call category you hadn't seen coming. For a dental clinic we follow, it was 'patient who wants to switch practitioners within the same clinic' — 11% of monthly volume, never named before, never treated as a distinct flow. Once identified, we adjusted the agent to proactively offer the other practitioners' availability.
Third effect, around week 4: you walk into a team meeting with numbers. '33% of our calls are pricing-questions. Why? Is our website unclear?' That's a conversation that never happened before because nobody could quantify it.
FAQ
Can you put multiple tags on the same conversation? Yes — the Conversations API supports multiple tags. But for SMB usage, stick to one primary tag per call. You'll avoid double-counting in your statistics.
How much extra does it cost on ElevenLabs? Tagging itself is included in the plan. What burns a bit is the post-call analysis running the LLM. Budget the equivalent of 5 to 10 extra seconds of TTS per call — negligible at SMB scale.
What if we change the tag list in 3 months? No problem. Old conversations keep their original tags. New ones use the new list. You can even retroactively tag a batch of old calls via the API if you want uniformity.
Can we do this without TECHMA's help? Technically yes — the docs are public. Practically, the post-call webhook that links the analysis to the tag means touching a bit of code and agent config. If you have a developer on hand, it's a one-afternoon job. Otherwise, we handle it.
Conclusion & next step
Conversation Tags aren't the flashiest thing ElevenLabs shipped in May 2026 — the four security locks in v2.47.0 got more buzz. But they're the feature that changes the operational life of an SMB with an agent already in production the fastest.
You go from a dashboard where 800 calls are an undifferentiated wall to a categorized board you can read in 15 minutes a week. Worth the quarter hour of setup.
If you'd rather we handle it for you — picking categories that fit your trade, configuring post-call analysis, the tagging webhook, doing the first audit together — grab 20 minutes with our team or browse our plans starting at $49/month. All our plans include tag setup.