🚧 This feature is currently in Alpha and available on request.
Alpha means the feature is in its earliest stage of testing and is being validated in live customer environments with real calls. You may experience changes, unexpected behaviour or occasional issues as we refine it. We welcome your feedback.
To get access, contact support@voyc.ai or your Voyc Account Manager.
If you have ever tried to capture a nuanced moment in a conversation using only a list of keywords, you will know how quickly it can turn into a guessing game. Did the customer say "frustrated" or "unhappy" or "fed up"? Did they say it at all, or just imply it?
Smart Detection changes that. Instead of building a keyword list and hoping it covers every variation, you describe what you want to catch in plain language, much like explaining it to a colleague, and the AI does the interpretation for you. It is faster to set up, more flexible in practice, and far better at catching the meaning behind the words rather than just the words themselves.
💡Tip: To learn more about the different types of Alert Conditions available please have a look at this article.
What Is Smart Detection?
Smart Detection is a new condition type available when setting up Alerts in Voyc. While the existing Phrase condition listens for specific words or phrases in a conversation, Smart Detection uses AI to evaluate whether the intent or content of a conversation matches what you have described.
Think of it like the difference between asking someone to find all conversations that contain the word "cancel" versus asking them to find all conversations where a customer is trying to cancel their service. The first will miss people who say "I want to leave" or "close my account". The second will not.
Smart Detection reads the conversation the way a human would, so you can focus on defining what matters, not on trying to predict every possible way someone might say it.
Setting Up Smart Detection on an Alert
Accessing Your Alerts
Smart Detection is added to an alert as a condition. You can add it to a brand new alert or to one you have already created.
Open the correct Channel.
Select Configurations.
Open the Alerts tab.
From here, either select an existing alert to edit it, or create a new one by clicking +New Alert and choosing an option.
💡Tip: If you have not set up an alert before, take a look at our guide on Setting Up Alerts on Voyc before continuing.
Adding the Smart Detection Condition
When viewing your alert, scroll down to the Conditions block.
Click +Add a Condition.
Select Smart Detection from the options.
In the Criteria field that appears, describe what should trigger the alert. Write it the way you would explain it to a new team member. For example: "The customer explicitly states they are struggling to make payments or says they cannot afford what they owe."
Click Save.
Optionally, add Supporting Phrases in the field provided. These are words or phrases that give the AI additional signals to work with. They are not required, but can improve detection for alerts where there are strong, predictable patterns in how something gets said. These can be viewed as examples the LLM can use when understanding what criteria would match the alert.
Select Done in the bottom right corner.
Writing Effective Smart Detection Conditions
The quality of your Smart Detection setup comes down to two things: what you name the alert and how clearly you write the condition. Both feed the AI, so both are worth getting right.
Naming Your Alert
The alert name is not just a label for you, it also gives the AI context about what it is looking for. A vague name creates ambiguity; a specific one sharpens detection.
A good alert name is:
Specific, not categorical. "Suicidal ideation" is clearer than "Vulnerable customer". "Customer threatening to cancel" is clearer than "Escalation".
Behavioural or signal-oriented. Frame it around what someone says or does, not a label you would apply to them afterwards.
Consistent with your criteria. If your criteria describes a customer threatening to self-harm, naming the alert "Customer complaint" creates a mismatch that confuses the AI.
❌ Too broad | ✅ Better |
Fraud | Customer denying authorising a transaction |
Escalation | Customer demanding to speak to a manager |
Vulnerable customer | Customer expressing inability to afford repayments |
Writing Your Criteria
This is the most important part. The AI treats the criteria as the source of truth, so vagueness here directly affects how well your alert performs.
Write criteria that describes what someone actually says or does, not abstract categories. A good gut check: if a colleague could read the transcript and immediately know whether the criteria was met, you are on the right track.
A simple structure to follow:
[Who] [does or says what] [in what context, if relevant]
Example: The customer explicitly states they cannot afford their repayments when discussing their outstanding balance
Rules of thumb:
Describe what is said, not how it feels. "The customer threatens to report the agent" is something you can point to in a transcript. "There is a hostile tone" is a judgement call. If you cannot quote it, it is probably too vague.
Say what needs to actually happen for the alert to fire (be explicit). If the customer needs to come out and say it directly, write "the customer states...". If it is enough for them to imply or suggest it, write "the customer implies or suggests...". The AI will follow your lead either way, so be intentional about which you want.
If multiple things need to happen, say so. If your alert should only fire when a customer both denies a transaction AND mentions fraud, write both into the criteria joined by AND. If you just write "the customer mentions fraud or denies a transaction", the AI may trigger on either one alone. Use AND when both must be present, OR when either is enough.
Say who is speaking. The same words mean different things depending on who says them. An agent explaining fire cover as part of a sales call is routine; a customer mentioning fire damage is a possible claim. If it matters whose words you are evaluating, name them: "the agent says..." or "the customer says...".
Put the detail in the criteria, not just the name. If your alert is called "Repeated complaints", that detail needs to live in the criteria too, not just the name. For example: "The customer references a previous complaint or states that they have raised this issue before." The AI reads the criteria, not the name, when deciding whether to fire.
❌ Vague | ✅ Explicit |
The customer is distressed | The customer explicitly states they are struggling mentally, mentions self-harm, suicide, or expresses that they cannot cope |
Agent is being unhelpful | The agent refuses to escalate, denies a request without explanation, or tells the customer there is nothing they can do |
Possible fraud | The customer denies having made or authorised a transaction that is being discussed |
Excluding false positives
If there are common patterns that should not trigger the alert, add an exclusion directly in the criteria. For example: "Exclude cases where the customer is discussing a third party's situation rather than their own." This keeps your alert focused and reduces noise.
💡Tip: Avoid criteria that is too broad or that could apply to many different conversations. If your criteria could match both a cancellation and a general complaint, you will end up with an alert that fires on both. Narrow it down.
Using Supporting Phrases
Supporting Phrases are not necessary but can greatly improve accuracy by helping the AI identify the parts of a conversation worth evaluating against your criteria. Think of them as signposts or examples, not gatekeepers. They point the AI toward relevant moments; the criteria then decides whether those moments actually qualify.
Good Supporting Phrases:
Are close to what someone would actually say: "can't afford", "I'll report you", "cancel my account"
Include informal or realistic variants: "gonna", "fed up", "dunno"
Cover synonyms when your criteria is broad: "distressed", "overwhelmed", "can't cope"
Avoid:
Abstract terms that would not appear verbatim in speech: "vulnerability", "escalation", "non-compliance"
Repeating the alert name as a phrase, it adds nothing
Adding too many. Focus on the 5 to 10 most distinctive signals rather than trying to cover everything
For example, if your criteria is "The customer explicitly states they cannot afford to make their repayments", useful Supporting Phrases might include: "can't afford", "don't have the money", "struggling to pay", "can't keep up with payments".
💡Tip: If your criteria requires a specific word or phrase to be present, write that into the criteria itself. Supporting Phrases cannot enforce logic, only the criteria can do that.
The Gist
Smart Detection lets you set up alerts by describing what you want to catch in plain language, rather than trying to predict every keyword. Add it as a condition on any alert, write a clear and specific description of the behaviour you are looking for, and optionally include Supporting Phrases to sharpen the results. The clearer your description, the more reliably your alert will surface the conversations that actually matter.