Improve your assistant
Your assistant answers, but doesn't quite do what you want yet? Almost every issue falls into one of three categories. For each one, this article describes what you see, what causes it and how to fix it, and ends with a testing routine.
The assistant ignores instructions
What you see
You've told the assistant in the prompt to always ask for the postcode, or never to quote a price, but in real conversations it still happens. Sometimes it follows the rule, sometimes it doesn't.
Cause
It's usually the prompt itself, the text on the Instructions tab:
- The prompt is too long or contains instructions that contradict each other. The AI model weighs all rules against each other; the more text, the less weight each rule carries.
- The instruction is buried in a long paragraph instead of standing on its own line.
- The rule is written descriptively ("we prefer not to mention prices") rather than prescriptively ("Never quote a price").
- The prompt is full of facts (rates, opening hours, service area) that belong in Knowledge or in your business details. Those facts crowd out the behaviour rules.
Fix
- Cut the prompt down to roughly 300 words or fewer. Remove anything that doesn't directly help the assistant run the conversation.
- Put every instruction on its own line: one instruction per line.
- Phrase rules as If … then …: "If the caller asks for a price, say that an engineer will quote on site."
- Move facts to Knowledge. The assistant always gets the questions and answers and looks up documents when needed, without cluttering the prompt. Opening hours and Service area are under Settings → Business; only an admin can change them.
- Your change saves automatically. Test the changed instruction via Try it out on the assistant (see below).
More on structuring a prompt in How to write a great prompt.
Data capture is sloppy
What you see
After the call, the caller's name, address or phone number is missing. Or the assistant asks for everything in one breath ("What's your name, address and phone number?") and the caller loses track. Postcodes and house numbers come through wrong.
Cause
- The prompt says the details are needed, but not when or in what order.
- Asking several questions at once doesn't work on the phone; the caller answers one and forgets the rest.
- Postcodes, house numbers and names are easily garbled by speech recognition if they're not confirmed.
- The details aren't stored anywhere structured, so they only exist somewhere in the transcript.
Fix
- Put the questions in the Conversation flow field on the Instructions tab as numbered steps, in the order you want them asked. In Advanced mode they go in Full prompt.
- On the phone, have the assistant ask one question per turn: name first, then address. In a chat it asks for all the details together in one message by itself, even when your prompt says to ask one question at a time. On calls, SMS and WhatsApp the assistant doesn't need to ask for the phone number: it already has it and only asks if the customer wants a callback on a different number. In webchat it only has the number if the visitor types it, so have the assistant ask for it there.
- Add a rule to repeat or spell postcodes and house numbers: "Repeat the postcode and house number and ask whether it's correct."
- Check on the Tasks tab that Save requests is on; it is by default. The assistant then saves the request, with name, address, problem, urgency, preferred time and the customer's number, and notifies you. You'll find the request under Inbox → Requests. With a mobile number, the name also goes on the contact with that number.
- Need extra details, such as the type of fault or the boiler brand? Have the assistant ask for them in Conversation flow and define them as Custom variables to extract on an action that runs after the call (for example the email action, on the Actions card of the Tasks tab). After the conversation the assistant takes the values from the transcript; anything that wasn't said stays empty. You use them as a variable in the action's content, for example
{{boiler_brand}}, and you see them on the conversation under Technical details.
See Technical settings and Actions.
The tone is wrong
What you see
The assistant is stiffly formal with regulars who are used to a casual chat, or the other way round. It talks for too long, repeats the caller's question back, or sounds flat and robotic.
Cause
Tone is set by three things: the formality setting, the voice you chose with its speaking style, and the rules in the prompt. If any of the three doesn't match your customers, the whole conversation sounds off.
Fix
On the Voice & greeting tab, set Formality to Formal or Informal. This carries through to what the assistant words itself. The Greeting and Outbound greeting stay as you typed them, so adjust those yourself if needed. Fixed platform lines, such as the line when transferring a call, don't change with it.
On the same tab, adjust Speaking style: Speed if it talks too fast, Tone, Sound, Pitch and Intonation if it sounds stiff or read aloud. The sound of the voice itself mostly comes from the chosen Voice, so also try another voice if it keeps sounding flat.
Add a few tone rules, for example:
- "Keep answers under two sentences."
- "Ask one question at a time." (calls only: in a chat the assistant asks for all details in one message)
- "Don't repeat the caller's question back."
- "Use everyday language, no trade jargon."
For calls, put them in Speaking style: choose the pencil next to Conversation and add to the text. The phone voice doesn't get the prompt directly; the prompt steers what the assistant does, such as asking for details, booking appointments and transferring calls. For SMS, WhatsApp and webchat, put them under Important rules on the Instructions tab (in Advanced mode, in Full prompt), or in Chat style if Separate style for chats is on.
Which voices are available per language is covered in Languages & voices.
Testing with Try it out
Test every change as text first, then by phone.
Open the assistant under Assistants and choose Try it out in the header. In Call you can test as text, without making a call. You type what a caller would say and see the answer immediately. If the assistant used a tool, a Tools used line under the answer gives its name, such as search_knowledge_base (searching documents) or book_appointment (booking an appointment). Questions and answers aren't listed there, because the assistant always has them. Test search lets you check what your knowledge finds.
With Start call you talk through your microphone and the assistant answers in the saved voice. It is not a live phone call: no real number is dialled. You can interrupt: the assistant stops speaking and your words show at once. It talks through a brief sound. After you hang up, Resume call continues without repeating the greeting; New call starts a fresh conversation. Chat is the same chat as the website widget, and it works even if webchat is still off.
In Call and Chat, bookings, emails and other actions are only simulated. Switch on Perform actions for real if they should really run.
Always the latest version
Try it out waits until your last change is saved. If that fails, the test does not start and you see the error on the page.
Then always test by phone as well. Timing, interruptions and speech recognition behave differently from text, and that's where things most often go wrong. Transfers and voicemail can only be tested this way: they aren't available in Call and Chat. The conversation there also runs as webchat: the assistant doesn't know the customer's number, and the rules for phone calls, such as spelling details back, don't apply. So check those by phone.
Call the number set to Answers incoming on the Channels tab; a phone number takes a paid subscription. If your plan includes outbound calling (Business, Pro and Enterprise), you can also use Real call in Try it out: the assistant calls the number you enter, from the number you choose under Call from number. It lists your numbers that can place outgoing calls. Real call counts towards your usage. A test by phone is a real call: bookings, emails and other actions really run.
Run through at least these scenarios each time:
- The happy path – a customer with a routine job who wants an appointment. Are all details asked for, and does the wrap-up make sense?
- An emergency – "Water is pouring through the ceiling!" Does the assistant recognise the urgency and transfer, or promise an immediate callback?
- An unknown question – something that isn't in your knowledge. Does the assistant say honestly that it doesn't know, or does it make something up?
- A request for a human – "I want to speak to someone." Does the assistant transfer according to your transfer rules?
- The caller switches language – starts in the assistant's language and moves to another one. Does the assistant do what you chose under Language switching on the Voice & greeting tab?
Note what goes wrong, change one thing at a time and test again. That way you know exactly which change made the difference.
You don't have to type these yourself each time. On the Tests tab, the AI writes conversations like these for your business, runs them by chat and suggests a fix for each problem. See Test your assistant.
Try changes on a copy
A bigger change, such as new instructions or another voice, and you'd rather not risk the live assistant? Choose Duplicate at the top of the assistant. The copy is not live: its webchat and WhatsApp are off. Change the copy and test it with Try it out.
The copy also gets the assistant's tests. On the Tests tab of the original, run the same tests against the copy under Compare with another assistant (Enterprise), and see per topic whether it does better (▲) or worse (▼). Happy with it? Make the same change on the original.