Connecting is the easy part. You paste in one address, sign in the way you always do, approve what the AI is allowed to touch – and then Claude, ChatGPT or Gemini is standing with a hand on your phone system. The steps for each of the three are in a separate article.
What is rarely written down anywhere is what to type next.
That is where people stop. The connector is green, the chat window is empty, and the first sentence tends to be «what can you do?» – which gets you a list of tools, not a finished job. Below are seven jobs from an ordinary week of running a voice agent. For each one: the sentence you actually type, what happens behind it, and the trap that costs the most time the first time round.
One thing first. The AI you are sitting in does not call anyone. It does what you used to do with the mouse: searches, reads, counts, proposes, prepares. The only thing that reaches the world outside is an outbound call – which is exactly why it sits behind a confirmation.
1. Monday, 8.40 am: what came in overnight?
What you type: «Show me every inbound call from 4 pm yesterday until now, local time. Which ones were handed to a human, and what were they about?»
What happens: the AI pulls the call history for the period, picks out the escalated ones, opens the transcripts and summarises them in plain language. You read five lines instead of clicking through twenty log rows to find the three that mattered.
The trap: two of them, and both are mundane.
Time is the most common source of confusion. Name the time zone and give the date. «Yesterday» is not unambiguous at 8.40 in the morning, and a window that is off by a few hours gives you a morning summary missing the evening before.
And: your own test traffic looks exactly like real traffic. The calls you place yourselves to check a change sit in the same list as the customers. Tell the AI to leave your own numbers out, once, and keep that sentence in a saved prompt.
2. Tuesday: why did we lose that one?
What you type: «Find the call that came in around 2.10 pm yesterday and ended in under a minute. Show me the last thirty seconds word for word.»
What happens: the AI finds the call, pulls the events and the transcript, and gives you the exchange as it happened.
This is the job that changes the most for people. Not because it saves time – finding one call in the interface is not slow either – but because the threshold for doing it drops to nothing. When it costs one sentence, you do it ten times a week instead of once a month. And one transcript tells you something a monthly average never will: it shows you the sentence where it broke.
As a rule that sentence is either the opening line or the moment the agent should have handed over and did not.
The trap: do not ask for «every call from last week, word for word». Transcripts are long, and you burn both your patience and the context window on material you were never going to read. Ask for the selection first and the detail second.
3. Wednesday: the calling round of eighty
What you type: «Here are eighty leads from the CRM. Create them as people in Threll, prepare calls with Harald tomorrow between nine and eleven, and show me the whole list before anything goes out.»
What happens: the people are created in one operation – up to a hundred at a time – and the calls are prepared in the next. Preparing calls no one. It produces a proposal naming the worker, the number it dials from, who is being called, when, and with what context. You read it. If you say yes, the whole round is queued together: either all of them or none. An approval can only be used once, so the same round cannot go out twice by accident.
The trap, and it is worth the whole article: anything that differs from customer to customer belongs in the context for that individual call – not in the instructions field.
The instructions field is not an addition to the script. It is the script. Write «remind them the contract expires on 1 November» there and you have effectively replaced Harald's entire script with that one sentence, for those eighty calls. The agent loses its opening line, the disclosure, the escalation rule and everything else somebody thought carefully about. Put the detail in the context instead, and the agent keeps its script while still knowing what it needs to know about this customer.
One small practical matter: scheduled calls go out within a couple of minutes of the time you set, not on the second. Schedule in windows, not in points.
Changed your mind? «Cancel the round at nine tomorrow» works right up until the calls are sent. After that it is no longer a scheduled job, and the agent is on the line.
4. Thursday: the script that stumbled
What you type: «Fetch Nora's active script, read it against her last twenty calls, and propose what should change in the section on pricing. Show me the proposal, do not write it in.»
What happens: the AI finds the active script document for that worker, reads it against real calls, and points at the places where the agent actually hesitated – not where you assumed it did.
How the change works: a new version is a new document, not an overwrite. The previous one has to be taken out of use, and it is the combination of active and linked to the worker that decides what the agent says on the next call. A new document takes a few seconds to finish processing before it applies. Always ask to see what changed, and ask for confirmation of which document is now the active one.
The trap: not every worker keeps its script as a document. Several inbound setups hold the lines in the worker's own configuration text instead. Ask where the AI found the script before you ask it to change anything – otherwise you can end up with a tidy new version that no call ever uses.
And what holds regardless: what the agent says can bind you. A proposed wording is a proposal to a human being, not a decision.
