Capacity and concurrency
The agent always answers — until everyone calls at once
A voice agent has no queue, and therefore no warning light either. Capacity is the lowest of four numbers, and the four are owned by four different parties.
Articles and updates
Knowledge sources
The agent answered. Where did it get that from?
An answer from a voice agent sounds equally certain whether it comes from an approved price list or from a note nobody has read since last year. Seven questions that decide which one the customer gets.
Latency and tool calls
Your voice agent can talk while it waits. Now someone has to decide what it says.
Google shipped two speech models on Tuesday that run tool calls in the background while the agent keeps talking. It removes the most awkward silence in a customer call – and moves the error somewhere harder to spot.
Conversation design
A manuscript is not a script. How to write one that survives a real call.
The manuscript a customer sends over was written by someone picturing the call going well. A script also has to cover the times it doesn't. Ten moves that separate the two.
Orchestration
Seven jobs you can move from the dashboard to the chat window
Connecting Threll to the AI you already use takes ten minutes. The hard part is knowing what to ask for afterwards. Here are seven jobs from an ordinary week – the sentence you type, what happens behind it, and what usually goes wrong.
Accessibility
“Voice is the most inclusive interface.” Half of that is true.
The argument for letting an agent answer the phone is often that the phone is the one channel everyone can use. That holds for more people than you would think – and not for a group that almost never makes it into the calculation.
Traceability and records
How to build the record of what your agent actually said
A customer calls in September and says your voice agent promised her something back in June. The question isn't whether she's right – it's whether anyone at your company can find out today.
Division of labour and staffing
The agent takes the easy calls. Someone has to take the rest.
Salesforce gave seven AI agents first names and job descriptions this week. None of those descriptions covers the human who picks up when the agent stops – and that is the job that changes most.
Call flow
One customer call, beat by beat. Seven links that can break.
OpenAI has now put a price on the voice layer alone: $0.05 a minute for the part that listens and speaks. Which makes it worth walking one perfectly ordinary call through link by link, and asking who actually owns each of them.
Operations and maintenance
Your voice agent doesn't know you closed at noon
Almost everything in a voice agent's configuration stays true until somebody changes it. Opening hours are the one exception: they stop being true on their own, without anyone touching a setting.
Dental care
An AI receptionist for a dental clinic: what it should answer – and what it never should
A dental front desk runs one phone queue with six completely different calls in it. Only some of them should reach a voice agent – and one of them it should never assess on its own.
Names and speech recognition
The agent gets the phone number right. The name is another matter.
On the measurements that exist of real agent conversations, names are the weakest category in every single model – several times worse than phone numbers. Not because models are bad at names, but because a name has nothing to check against.
Model selection and testing
Two new speech models in three days. Neither score was measured on your phone line.
Meta and Microsoft each released a speech recognition model in the first week of September, with leaderboard places and error rates to show for it. The numbers are real enough. They were just measured on something quite unlike a Nordic phone line.
Liability and binding agreements
The checklist before your agent says something that binds you
"The AI said it, not us" has been tried as a defence, and it failed. Six questions that should have an answer before your voice agent speaks about price, dates or coverage.
Context and handovers
Context dies at the handover. A new layer is not the cheapest fix.
Two numbers sit side by side in the same Genesys report: 95 percent of consumers expect not to have to repeat themselves, and 48 percent of companies pass nothing at all from the virtual agent to the human. The industry's answer is a coordination layer. That is right for some – and an expensive answer to a question most companies can ask much earlier.
Estate agency
The bid cannot come by phone. Everything else does.
In an estate agency, the one call that matters most is also the only one the regulation forbids taking verbally. The rest of the phone volume is something few offices have decided who should handle.
Accessibility
The agent heard everything. It was the caller who didn't.
Almost all the work of making a voice agent work in a Nordic language runs in one direction: getting the machine to hear the customer. This call goes through the other one.
Orchestration
How to orchestrate Threll from your favourite AI
Threll has an MCP server. That means you can run your voice agents from Claude, ChatGPT, Gemini or Microsoft Copilot – in the same chat you already work in, without writing a line of code.
Scope and boundaries
“We can't put it live until it's connected to the back-office system”
The sentence turns up early in almost every project, and it sounds responsible. It is also why a fair number of voice agents never make it out of a meeting room.
Audio and speech recognition
The phone line throws away half your voice
The demo understood everything. Then the agent went onto a phone number, and now it can't tell an F from an S. The model is the same one – the audio reaching it is not.