Designing AI that sounds like a junior associate, not a chatbot
Notes on the conversation design behind our response layer - tone, hedging, and the deliberate use of silence.
Most AI outbound still reads like AI outbound. The em-dashes, the 'I hope this email finds you well', the over-eager triple exclamation when a prospect shows the faintest interest. Property developers - who get dozens of these a week - have learned to filter them in under a second. The work we've spent the last twelve months on at Capital Edge is largely about the opposite: designing a voice that sounds like a smart, slightly cautious junior at a brokerage.
Tone is a specification, not a vibe
Our system prompt does not say 'sound friendly and professional'. It specifies sentence length distributions, the ratio of statements to questions, when to use 'we' versus 'I', when to drop the greeting entirely, and a hard ban on twenty-three specific phrases that any developer reading their inbox will pattern-match to a script.
Tone is downstream of those constraints. When you specify the mechanics of how a good junior writes - short sentences, one question per reply, no recap of what the prospect just said - the warmth falls out for free.
Hedging is a feature
The default failure mode of a sales-tuned LLM is to over-promise. A developer asks 'can you do 75% LTV on a part-built scheme?' and an under-designed model will say 'absolutely, we can definitely arrange that'. A good junior associate says 'on a part-built scheme it depends on the QS report and the exit - happy to walk it through with one of our specialists Thursday at 11?'.
Hedging, used well, is one of the strongest trust signals in finance. We treat it as a first-class behaviour in the model, not a safety bolt-on.
The deliberate use of silence
Not every message deserves a reply. A prospect who has said 'not now, maybe Q3' should hear from the model again in Q3 - not on day three, day seven and day fourteen with bumped subject lines. We spent a lot of design effort teaching the model when not to send.
Counterintuitively, dialling outbound volume down on uninterested prospects has been one of the larger lifts to overall reply rate this year. Inbox warmth is finite. Spending it on prospects who said no is a tax on the ones who haven't replied yet.
What we removed
The two largest single improvements we made to perceived quality this year were both deletions. First, we stopped letting the model use the word 'just'. Second, we banned any sentence that began 'I wanted to'. Both are tells. Both were collectively responsible for more 'this is AI' replies than any other phrasing.
Conversation design in 2026 is largely subtraction. The model knows how to write. The job is teaching it what not to say.
The bar for AI in outbound is no longer 'can it write a coherent email'. It is 'would a developer scrolling on the train recognise this as a person'. That is a tone problem, not a model problem, and it is mostly solved with specification rather than scale.
Capital Edge publishes one note a month on UK bridging finance, paid acquisition, and AI-led outbound. Written for brokers, by the team running the playbook.