Somebody on a phone call says a website address, and the research that used to follow the call happens during it.

A voice agent on a phone call is a much narrower thing than a chat agent. It cannot show a table, it cannot open a document, and it cannot make six tool calls while somebody waits on the line listening to silence.

It also cannot read JSON aloud. Anything handed back to it has to already be a sentence, in the right order, with no formatting in it, because the next thing that happens to that text is a synthetic voice saying it.

So the usual shape of an integration, a set of small composable tools the model orchestrates, is exactly wrong here. Every extra call is a pause the person on the phone can hear.

One tool. It takes a URL and returns a paragraph. Everything else is hidden behind it.

Inside that single call the site is scraped, the business type and the technologies are worked out, three opportunities are drafted, a formatted report is generated, the report is emailed to the address found on the site, and the lead is written into the CRM. The caller waits once.

There is a second tool for one situation only: when no email address could be found on the site and the person on the call offers one. It sends the report from the previous analysis rather than running the whole thing again, so the caller does not sit through a second scrape for a report that already exists.

The output is deliberately prose. It opens by naming the company, states the industry, numbers the three opportunities, and finishes by saying where the report was sent. That structure is not decoration, it is the running order of a sentence somebody is about to speak.

It is deployed and connected, and calling it returns a real answer: asked about two of our own sites it came back with each one's actual page title, so the reading half of the pipeline is genuinely running.

The analysis half is not currently producing recommendations. Both checks fell through to the written fallback, which is what the code returns when the drafting stage does not complete, and both reported the industry as the default value rather than a detected one. The system degrades exactly as it was built to, which is the part that works. It is not delivering reports today.

The design idea is the transferable part and it survives the status. A voice agent is not a chat agent with a speaker attached. It gets one call and one paragraph, and building for that constraint means putting the orchestration on the server where the caller cannot hear it.

Two versions of it exist, identical underneath and pitched differently: one offers automation opportunities and files the lead in the CRM, the other offers landing page improvements and posts to a webhook.