The most interesting thing about Caretta is what building it admits.

A large part of this year's AI crop is aimed at getting rid of salespeople. Agents that dial, agents that qualify, agents that follow up, systems that can make more calls in an afternoon than a floor of humans could manage in a week. Caretta is aimed the other way. It exists to make one human on one call slightly better at it.

That is a position, not a compromise, and it is worth taking seriously.

What they built

Caretta works across three parts of a sales call. Before it, it briefs the salesperson on the account and its history. During it, it listens and helps them handle hard questions and objections. Afterwards it does the transcription, the notes and the system updates that usually eat the rest of the hour.

It plugs into the ordinary furniture of a sales team, Zoom, Google Calendar, HubSpot and Slack, and offers APIs, webhooks and an MCP server for teams that want to build on top.

Who else is in this field

Selling software to salespeople is one of the most fought-over categories there is. Gong and Chorus built large businesses recording calls and telling managers what happened afterwards. Outreach and Salesloft own the sequence of emails and calls. Clari forecasts which deals will close. Apollo and ZoomInfo supply the people to call in the first place. Avoma and Fireflies take the notes. Second Nature lets representatives rehearse against a simulated buyer, and Attention works on filling in the system afterwards.

Nearly all of it happens before the call or after it. The call itself has stayed a private conversation between two humans.

The point: the failure rate is the market

Automation arrives in every industry with the same promise, which is that the human step will eventually not be needed. It is almost never wrong about the direction and almost always wrong about how complete it will be.

What actually happens is that automation soaks up the volume and leaves a remainder: the calls that go sideways, the objection nobody scripted, the customer who is nearly persuaded and needs one right sentence. That remainder is small as a percentage and enormous in value, because it holds most of the decisions worth money.

As long as there is a failure rate, there is a business in helping the human who has to handle it.

Notice what that claim does not need. It does not need automation to disappoint, or the technology to stall, or anyone's predictions to be wrong. It only needs the last few per cent to stay hard, which is the one thing that has been reliably true of every wave of automation so far. This is not a hedge against AI working. It is what happens when AI works and then stops short.

Sales is the sharpest version of this, because closing is close to the definition of human commercial work. An agent can find the prospect, reach them, qualify them and book the meeting. What it has not shown is what a good salesperson does in the twenty seconds after a customer says something unexpected, which is to work out what they actually meant and respond to that rather than to the words. Volume has been automated. Persuasion has not.

Yes, it will be crowded

There is no point pretending this is an empty field. Several companies in this batch alone are working on sales conversations from one angle or another, and people have been trying to build assisted selling for a decade.

The crowding matters less than it looks, for one reason. Every company in the category is selling the same underlying thing: a better conversion rate on calls the business is already paying for. That is a number the buyer already tracks, already cares about, and can trace back. A crowded market where the value is directly measurable is a much better place to be than an empty one where it is not.

What will separate the winners is not model quality, because everyone will have the same models. It will be restraint. A salesperson in a difficult conversation has almost no spare attention, and a tool that puts too much on screen makes them read instead of listen, which is worse than no help at all and obvious to the customer. Knowing what not to show is a matter of taste, and it is only learned by watching a great many real calls go badly.

Where else this points

This next part is my opinion, not something the company has said.

The useful move is to look wherever full autonomy is being promised loudest, and ask what the tool for the remaining humans looks like. Those markets are priced today as though the human step goes to zero, which means nobody is building for the case where it does not.

Self-driving taxis are the obvious example. An enormous amount of money is committed to the idea that the driver disappears completely. Our view is that this will turn out to be less complete than it currently looks, and that remote operators, depot staff, exception handlers and safety supervisors will stay in the loop far longer than the forecasts assume. Every one of those people is doing a hard job under time pressure with partial information, which is exactly the shape of problem Caretta is solving somewhere else.

And the bet works either way. If autonomy stalls, the tool for humans is essential. If autonomy mostly succeeds, the remaining exceptions get rarer and harder, which makes helping the human who handles them worth more rather than less.

That is the general shape worth carrying out of this file. Wherever an industry is being sold a story about removing people, there is a quieter business in equipping whoever is left.

Why it is bigger than it sounds

The short description is sales AI. What matters is which half of the call it is built for.

  • Live help is the harder and more valuable half. Notes written afterwards have no time pressure and a full recording to work from. Helping someone inside a difficult conversation has neither.
  • The value is directly measurable. It is the conversion rate on calls the business already pays for, which the buyer already tracks, so the purchase is easy to justify.
  • Restraint is a real skill. Knowing what not to put on screen is learned by watching a great many real calls, and it does not become a commodity the way model quality does.

What to watch

One measure decides this, and it is not seats sold. It is whether people keep using it once they are no longer told to. Sales tools are routinely bought by management and quietly abandoned by the people they were bought for, so adoption in the first quarter tells you almost nothing.

If salespeople keep the live layer switched on when nobody is checking, the product is real and the argument holds. If they keep the notes and turn the live help off, then Caretta has built a better version of the easy half, which is a business, but not this one.