Perplexity gives AI agents a standing job, and Box rebuilds how its AI searches files

Teammates are built on Computer, the agent that Perplexity launched in February, which works through tasks that a person gives it. A Teammate keeps its job instead. It has a role, it watches a source of work such as an inbox or a Slack channel, and it starts when something arrives there. When it needs a decision, it asks a colleague in Slack. It needs Perplexity's $200-a-month Max plan, and OpenAI and Anthropic offer similar agents.

That moves the hard part from writing a prompt to writing a job description. A standing job needs four things on paper: where the work arrives, what a finished piece looks like, what the agent may do alone, and when it must ask. A new colleague learns these by watching for a week; the agent has only what you wrote.

For example, an accounts payable team could give a Teammate the supplier inbox and the list of invoices due for payment. When a supplier asks when an invoice will be paid, it finds the invoice, drafts the reply and asks the clerk in Slack before sending. At 40 such emails a day, four minutes each by hand and 30 seconds to approve a draft, the clerk gets more than two hours a day back: over 40 hours a month for $200.

Each approval is also a record. After a few weeks, it shows which replies the clerk never changes, and those are the ones the agent can start sending alone.

Archy takes drive-thru orders and gets more than 90% of them right. McDonald's ended its first attempt, with IBM, in 2024 after too many wrong orders. This time the case rests on staff time: at least 50 hours a week in each restaurant.

That is about seven hours a day, roughly one full shift. The hours save money only when that shift comes off the schedule. Spread across the crew as a free minute between cars, they leave payroll unchanged. Any AI tool deserves the same test: its fee against the shift or the purchase that disappears.

The value of new contracts that Palantir signed in the third quarter rose only 6%, while revenue rose 96%. A shorter contract books less value on the day it is signed, even when the customer goes on to spend more.

A short contract limits the buyer's loss if the tool fails. If it works, the team builds its routine around it within weeks, and the bigger contract is negotiated when leaving would mean undoing that work. So put the expansion price, the exit terms and a monthly spending cap into the pilot contract, while walking away is still easy.

Every answer that Box's AI gives about company files starts with a search for the right passages. Early next year, a new engine will do that search by meaning instead of exact words.

That step matters more than the model. The assistant writes only from the passages it found, so a clause that the search missed cannot appear in the answer, and nothing shows that it is missing.

In 1985, lawyers searching documents for a lawsuit believed they had found more than 75% of the relevant ones. They had found about 20%, because the documents used words the lawyers never thought to search for. An AI assistant hides that gap better, because it writes a confident paragraph over it.

So test any file assistant with ten questions whose answers you know, and word some of them differently from the documents.