Prentis raises $100M for computer-use office agents

The interesting part of this funding round isn't the money. It's the training data. Prentis, an AI lab building computer-use agents, is in talks to raise $100 million at a $1 billion valuation, according to TechCrunch — and its product is a model trained on how office workers actually navigate the documents and systems you already pay for. Not an API you call. An agent that watches the work, then does the work.

What actually happened

Prentis launched in April 2026, co-founded by CEO Ritankar Das alongside Reid Hoffman and Mark Pincus. It has more than 25 employees, including researchers who came from OpenAI, Google DeepMind, Meta, Tencent, and Alibaba. TechCrunch reports the company has signed customer contracts worth up to $50 million, with early customers including a healthcare management services organization and manufacturers.

The work it targets is deliberately unglamorous: insurance claims processing, customs duty refund exceptions — the queue nobody wants to own. Prentis says its in-house model beats frontier models on computer-use benchmarks; that's a vendor claim on a vendor-chosen benchmark, so file it under "reason to run a pilot," not "result."

The number worth staring at is the pricing model. TechCrunch reports investor materials projecting a $75 million annualized run rate by Q3 2026, based on capturing roughly 20% of the savings the agents produce. The round is in talks, not closed. The revenue is a projection built on a percentage of a number the vendor helps define.

Why it matters for your business

Two things follow, and neither depends on whether Prentis specifically wins.

Outcome pricing is a measurement contract. "We take 20% of what we save you" sounds risk-free until you ask who computes the baseline, who counts a task as completed, and what happens when volume drops for reasons that have nothing to do with the agent. If you sign one of these — from Prentis or anyone else — the definition of "savings" is the entire negotiation. Get the baseline measured before the pilot starts, by you, in a system you control.

Your workflows are the asset being trained on. A computer-use agent learns by observing screens: your ERP, your claims portal, your vendor's admin panel, and whatever is on those screens. That's customer data, pricing, and process IP flowing into someone else's model. Ask three questions in writing — does our data train a shared model, can we get an export of what was captured, and what happens to it when we leave. If the vendor can't answer plainly, that's the answer.

The underlying shift is real and worth acting on: the automation frontier has moved from systems with clean APIs to the swivel-chair work between them. That's where most small-business back offices actually lose hours. You don't need a $1B lab to start on it — you need one process mapped, one baseline measured, and a human approving the output.

Sources: TechCrunch.