Hone, a San Francisco-based artificial intelligence startup, has closed a $60 million seed round at a $285 million valuation, according to Bloomberg — cited by RuntimeWire — and FinSMEs between October 8 and 9, 2026. The round was led by Benchmark and co-led by Index Ventures, with participation from Elad Gil, Hanabi, Definition and Diffusion. It is an unusually large figure for a seed round, even in an AI market used to big early-stage checks.
The company was founded by Moritz Stephan, Oliver Brady and Carlo Kobe. Stephan, the CEO, was chief of staff at Cognition, the company behind the Devin coding agent, and according to Hone’s website the team comes from Cognition, Mercor and Jane Street. Bloomberg’s headline highlighted that former OpenAI and Cognition staffers want AI to help run a business. Hone describes itself as an applied research lab focused on closing the gap between what models can do and the economic outcomes they drive.
The product is called Engines. Unlike chatbots, which answer questions, and agents, which complete assigned tasks, an Engine is given a business outcome to achieve — for example, improving customer retention, raising lead conversion or cutting credit losses — and works on it continuously. The company sums up the difference this way: with chatbots, you prompt them; with agents, you assign the work; with Engines, you set the goals and guardrails.
According to Hone, each Engine onboards itself into the organization: it connects to its systems, gets to know the people and processes it will work with and, before touching anything live, runs against the company’s history — old tickets, past deals, previous decisions — so the team can see what it would have done, correct it and decide when it is ready. After it acts, it watches what happened and learns from the result. Engines work in Slack, Teams and email.
The platform includes organizational memory that stores user preferences and company best practices, custom policies so sensitive actions require prior approval, and a log of every action and permission so execution is auditable. Engines write the routines that handle their work and refine them in a versioned repository, so repeatable work becomes reusable code. Changes are tested against past cases in simulation before going live.
Examples published by the company include an inbound lead conversion Engine, a vendor management Engine that analyzes spend and renegotiates contracts, a credit risk Engine, a product adoption Engine and one that tracks customer commitments from pre-sales through delivery. Cognition and inference platform Modal are listed as early customers, according to RuntimeWire, and Hone’s website features testimonials from employees at those companies and at Lucanet.
Stephan laid out his diagnosis of the market on the TBPN show. “The ROI of AI outside of engineering today is quite disappointing. That’s what we’re here to solve,” he said. In his view, most organizations use agents to hand off individual tasks, micromanage outputs and stitch together workflows that do not clearly map to business outcomes. Hone’s pitch is to “put AI in charge of outcomes, dial up the autonomy.”
The first commercial beachhead Stephan describes is revenue, specifically account management: watching signals across customer relationships and acting on them automatically. He considers this kind of use case common in sectors such as banking and insurance. Rather than building bespoke solutions for each customer, Hone is developing template Engines that can be deployed with little setup, and uses forward-deployed engineering work with customers as a source of learning to improve those templates.
The company is also exploring budget-based pricing: instead of billing per task or per seat, the customer allocates a fixed budget and the Engine decides how to use it to reach the outcome, including choices about how much context to gather or how many alternatives to explore. Stephan cited Sierra as a reference point for outcome-based pricing, while acknowledging that the diversity of Hone’s use cases makes direct replication difficult.
On security and compliance, Hone says it does not use customer data to train models, that execution is sandboxed, that data is encrypted in transit and at rest, and that it holds SOC 2 certification. It offers a fully managed service in its own cloud or deployment inside the customer’s private cloud, with customer-controlled keys. On model strategy, Stephan was pragmatic: if frontier labs are likely to close a capability gap within a few months, investing heavily in custom training now would probably be a mistake.
The round reflects investor appetite for the next phase of enterprise agents. After the wave of AI-assisted coding tools, capital is looking for companies able to bring AI to functions such as sales, procurement, risk or operations, where value is measured in business metrics. The fact that Benchmark and Index Ventures, two of the best-known venture firms, led a seed round of this size underlines the bet.
The doubts are also clear. As RuntimeWire notes, Hone has not disclosed revenue, customer outcomes or independent performance data. Handing an autonomous system responsibility for metrics such as retention or credit losses requires strong safeguards, especially at a time when several labs have documented agents bypassing restrictions to meet their goals. The prior approvals and auditing Hone promises will be decisive in winning the trust of regulated customers.
With $60 million in its first round, Hone has the resources to hire researchers and engineers and to test whether its Engines concept works at scale. If it can show measurable improvements in business results, it could mark the shift from agents that execute tasks to systems that take on whole objectives. If not, it will be another example of the high expectations surrounding enterprise AI in 2026.
