Mercor raises $750M to expand enterprise AI evaluations

Mercor announced a $750 million Series D on November 10, 2026, valuing the company at $20 billion. It plans to invest in evaluations, post-training and simulated workplaces where AI agents can practice business tasks. The round included an employee tender, alongside investors including Benchmark, NVIDIA, Salesforce and Zoom.

For companies deciding whether to deploy an agent, the relevant question is what it can reliably finish in their own systems. Mercor is expanding its business around that measurement problem, rather than announcing a new general-purpose model.

What Mercor offers enterprises

Its enterprise offering spans workflow assessments, agent deployment and improvements to agents already running. The company says it combines domain experts, engineers and a platform for connecting context, tools and evaluations. The enterprise page directs prospective customers to book a demo or contact sales and does not publish a price list.

A separate data-licensing service is described as optional. Mercor says customers choose whether any data is licensed. Buyers should still establish exactly which records an evaluation needs, who can see them, how long they are retained and whether the contract permits reuse.

Investor Sapphire Ventures describes the strategy as combining expert recruitment, task design and realistic software environments. It points to Mercor’s July acquisition of Deeptune, whose team recreated enterprise applications for agent training. Sapphire participated in the funding round, so its assessment is an investor’s view rather than an independent product review.

A concrete example, with a small sample

Mercor’s earlier October 29 account of its internal incident-response agents shows what these evaluations can involve. Its evaluation process turns closed incidents into fixed tests, preserving the information available during the investigation while keeping later findings out of the agent’s evidence.

Two agents using the same underlying model achieved average graded scores of 0.79 and 0.55. Mercor attributed much of the difference to their tools and skills. These are scores on its own checks, not percentages proving general accuracy. The initial suite contained just five incidents with one attempt each, a limitation the authors explicitly acknowledge.

The example is useful because it separates model choice from the rest of the application. If an agent cannot retrieve the evidence needed to resolve an incident, swapping its model may leave the main problem intact. ByteForward has not independently reproduced Mercor’s results.

What a useful pilot should measure

A buyer can turn that distinction into a focused evaluation plan. Choose a workflow with an observable result, agree on failure conditions before running it, and reserve cases that were not used to tune the agent. The following are suggested acceptance checks, not claims about Mercor’s measured performance.

Include failures and human review in the cost calculation. A low price for one successful run says little if difficult cases require repeated attempts or lengthy cleanup. Record the model version, tool permissions and test inputs so a later comparison measures the same job.

Our coverage of Hugging Face’s harness adapters illustrates a related caveat: fewer tool calls on tasks both systems solve does not establish a lower bill across all attempted work. The denominator matters as much as the headline result.

What the funding changes

The announcement gives Mercor more capital to build its evaluation and training business. It does not supply a customer-specific return forecast, a deployment guarantee or evidence that a particular company should automate a workflow. For prospective customers, the strongest next step is a narrowly scoped trial with agreed outputs, review rules and a complete cost comparison.