Evals and accuracy

If you cannot grade it, do not automate it

Amol Patil, Founder · 24 July 2026 · 8 min read

Explore with AI

ChatGPTPerplexityClaudeGoogleGrok

The short version

• Before anyone builds an AI agent for a process, the people who run it should grade real past cases, with the right outcome and the reason for each.
• If two of them grade the same case differently, the process is not defined yet, and an agent would only repeat the disagreement faster.
• OpenAI's Presence, launched on 22 July, lets teams check agents with graders and simulations before they meet users. Those checks need written answers to run against. We think writing them is the first job of a forward deployed engineer.
• Use the worksheet below. If nobody can write the grades, the first project is to define the process, not to automate it.

We think nobody should build an AI agent for a process until the people who run it have graded real examples of it. A process nobody can grade is not ready to automate, because nobody will be able to tell whether the agent is doing it right.

In our view, writing those grades is the first job of a forward deployed engineer, the engineer who works inside a company until an AI system runs part of its daily work. It comes before any integration or build.

What OpenAI shipped on 22 July

On 22 July, OpenAI launched Presence, a product for putting voice and chat agents to work in customer and internal workflows. Each deployment starts with one specific job, such as resolving billing issues. OpenAI's forward deployed engineers and selected systems integrators lead every deployment, and the product is not yet self-serve.

Before an agent reaches users, simulations and graders check whether it reached the right outcome, followed policy, used tools correctly and escalated when it should have. After launch, teams can test each proposed change against the version in production, then approve a controlled rollout. In screenshots OpenAI shared with VentureBeat, administrators run a batch of simulations against a revised refund policy.

Every one of those checks needs a written answer for each case. A grader cannot say an agent reached the right outcome until someone has written down what the right outcome was. No product can write those answers for a business. They come from the people who do the work.

OpenAI's own support line shows the order. OpenAI says the agent on its phone line met or exceeded, within weeks, the benchmarks OpenAI uses to grade the quality of its human frontline support. The bar was one OpenAI already applied to people.

Do not automate an invention

A day later, Nate Herk made the builder's case in Don't automate an invention. He argues that automating a broken process scales the problem, and that automating a process the business has never run is worse.

He adds that an automation gives no warning when it goes wrong, so it needs evals, and evals need a golden dataset of known correct answers. His advice is to confirm that such data exists before building, and to get the process working with a person first.

We agree, and we would make the test stricter. A process can have a manual and years of history and still fail, because nobody ever wrote down the right outcome for the hard cases. The test is not whether the process is documented. It is whether two people who run it give the same grade to the same case.

Why grading comes first

Anthropic's guide to agent evals says evals help most early, because they force a team to define what an agent should do. A team that waits ends up working out its success criteria backwards from a live system. The guide calls a task well formed when two domain experts would independently reach the same pass or fail verdict.

A 2024 study explains why the grades cannot all be written in advance. Shreya Shankar and colleagues, in Who Validates the Validators?, found that people need criteria to grade outputs, yet grading outputs is how they work out their criteria. The authors call this criteria drift.

Our reading is that a team discovers the rules of a process by grading its real cases. Grading is therefore the first step of an automation project, not a test at the end of it. It costs the time of people who already do the work, and it exposes their disagreements before anyone pays for a build. A missing definition of accurate is the first cause we list in why AI pilots die.

The grading worksheet

Run it on the process you most want to automate, with the person who does that work today.

What the grades tell you

We would want the two graders to agree on at least 9 cases in 10 before anyone builds. Agreement does not make the process perfect. It gives a grader, human or automated, something to check against, so every later change to the agent can be tested on the same cases. The same cases also set the pass mark for an agent's probation, the period when a person reviews every output before the agent may act alone.

Start with 20 cases

Pick the process you most want to automate. Ask the person who runs it to grade 20 recent cases with the right outcome and the reason, and ask a colleague to grade the same 20 without seeing the first set. If they disagree on more than two, define the process before anyone builds.

extendfuture's AI process automation method starts the same way, with a baseline of what the process costs today and graded cases to measure every change against.

Sources

• OpenAI, Introducing OpenAI Presence, 22 July 2026
• Carl Franzen, VentureBeat, OpenAI unveils Presence, a new platform that lets enterprises launch and manage realtime voice agents and chatbots, 22 July 2026
• Nate Herk, Don't automate an invention, 23 July 2026
• Anthropic, Demystifying evals for AI agents, 9 January 2026
• Shreya Shankar and colleagues, Who Validates the Validators?, arXiv, April 2024
• Hamel Husain, Using LLM-as-a-Judge For Evaluation: A Complete Guide, October 2024