Delegation Rate

What It Is

Delegation rate is the proportion of a team’s productive output that is produced by an AI agent rather than by a human. It is measured against whatever unit your work produces: output tokens for a typed-text business, pull requests for an engineering team, drafts for a legal team, contracts processed for a finance team. The metric stays the same. Take the share of that output your team’s AI tooling now generates and divide by the total.

OpenAI’s June 25, 2026 internal-data report (“How Agents Are Transforming Work”) established the first publicly visible delegation rate at a frontier lab, citing Codex producing more than 85% of the average employee’s output tokens across departments including Legal, Finance, and Recruiting. Before that publication, “AI is doing more of my work” was a vibe. After it, the same sentence became a number. The phrase delegation rate names the number so it can be tracked, compared, and contested.

How It Actually Works

The rate is simple in form, awkward in practice. You count what your team’s AI agent produced this week, you count the total your team produced, you divide. The trouble is the denominator. Output tokens are easy when everyone uses one assistant logged into one account. They get harder when half the team uses Claude inside their IDE and the other half copies prompts into ChatGPT on a personal phone.

Most teams measuring delegation rate today triangulate from three sources: enterprise usage dashboards, individual self-reporting, and Git or document history that flags AI-touched artifacts. There is no industry-standard accounting yet. OpenAI’s report used output tokens as the unit because tokens are what its agentic-coding product emits and what its billing meters anyway. A consultancy might use billable hours of drafts produced. A sales team might count outbound emails. The right denominator depends on what the team gets paid to produce.

Why It Matters Right Now

For most of the past year, the frontier-lab pitch was that AI agents would replace specific kinds of work soon. The frontier labs were not measuring the replacement themselves. They were selling the future of it. The June 25 OpenAI report changed that. The company now publishes the proportion of its own internal work that runs through Codex, including the parts not done by engineers. Once one company at that altitude publishes the metric, every other software company will be asked the same question in board meetings, hiring conversations, and investor calls.

The implication is that delegation rate becomes a hiring signal. A team with a low rate looks slow. A team with a high rate looks under-managed. The middle is the new operating zone, and “middle” has no agreed-on definition yet.

The Cost / Tradeoff

A high delegation rate is not automatically good and a low one is not automatically bad. The cost shows up two places. First, in agent-loop-cost, since each delegated task burns tokens and inference time even when the agent’s output gets discarded. Second, in review burden, since output produced by an agent still needs to be read, verified, and signed off on by a human if it leaves the team. The shadow cost of a 90% delegation rate is a 90% review load, and review at that scale is its own role.

How TWO Uses It

TWO treats delegation rate as a leading indicator that runs ahead of headcount changes. When Scott reviews a small business or an internal team, the first question is no longer “how many people do you have.” The first question is “what fraction of your weekly output now runs through an agent you are not paying a person to do.” The number tells you what the team should be hiring for, what they should stop hiring for, and where their hiring bar should rise.

A team with a 15% delegation rate has hiring leverage. A team with a 60% delegation rate has training leverage. The work they have left is harder to teach and harder to delegate. A team at 85%, OpenAI’s own internal benchmark for a typical worker per the June 25 report, is in a new operating mode where the role’s job description is mostly editing and witnessing the agent’s output. That is not failure. It is a different role, and the role description needs to say so.

A Concrete Operator Scenario

You run a six-person marketing team. Three of your writers have quietly shifted to drafting outlines in Claude and finishing prose in ChatGPT. Two have not. Your weekly drafts are up, your meeting load has not changed, and you are about to back-fill an open position. Delegation rate is the number that decides whether to back-fill at all.

If three of your six writers are now operating at a 70% delegation rate, your team has roughly one full-headcount of latent capacity sitting inside the existing payroll. You promote one writer to a senior-editor role focused on agent oversight. You do not back-fill. You reallocate the savings to the part of the team where delegation rate is structurally low. The decision is not “fire one writer.” The decision is “stop hiring the writer you would have hired before measuring.”

What to Watch Next

Three signals tell you delegation rate is changing inside an industry. A publicly traded company breaks the metric out in an earnings call. A regulator asks an enterprise customer for it during a vendor-risk review. A labor union mentions it in a collective-bargaining position. The first will normalize the disclosure. The second will turn it into a compliance number. The third will turn it into a wage conversation. All three are plausible inside 2026.

Related Concepts

ai-agent, agentic-coding, agent-loop-cost, default-model, token.