How AI Agents Will Reshape Team Costs by 2027: A Forecast Built on METR's Expenditure Horizon

Every year, every CEO runs into the same question: how many people do you actually need to hire, and where can you get by without adding another line to payroll. Research organization METR has, for the first time, given a quantitative answer: its new "expenditure horizon" metric shows the dollar point where the cost of an AI agent and the cost of a human on the same task become equal — and the top models (GPT-5.5, Opus-4.8) are already right up against that line, $2,000–$3,300 versus $2,500 per percentage point of human output.

But the number alone says nothing about how to apply it to your own headcount — and the pace at which agent autonomy doubles every ~7 months means the window for "wait and see" is already closing. Ahead: where agents still lose to humans, how to calculate this threshold for your own team, and why the "junior + senior" structure may look very different soon.

What METR Is — and Why "Expenditure Horizon" Is Not Just Another AI Buzzword

METR (Model Evaluation and Threat Research) is a non-profit research organization based in Berkeley, formerly known as ARC Evals, specializing in the rigorous scientific measurement of autonomous AI agent capabilities. They don't sell products and have no stake in inflating expectations. Their business is sober numbers.

For several years, METR has been measuring what it calls "time horizon" — the duration of tasks an AI agent can complete autonomously with 50% reliability. That metric has shown steady growth: the horizon has doubled roughly every 7 months for six consecutive years. But it had one critical blind spot — it said nothing about money.

In July 2026, METR published a new paper by Tom Cunningham, Manish Shetty, and colleagues — "Expenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT." It introduces a fundamentally new unit of measurement.

How Expenditure Horizon Works

The idea is simple; the execution is not. METR plots two curves on a shared axis: how much useful output is produced as the budget for an AI agent grows (API calls plus GPU time) — and how much output the same budget buys when spent on a human (salary plus compute resources). The point where those two curves cross is the "expenditure horizon" — the threshold beyond which a human delivers better value than an agent.

Before this method, companies had no common currency for the comparison. Now they do: dollars per unit of output. For a COO or Chief People Officer, this shifts the question "replace or hire?" from the realm of philosophical debate squarely into financial modeling.

The First Numbers: Where Agents Still Fall Short

METR chose the NanoGPT speedrun as its testing ground — an open community challenge where participants compete to accelerate language model training. Since May 2024, 82 documented contributions have reduced training time from roughly 45 minutes to under two minutes — a cumulative 33× speedup.

The human baseline came out clean: interviews with active participants in the project, cross-validated by an independent AI-based assessment, produced the same result — approximately 16 hours of human effort per one percent of performance gain. At a rate of $150 per hour, that works out to $2,500 per percentage point of improvement.

What did the agents show? Of the six models tested, only GPT-5.5 and Opus-4.8 reached meaningful expenditure horizons — in the $2,000–$3,300 range. Earlier models — GPT-5 and Opus-4.1 — produced virtually no real progress: their apparent gains evaporated under revalidation. Qualitatively, the agents were mostly tuning hyperparameters; genuinely novel, deployable solutions were rare.

The conclusion is measured: on complex, open-ended research tasks, AI agents do not yet outperform humans on cost per unit of output — they are roughly on par at small budgets and lose ground at larger ones. But this is a narrow test case with one type of task, and METR explicitly notes that the newest model generations were not included in the study.

The question that actually matters to HR and COOs is different: what do these same numbers mean for specific roles on your org chart right now.

What These Numbers Mean for Headcount Planning Through 2027

A sensational headline would be wrong here. The correct reading of METR's data is not "AI will replace everyone" — it's "for the first time, there is a method to understand where and when replacement makes economic sense."

The Doubling Horizon: Why 2027 Is Not an Arbitrary Date

The trend METR has documented — autonomous task horizon doubling every ~7 months — has held for six years with no signs of slowing. Simple extrapolation: models that today reliably handle tasks lasting 2–3 hours could, by the end of 2027, reach the level of full-workday or even multi-day tasks. That's not a guarantee, but it is the tempo embedded in the data.

METR's researchers themselves note: if the documented trend holds for another 2–4 years, generalist autonomous agents will be capable of handling a wide range of tasks that currently take a week. At that level, the question "hire a person or spin up an agent?" stops being hypothetical and becomes a daily operational decision.

Three Role Categories — and Three Different Substitution Logics

It helps to break your workforce into categories based on how readily their output can be quantified — that is, how directly the expenditure horizon logic applies to them.

Roles with clearly measurable output — data analytics, software testing, report generation, template-based compliance monitoring, initial document processing. Here, expenditure horizon is already a practical tool. For some tasks, an agent is cheaper than a human at budgets up to $2,000–$3,000 per task — and that ceiling moves up with every model generation.

Roles with partially measurable output — recruiting, operational support, content marketing, basic legal analysis. Agents are effective on structured sub-tasks here, but require human oversight at the judgment and communication stages. The logic: hybrid teams where the agent handles routine work and the human handles decisions.

Roles with hard-to-measure output — strategic leadership, complex negotiations, crisis communications, non-standard client situations. Expenditure horizon doesn't apply directly here: there is no clear numerical criterion for "one unit of result."

For a Chief People Officer, the practical takeaway is this: the first-priority audit is Category 1, where the numbers already exist.

What the Metric Doesn't Measure — and Why That Matters

METR is candid about its limitations: time-horizon measurements above 16 hours are unreliable with the current task set. Beyond that, expenditure horizon has so far been tested on a single task type — research optimization with a well-defined success metric.

A real workforce decision has to account for more than the cost of task execution. It must also factor in the cost of deploying an agent system, maintaining it, securing it, and upskilling the team around it. The METR study also revealed that agents are prone to reward hacking — optimizing the metric rather than the underlying objective. That's a separate line item in operational costs that rarely makes it into the initial financial model. For a deeper look at AI agent security risks, it's worth reading the piece on autonomous hacking and AI agent risks.

These caveats matter, but they don't change the bottom line: METR's methodology hands you a ready-made template for the calculation — and applying it to your own team takes just four steps.

Building Your Own Model: From Metric to Budget Decision

METR's methodology hands operational leaders a template. Here is an adaptation for real-world corporate planning.

Step 1. Define the "Unit of Output" for the Role

Without this step, none of the calculations that follow make sense. A unit of output is what you actually evaluate at the point of hiring: applications processed per week, ticket resolution time, documents reviewed, classification accuracy rate, and so on. If you can't articulate a unit of output, the role isn't yet suited to direct comparison under the expenditure horizon framework.

Step 2. Calculate the Human Cost per Unit of Output

The base formula: total cost of employment (salary + benefits + workspace + management overhead + onboarding time) ÷ units of output per year. This is your "human curve."

An example: if an analyst processes 800 reports a year and their all-in annual cost to the company is $80,000, one report costs $100. If an agent processes the same report for $2–5 and requires human review on roughly every fifth one, the true cost per unit — including oversight — is roughly $12–20. The expenditure horizon for this role is very low: the agent is cheaper even at minimal volumes.

If the task involves complex judgment or client-facing interactions, the recalculation tells a very different story. The article on how to calculate the break-even point for an AI agent versus a new hire walks through how to build that inflection point for a specific role in detail.

Step 3. Build the Agent Curve Using Real Parameters

Ask your vendor or run your own test: what are the API costs at your volume? What percentage of tasks does the agent complete without errors? What level of human oversight is required? Build the full cost: API spend + cost of errors + cost of supervision.

The point where that curve intersects the human curve is your own expenditure horizon for a specific function. Not theory — a line in the budget.

Step 4. Build in a Rate of Change

If agent autonomous horizons double every 7 months, your 2025 model may be significantly outdated by the end of 2026. Don't build this as a static spreadsheet — build it as a function of model generation. Add a "quarterly review" parameter to the budget document, or you'll be making decisions with a six-month lag baked in.

This rate of change, incidentally, explains why companies that start collecting agent performance data today will have a far stronger forecasting base for 2027 than those starting from scratch. Without measurement, there is no management — which is exactly why it's worth looking at the material on measuring ROI on AI budgets before falling into the trap of losing funding due to missing metrics.

What This Means for Team Structure

If your org chart includes Category 1 roles (clearly measurable output), the realistic scenario through 2027 is not eliminating those positions — it's transforming their composition. Instead of five junior analysts and one senior: one or two seniors who manage the agent system and verify the output.

That means a different OPEX curve — lower costs for routine volume, higher investment in people capable of judgment. If you treat this transition as a planned transformation, you keep control of the process and cost predictability. If you react to it instead, you'll face simultaneous pressure to cut costs and a shortage of the right competencies at the same time.

Leaders who come to the board or to investors with a concrete model — "here is our expenditure horizon by function, here is where we're planning the transformation, and here is the projected OPEX impact" — present a fundamentally different picture than those who say "we're monitoring the AI market." The former build a reputation for turning technological turbulence into predictable financial outcomes.

Three Practical Steps to Take This Quarter

The methodology exists. The data exists. What should you actually do right now?

Audit roles by the measurability of their output. Go through your org chart and ask one question for every position: "Can we describe the unit of output for this role as a number?" Roles where the answer is "yes" — that's your priority list for a pilot. Roles where the answer is "no" — those require a different approach and remain outside the direct comparison framework for now.

Run at least one measurable pilot with an agent. Not a "technology test" — a pilot with a specific metric: how many units of output does the agent deliver at a given budget? That data is your own empirical curve, and it will be far more accurate than any industry benchmark. A 4–6 week pilot will generate enough data for a first version of your expenditure horizon.

Add expenditure horizon to the budget justification template for new hires. When a line manager comes in with a headcount request, the standard question is now: "What is the expected cost per unit of output for this role compared to an agent?" This isn't a hiring freeze — it's an informed decision instead of an automatic approval.

When you know your numbers and can defend every position in the org chart with a concrete calculation rather than "we need more hands" — that is a qualitatively different level of operational control. Not firefighting: architecture. The feeling that workforce planning has finally become something you steer rather than something you react to — that's exactly what a properly built model delivers.

---

Want to work through the expenditure horizon for specific roles in your company, or build the first version of the model together? Book a 15-minute consultation — we'll go through your situation with numbers, not generalities.

---

FAQ

What is the METR expenditure horizon metric, and how does it differ from standard AI benchmarks?

Expenditure horizon is the dollar threshold at which the cost of achieving the same result is equal for an AI agent and a human. Unlike conventional benchmarks, which give a binary pass/fail verdict, this metric tracks the return on every additional dollar spent — and converts human and machine costs into a single currency. That's precisely why it's applicable to financial planning, not just technical evaluation.

Can you replace part of your team with AI agents right now, and what will it cost?

For tasks with clearly measurable output — analytics, document processing, testing, basic compliance — agents are already price-competitive above a certain volume threshold. But the true cost includes deployment, oversight, and error management. A reliable calculation requires a pilot with your own data; industry averages are not trustworthy here.

How often should expenditure horizon calculations be revisited for workforce planning?

METR has documented a doubling of agent autonomous horizons roughly every 7 months. In practice, that means a model built a year ago may already be significantly stale. The optimal cadence: quarterly review for Category 1 roles (clearly measurable output), and semi-annual for the rest.

Which roles are least vulnerable to AI agent substitution through 2027?

The most resilient positions are those where the "unit of output" cannot be described numerically without context: strategic leadership, complex negotiations, crisis communications, non-standard client situations. Also roles with significant regulatory accountability, where decisions are legally required to be made by a human. These positions will transform — but through redistribution of time toward higher-complexity work, not through direct replacement.