AI will automate the tasks whose outputs are cheaper to verify than to produce. The larger that gap, the sooner the task yields.
That is a better forecast than a list of threatened job titles. Model ability matters, but ability becomes dependable automation only when a system can reject a bad answer quickly, cheaply, and with enough confidence to try again. Without that check, AI can still draft. It cannot safely own the result.
Three questions sort the work. Can a model produce a plausible answer? Can the workflow tell a good answer from a bad one? Can the business afford the mistakes that pass the check? A yes to all three puts a task near the front of the automation queue. Most forecasts stop after the first question and overstate what is coming.
Tasks are automated before jobs
A job title is too large a unit to predict. A job is a bundle of tasks with different rules.
A bookkeeper may enter transactions, reconcile accounts, investigate exceptions, explain a discrepancy to an owner, and decide how to classify an unusual purchase. Data entry and reconciliation have crisp checks. The explanation draws on context. The classification may carry legal or tax consequences. One role contains work that can run unattended, work that benefits from a draft, and work that keeps a responsible person in the decision.
The International Labour Organization reached the same level of analysis in its 2025 exposure index. It scored nearly 30,000 occupational tasks rather than declaring whole occupations automatable. Its conclusion was that job transformation is more likely than wholesale replacement because most occupations still contain tasks that require human input.1
The economic question is how much of each role consists of outputs that a machine can check. A role made mostly of those outputs changes quickly. A role with one such output changes less.
The verification ratio predicts what goes first
For any task, compare the effort required to produce an acceptable output with the effort required to check one.
A report that takes four hours to assemble but five minutes to reconcile against source data has a large verification advantage. A strategy memo that takes an hour to write and nearly an hour of senior attention to judge has almost none. AI can make both drafts faster. Only the first removes much labor from the full workflow.
This gives a useful directional ratio:
Automation pressure rises as production effort rises and trustworthy verification effort falls.
The ratio predicts where a new model capability will become economically useful. It does not say whether today’s model can perform the task, or on what date a company will adopt it. Capability sets the earliest possible date. Verification determines which capable tasks can run reliably at scale.
| Task | Available check | Likely result |
|---|---|---|
| Move invoice fields into an ERP | Schema, totals, duplicate rules, source record | Automation with exception review |
| Route a support ticket | Policy rules, confidence threshold, reopen and escalation data | Routine cases automate first |
| Draft a contract clause | An attorney must reason through the same risks | Drafting gets faster; review remains |
| Choose a product strategy | Revenue arrives months later and has many causes | AI helps with analysis; a person owns the decision |
The first two have a short correction loop. The last two can produce polished work, but polished is not a quality signal.
A useful verifier has four properties
Jason Wei’s verifier’s rule describes the traits that make a task easier for AI to learn: agreement on what good means, fast checks, checks that run at scale, and a signal closely tied to actual quality.2 The same traits work as a business test.
Agreement. Two competent reviewers reach the same answer from the same evidence. A valid JSON object has this property. “A strong point of view” usually does not.
Speed. The result arrives soon enough to correct the current attempt. A compiler returns an answer in seconds. Quarterly revenue arrives too late to guide each sales email, even if revenue is the outcome the company cares about.
Coverage. The check catches the failures that matter. A schema can prove that an invoice total is formatted as a number. It cannot prove that the number matches the invoice. Format validation alone creates a crisp signal with very little truth in it.
Scale. The check is cheap enough to run on every output. A senior reviewer may be accurate, but a review that consumes the time saved by generation supports assistance rather than automation.
One more variable determines whether the output can be released without a person: the cost of a false pass. A low-cost classification error can be corrected. A bad payment, medical instruction, hiring decision, or access-control change has a different threshold. High stakes keep a person or a separate control at the release point until the verifier earns enough trust. The Managing Risk guide covers that release decision in more detail.
That produces a simple three-gate test:
| Gate | Question | What failure means |
|---|---|---|
| Generation | Can the model produce a candidate? | The task waits for better capability |
| Verification | Can the workflow reject a bad candidate cheaply and quickly? | The task remains human-reviewed |
| Release | Is a false pass cheap or contained? | The output needs approval before it acts |
A demo proves the generation gate. Production automation requires all three.
AI benchmarks favor work with built-in checks
Capability headlines make the future look less predictable than it is because their tasks are unusually easy to score.
METR’s task-completion time-horizon research measures agents on more than a hundred software, machine-learning, and cybersecurity tasks. The tasks are self-contained, well specified, and designed around clear success criteria. METR also warns against translating those results into a claim about whole jobs: real work depends on tacit context, interaction with people, and success measures that cannot always be scored by an algorithm. Agent performance drops when evaluation becomes holistic rather than mechanical.3
That caveat explains a lot. The benchmark is measuring work that already has a verifier. Lengthening the task from minutes to hours is real capability progress inside that category. It says much less about an eight-hour management decision whose quality becomes visible next quarter.
Actual usage is concentrated too. In Anthropic’s November 2025 Economic Index sample, computer and mathematical tasks accounted for 46 percent of first-party API traffic, while the ten most common tasks accounted for 32 percent.4 That vendor-specific sample does not prove the verifier rule. It is consistent with a market finding repeatable value in a narrow set of digital tasks before capability spreads evenly across work.
Business outcomes are often too slow to verify the work
“Did the deal close?” is a scorecard. It is usually a poor verifier for a particular sales email. The answer arrives weeks later and mixes the email with price, product fit, timing, the salesperson, and the buyer’s budget. The signal is delayed and noisy.
The same problem appears in underwriting. Whether a loan performs matters, but the answer may take years. An automated underwriting step needs closer checks: required evidence is present, calculations reproduce, policy rules pass, similar historical cases behave within expected bounds, and exceptions reach someone with authority.
This distinction separates measurement from verification. A metric tells leadership whether a process is working. A verifier tells the process whether this output is acceptable while correction is still cheap. Good automation needs both.
Businesses can move a task forward by building the check
Verifiability is partly a property of the task and partly a product of process design. Code became friendly to automation because people spent decades building types, tests, linters, sandboxes, review rules, and rollback systems.
The same work can happen elsewhere. A vague request to “review this expense report” becomes more verifiable when every amount links to a receipt, arithmetic reconciles, policy limits are encoded, and unusual cases route to a named owner. AI may still misread a receipt. The workflow now has several ways to notice.
Long tasks often become tractable when the process checks intermediate outputs. Waiting until the end of a ten-step agent run asks one reviewer to reconstruct the whole path. Checking the source list, calculations, permissions, and proposed action at their own steps makes each failure smaller and cheaper to correct. Research on process verifiers follows the same logic: feedback at intermediate reasoning steps can be more useful and compute-efficient than a single final score.5
A rubric helps only when applying it is cheaper than doing the work. “Accurate, strategic, and on brand” is a description of taste. A source requirement, numerical bound, prohibited claim list, or reconciliation rule can reject an output. A model grading its own prose against a vague rubric has no independent standard to apply.
The pattern is predictable
The next work to automate will be digital, repeated, and surrounded by fast checks. Structured back-office processing. Code changes with tests. Extraction tied to source records. Reconciliation. Monitoring. Scheduling against explicit constraints. Variants produced against a measurable specification.
Work built on taste, causal judgment, tacit history, physical presence, interpersonal trust, or formal accountability will change differently. AI will prepare material, surface options, and remove routine steps. A person will remain at the point where the organization decides what is true, acceptable, or worth the risk.
For a business, the useful inventory is one line per recurring output: the time spent producing it, the acceptance check, the time and delay of that check, the cost of a false pass, and the owner of exceptions. Sorting that inventory by verification advantage gives a more credible automation plan than ranking departments by payroll. It is the task-level companion to Recognizing Leverage, which finds the people already doing this well.
Model releases move the capability line. Verification determines what the new capability can attach to. If a task has a cheap, fast, high-coverage check and a safe path for exceptions, it is close. If checking takes as much judgment as producing, the task will stay in human hands longer no matter how fluent the demo looks.
Footnotes
-
International Labour Organization and NASK, “Generative AI and Jobs: A Refined Global Index of Occupational Exposure”, May 2025. ↩
-
Jason Wei, “Asymmetry of verification and verifier’s rule”. ↩
-
METR, “Task-Completion Time Horizons of Frontier AI Models”, updated May 2026. ↩
-
Anthropic, “Economic Index: Economic primitives”, January 2026. The report analyzes sampled Claude.ai and first-party API traffic, not the whole AI market. ↩
-
Amrith Setlur et al., “Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning”, ICLR 2025. ↩