The Verification Tax
Vishal Sachar
Co-Founder & CEO of CLRT
The numbers describing delegation to AI have stopped being soft. Anthropic's Economic Index, built on telemetry across a million conversations rather than on what people say in surveys, shows directive conversations, where a person hands the model a task and takes the result with minimal back and forth, rising from 27% of Claude.ai usage in late 2024 to 39% by August 2025. Automation-shaped usage has reached about half of consumer traffic and 77% of business usage through the API. Microsoft's own telemetry counts a fifteenfold rise in active agents inside Microsoft 365 in a single year. Delegation is real, it is deepening, and it is measured. What is not measured, in almost any organisation I speak to, is where the saved time actually went.
It is worth pausing on why this round of evidence is different from the adoption claims of the last two years. Surveys measure enthusiasm, and enthusiasm has been near universal for a while. Telemetry measures behaviour. When the same instrumentation that watched people co-write with a model through 2024 now watches them hand over the task whole, something structural has moved. Anthropic's June 2026 follow-up shows what the handover compresses: producing a blog post took a median of thirteen human turns in a chat interface and a single prompt in an agent harness, and 93% of conversations now produce an identifiable work artifact, a document, a dataset, a piece of code. The doing is collapsing into the instruction. On the production side of the ledger, the agent era has already arrived.
The other side of the ledger is the one nobody puts on a slide. The assumption under every AI business case is that hours saved doing return as free capacity. They do not. Google's DORA research programme, working through 1,110 practitioner responses published in March 2026, described what actually happens: the time saved generating is re-spent on verification and on shaping the next attempt, and teams with higher AI adoption report more cognitive load in code review, not less. The largest developer survey in existence agrees on the texture of that load. Of nearly 49,000 respondents to Stack Overflow's 2025 survey, 66% named AI output that is almost right, but not quite, as their top frustration, 45.2% said debugging AI-generated code is more time-consuming, and 75.3% still take an answer they do not trust to a human being. The week does not shrink. It converts, from hours spent producing into hours spent judging what was produced.
You might expect the teams furthest along to have automated the judging too. They have not. A survey of 1,340 agent builders, fielded in late 2025 and published this month by a vendor in the agent tooling market, worth reading with that in mind, found 57.3% now run agents in production, and yet the most common way those teams evaluate quality is human review, at 59.8%, ahead of the automated LLM-as-judge approach at 53.3%. Sit with that: the people who build agents for a living, with every tool and every incentive to automate the checking, still put a person at the gate. Which means every additional agent an organisation switches on adds output to a funnel that narrows to the same scarce resource, a senior person's hour of judgment. Delegation multiplies the producers. It does not multiply the reviewer.
This is why the span-of-control question of the agent era is not how many agents a team can run but how many it can afford to check. Call it the review budget: the verification hours your trusted people actually have, set against the verification cost of each workflow you delegate. Boards will soon ask for the headline ratio, agents per human, and the uncomfortable truth is that no public dataset on oversight ratios exists as of this writing. Anyone quoting you a market benchmark is inventing it. The disciplined answer is to design your own ratio, workflow by workflow, because the tax is not uniform. It moves with the blast radius of a mistake, with whether the action is reversible, and above all with whether wrong is objectively checkable before a person ever has to look.
Delegation multiplies the producers. It does not multiply the reviewer.
A deeper dive
The reason the tax is structural rather than a passing cost of immature models sits inside that Stack Overflow phrase, almost right, but not quite. Nearly correct work is the most expensive kind to review. An obviously bad draft is cheap to reject; a plausible one has to be read in full, checked against context the model never had, and traced for the one confident error hiding in competent prose. And the economics compound: verification cost scales with output volume, while judgment capacity does not scale at all. The agent harness that turns thirteen turns of doing into one prompt compresses the production and compresses none of the judging, so the binding constraint migrates quietly from making to checking. The obvious counter, use a model to judge the model, does not remove the tax either. It relocates it, because an automated judge is itself a system whose calibration has to be verified against human judgment before anyone can trust its verdicts, which is exactly why the teams running agents in production still rank human review first. What genuinely lowers the tax is unglamorous engineering: deterministic checks that run before human eyes so a person never reviews what a test could have rejected, review depth tiered to the blast radius of the action, sampling regimes for workflows that have earned them, and escalation thresholds decided in advance rather than under pressure. None of that is model work. It is the design of where judgment gets spent.
The missing ratio deserves more attention than it gets, because it will quietly define the next planning cycle. Microsoft's 2026 Work Trend Index, vendor research but built on 20,000 workers plus the company's own telemetry, found organisational factors explain 67% of realised AI impact against 32% for individual capability, which is a precise way of saying the constraint is the system around the model, not the person prompting it. The oversight ratio is that system's most important unset parameter. For a workflow where wrong is checkable by rule, reconciling invoices, validating a data pipeline, the right ratio might be hundreds of agent runs per sampled human review, because a deterministic check stands in front of the person. For anything that touches a customer, a contract, or money, the honest ratio today may be one to one, every output read before it leaves. The inputs to the design are knowable: the cost of a miss, the reversibility of the action, the availability of an objective check, and the loaded price of the reviewing hour itself. What is not knowable is a shortcut. Treating the ratio as one number for the whole organisation is how the review budget gets silently overspent, because the verification load lands, by default, on the two or three people whose judgment the business trusts most, which is precisely where the hours are most expensive. The backlog that forms in front of them looks like slowness. It is actually an unfunded tax.
Key terms
- Verification tax
- The hours an organisation re-spends reading, checking, correcting and re-prompting AI output; the conversion of time saved doing into time spent verifying.
- Review budget
- The finite verification hours an organisation's trusted judgment can supply; the real ceiling on how much delegated output it can safely absorb.
- LLM-as-judge
- Using one model to grade another model's output. It scales cheaply, but its calibration must itself be verified against human judgment before its verdicts can be trusted.
Work with CLRT
This is the work CLRT does. We do not install agents and hope the checking sorts itself out. We design the review budget with you: which workflows an agent may run behind a deterministic check, which need a sampled human review, which must never ship without a person holding the pen, and what the oversight ratio should be for each, in numbers you can defend to a board. If your organisation is saving time everywhere and somehow has less of it, that is the verification tax going unbudgeted. Bring us the workflow where the queue is longest, and we will design the ratio that clears it.
Vishal Sachar
Vishal Sachar is the Co-Founder and CEO of CLRT, where he helps UAE businesses make sense of applied agentic AI and put it to work. He writes on agentic systems, AI governance, and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.
