The Distrust Tax: Why Your AI Agent's ROI Is Leaking Into Bot-Sitting

Quick answer. Two-thirds of B2B software buyers now run AI agents in production, yet 45% say oversight and rework consume more than a quarter of the gains they were promised.[1] G2 calls this the Distrust Tax. It is a governance gap, not a model problem — and you close it by putting the agent's approval and audit trail in your system of record.

The number that indicts the pitch

Adoption stopped being a forecast this year: 86% of B2B software buyers bought agent software or services in the past twelve months, and 66% now run agents in limited or full production.[1]

Then the same survey measured the other side of the ledger. Nearly half of those buyers report losing more than a quarter of expected gains to oversight and rework, and 46% say operating costs came in higher than planned.[1]

The agents work. The business case still leaks.

What bot-sitting actually looks like

"Bot-sitting" sounds trivial. It is three jobs nobody budgeted for.

Exception handling: the agent gets most of a workflow right and drops the remainder on a human with no context. Re-verification: a manager re-checks completed work because the mistake, if it lands, is expensive. Rework: a wrong figure travels two steps downstream before anyone catches it.

Nearly a third of buyers say staff spend six or more hours a week supervising agent output, and larger organisations feel it worst — 27% need significantly more oversight, roughly double the SMB rate of 15%.[1]

Why the tax never appears in your business case

Most agent proposals price the licence, the integration and the training. Almost none price the supervision. That omission is why pilots look excellent and renewals disappoint.

The arithmetic gets skipped. An agent that saves a team 20 hours a month looks like a win. If checking its output takes six hours a week, that is 26 hours of supervision a month — net negative by six hours, before the licence.[1]

Supervision is a real cost at a real hourly rate. Until it sits in the ROI model, every agent you deploy borrows against a number nobody calculated.

The tax is a governance problem, not a model problem

If oversight were technical, a better model would fix it. The governance data says otherwise.

Almost half of buyers admit agent ownership and controls are informal or department-led, against 23% centrally governed. Only 27% keep a central agent registry, and just 14% have real-time observability into what agents do, spend or touch.[1]

The market has noticed. Creatio's survey of 600-plus IT and business leaders found organisations with established AI governance are more than four times as likely to feel very confident scaling AI — governance as an accelerant, not a brake.[2]

Capgemini data cited by TGVP puts 79% of companies adopting agents against only 2% deployed at scale, and names governance the primary gate: identity systems were built for people who log in and out, not for ephemeral, non-deterministic agent identities needing continuous authorisation.[3]

Your access model was designed for humans with shifts and managers. An agent has neither.

Where the tax gets paid down

You cut the tax by removing the need to re-check work — not by asking humans to check faster. Four things, per workflow:

1. A named approver and a threshold. A purchase order above a set value, a payroll adjustment, a journal entry, a customer refund: each needs a decision class, a limit, and an accountable person.
2. An exception queue that carries context. When the agent stops, the human should get the record, the reason and two options — not a blank prompt.
3. An audit trail that survives scrutiny. Who or what changed it, on what basis, approved by whom. If that record lives in a chat thread, you have no record.
4. Scope matched to appetite. Only 7% of buyers are comfortable with full autonomy, 48% want guardrails, and 29% want agents limited to specific tasks.[1] Design for the middle of that distribution.

This is the practical argument for the unified record. Whether you consolidate onto one platform or build an intelligence layer over legacy systems, approvals and audit trails must land where finance and operations data already lands. A governed ERP provides that; a spreadsheet plus a clever agent cannot.

The honest counterweight

This is not a story about agents failing. G2 also found 83% reporting better employee satisfaction, 35% redeploying people into higher-value roles, and only 14% reporting headcount reduction.[1] Creatio agrees on direction: 9% expect significant cuts, while 89% expect agents to expand capacity, augment teams or create roles.[2]

One caveat: G2 is a software marketplace surveying its own review community, and Creatio sells a workflow platform. Both datasets are self-reported, so two surveys agreeing makes the direction credible but the magnitudes indicative.

Bottom line

The Distrust Tax is the most under-budgeted line item in enterprise AI, and the reason successful pilots fail to scale.

Run a one-week audit before your next deployment. List every agent in production. Log the hours a human spends checking each one. Multiply by a loaded hourly rate. Then decide, per agent, whether to widen, narrow or switch it off.

If you cannot answer where the approval lands and what the audit trail says, you have not deployed automation. You have deployed an unsupervised employee with no signature authority — and you are paying a manager to shadow them. That is a governance problem, and unlike a model problem, it is fixable this quarter.

Related reading: How SMEs Can Measure AI ROI Without a Finance Degree, ERP Project Governance and The Phantom ERP Investment Gathering Dust.

Sources

[1] https://www.prnewswire.com/news-releases/new-g2-research-ai-agents-have-moved-from-pilot-to-production-but-oversight-comes-at-a-cost-302908654.html — G2 2026 AI Agents Report — press release, PR Newswire, 14 Oct 2026
[2] https://www.prnewswire.com/news-releases/creatio-releases-the-state-of-enterprise-ai-adoption-report-an-in-depth-look-at-where-enterprise-ai-stands-302913541.html — Creatio — The State of Enterprise AI Adoption Report, press release, PR Newswire, 20 Oct 2026
[3] https://finance.yahoo.com/technology/ai/articles/tgvp-infrastructure-not-model-quality-100544737.html — Blair Hayes — Infrastructure, Not Model Quality, Is the Primary Bottleneck to Enterprise AI Agent Adoption, Yahoo Finance, 19 Oct 2026