The 'Distrust Tax': G2 Finds AI Agents Are Working, but Babysitting Them Is Eating the Gains

On October 14, G2, the B2B software review company that serves more than 200 million buyers a year, published its 2026 AI Agents Report, titled "Crossing the Agentic Chasm." The report is based on a survey of 1,019 B2B decision-makers fielded in July 2026 across North America, EMEA, and APAC, plus more than 2,500 reviews from G2's AI Agent Builder category. Its headline finding is that the corporate AI agent era has moved decisively out of the lab: 86 percent of B2B software buyers purchased AI agent software or services in the past year, up 25 percent year over year from 69 percent, and 66 percent now have agents running in limited or full production.

The celebration stops there. The same survey finds that nearly half of those buyers are losing a significant share of the expected productivity gains to the cost of watching, checking, and fixing the agents' work. G2 calls this erosion the "Distrust Tax," and it may be the defining management problem of the agentic workplace: the technology works well enough to deploy, but not well enough to be trusted.

From pilots to production

The report documents one of the fastest enterprise technology ramps in recent memory. Technology and retail companies showed the biggest jump in budgets of five million dollars or more, while healthcare leads on absolute spending, with 57 percent of healthcare organizations spending at least one million dollars a year on agents. Among organizations using agents in high-volume workflows, four in five report cycle-time improvements of at least 11 percent, and nearly half report gains of at least 25 percent.

That performance is now shaping purchasing decisions across the entire software stack. Eighty-four percent of buyers say agent capabilities will influence their 2027 software choices, and 34 percent call them a primary deciding factor. Agents are no longer a feature buyers ask about; they are becoming a criterion buyers select on.

The Distrust Tax

Here is where the numbers turn sobering. Nearly one-third of B2B buyers (31 percent) say employees spend at least six hours each week overseeing agent work, a practice the industry has taken to calling "bot-sitting." Forty-five percent report losing more than a quarter of their expected productivity gains to oversight and rework, and 46 percent say agent operating costs came in higher than expected.

"AI agents are showing real signs of value, but much of it is being erased by oversight, rework, and exception handling," said Tim Sanders, Chief Innovation Officer at G2.

The burden scales with size. Among large enterprises, 27 percent say their agent deployments require significantly more oversight than planned, nearly double the 15 percent rate reported by small and midsize businesses. Bigger fleets mean bigger supervision bills, which is why the report argues the trust problem, not raw capability, is now the bottleneck: absent trust, agents stay confined to isolated use cases and mainstream adoption stalls. For agents to cross the chasm, buyers need repeatable proof of credibility, not more marketing promises, plus the controls that make autonomy safe to extend.

Shadow agents: the new shadow IT

Compounding the problem is a visibility crisis. Thirty-eight percent of organizations with agents in production now operate more than 100 of them, yet only 23 percent say agent ownership and controls are centrally governed. Almost half (47 percent) manage agents informally or leave ownership to individual departments. Just 27 percent maintain a central registry of their AI agents, only 18 percent use automated discovery to find sanctioned and unsanctioned agents, and a mere 14 percent have real-time observability into unusual behavior, costs, or risks.

G2 frames this as the arrival of "shadow agents," the agentic successor to shadow IT. Employees can spin up an agent with a credit card and a prompt, and central IT often finds out months later. Central IT teams and AI, data, and machine learning groups now manage 49 percent of agent operating costs and carry 60 percent of governance accountability, but the discovery and inventory tooling has not caught up with deployment speed. Every unguided agent is a potential security, compliance, and cost incident waiting for an owner.

The workforce story is better than feared

For all the anxiety about automation and jobs, the report's labor findings lean toward augmentation rather than replacement. More than eight in ten buyers (83 percent) say agents improve employee satisfaction by taking over repetitive work. Thirty-five percent say they have redeployed employees into higher-value roles. Just 14 percent report reducing headcount as a result of agent adoption.

That 14 percent figure is worth pausing on. In a year when headlines have been dominated by AI-driven layoffs, the buyers actually deploying agents at scale are mostly using the freed-up time to move people up the value chain, not out the door. The emerging pattern is not replacement but reallocation, though the report does not track whether those redeployed roles come with the same pay, which is the question workers will ask next.

The seat license is dying

Agents are also breaking the software industry's favorite pricing model. Only 9 percent of buyers consider per-user, per-seat pricing the fairest way to pay for AI agents. Pay-as-you-go models, including consumption pricing and payment per completed action or task, rose from 38 percent in 2025 to 57 percent in 2026. Seventy-two percent of buyers say a strong ecosystem of connectors, actions, skills, and marketplace offerings is very important or a must-have. And 56 percent expect agents to be orchestrating work across their entire SaaS stack by 2030, which would make the agent layer, not the application, the center of gravity in enterprise software.

What comes next: proof over promises

Perhaps the most telling finding is how buyers now vet agent vendors. Half of recent buyers use their own data science teams to verify vendor claims. Four in ten ran hands-on tests or proofs of concept, made reference calls, consulted analyst evaluations, or used independent benchmarks. Before production, 52 percent require human quality review, 48 percent require security testing, and 41 percent require legal, risk, or compliance review. After deployment, 77 percent re-evaluate agents at least monthly.

That rigor reflects a maturing market. Sixty-seven percent of buyers say they have high trust in their organization's AI systems, yet only 7 percent are comfortable with full agent autonomy; 48 percent prefer agents with guardrails and 29 percent want them limited to specific tasks. The message to vendors is clear: the race is no longer about who has the smartest agent, but who can prove their agent behaves. The companies that build registries, observability, and evidence of reliability first will be the ones whose agents get promoted from pilot to production without paying a crippling tax.