The hours your people spend checking AI agents are the next big productivity win, and for most organisations they are also the easiest one to collect. On 14 October 2026, G2 published its 2026 AI Agents Report, based on a survey of more than 1,000 B2B software buyers alongside its own review data. The headline is encouraging: 66 per cent of organisations now run agents in limited or scaled production, adoption is up 25 per cent year on year, and four in five teams running agents in mature workflows report cycle-time gains of at least 11 per cent.
The report also gives a name to something many CRM teams will recognise. G2 calls it the distrust tax, which its chief innovation officer Tim Sanders describes as the ongoing cost of running agents you don’t fully trust. In practice it means people reviewing, correcting and approving agent work by hand, and the report suggests that this time is now large enough to measure.
That is good news, because time you can measure is time you can win back. The agents are already in production and already delivering. The teams pulling ahead are the ones turning oversight hours into capacity, by giving each agent a clear path to earn more autonomy as it proves itself.
What is the distrust tax on AI agents?
The distrust tax is the share of an AI agent’s expected productivity gain that is spent on human oversight instead: reviewing outputs, fixing mistakes, approving actions and re-checking performance. G2’s October 2026 research quantifies it. Thirty-one per cent of organisations say employees spend six or more hours a week on what the report calls bot-sitting, and 45 per cent say oversight and correction consume more than a quarter of the gains they expected. Forty-six per cent report higher operating costs than planned. The distrust tax is not a sign that agents fail. It is the cost of running capable agents without the controls, data and evidence that would let people step back with confidence.
Seen that way, the figure is an invitation rather than a verdict. Fifty-two per cent of organisations require a human quality review before an agent goes into production, and 77 per cent re-evaluate agent performance at least monthly. Those habits are sensible, and they are exactly why agents have reached production so quickly. The opportunity now is to make the checking smarter: review where the risk sits, automate the checks that a rule can perform, and let the review effort fall naturally as each agent builds a track record.
Oversight hours are the most recoverable cost in your AI budget
Measuring CRM ROI in the agent era means counting the time agents save and subtracting the time people spend supervising them. Start with a simple baseline for each agent: how long the task took before, how long it takes now, and how many minutes of review, correction and approval each output receives. Multiply by volume and you have a net figure finance can trust. Add the outcome measures the business already watches, such as response times, pipeline created, cases resolved or forecast accuracy. An agent with modest raw savings and little oversight can beat a flashier agent that needs constant checking, and that comparison is what tells you where to invest next.
The arithmetic is striking. Six hours a week across a working year comes to roughly 270 hours per person, close to seven working weeks. Licences, models and integration are largely fixed once an agent is live, but oversight time responds directly to good design. That makes it the most recoverable line in an AI budget, and the one where a few weeks of focused work show up fastest in the numbers. This blog looked at the measurement side in Measured AI Agents Are the Ones That Scale, and oversight minutes belong on that same scorecard.
Clean CRM data shortens every review
CRM data quality is important because every agent decision inherits it. An agent that drafts a renewal email, updates an opportunity or routes a case is only as reliable as the account, contact and history it reads. When records are complete, current and free of duplicates, agent outputs are right more often, so reviewers approve faster and correct less. When data is patchy, reviewers have to verify the facts as well as the judgement, which is where most oversight time goes. Good data quality also improves the reports, forecasts and routing rules people already rely on, so the investment pays back twice: once in everyday CRM work and again in how much autonomy agents can safely be given.
This is where many teams find their quickest win. Pick the agent with the heaviest review load, look at the last fifty corrections, and sort them by cause. Usually a handful of fields, a duplicate pattern or an outdated picklist explains most of them. Fixing those sources removes whole categories of correction at once, and the improvement holds for every agent and every person who touches the same records afterwards.
Earned autonomy turns review time into capacity
Earned autonomy is the practice of giving an agent more independence step by step, based on evidence. A new agent starts with every output reviewed. Once its accuracy on a task type holds above an agreed threshold for an agreed period, review moves to sampling, then to exceptions only, then to periodic audits. Higher-risk actions, such as changing prices, closing deals or writing to customers, keep a human approval for longer. G2’s own recommendations to vendors point the same way: make oversight costs visible and set clear escalation thresholds that allow more autonomy over time.
The platforms now make this practical. Dynamics 365 agents can pause for approval before sensitive actions, Salesforce’s newer sales agents show their reasoning and route judgement calls to the rep with an audit trail, and HubSpot lets administrators set limits per feature. The capability is there in all three, and the teams getting the most from it are writing down their thresholds explicitly, agent by agent and action by action, so that autonomy grows on a schedule everyone has agreed rather than on instinct.
What is shadow AI, and why do unmanaged agents need more watching?
Shadow AI is the use of AI tools, models or agents inside an organisation without the knowledge or approval of the teams responsible for data, security and governance. It ranges from employees pasting customer data into public chatbots to departments switching on agents in their own SaaS tools. G2 describes unmanaged agents as the new shadow IT, and its numbers show why the topic matters: 38 per cent of organisations with agents in production run more than 100 of them, and 47 per cent still manage them informally. The answer is not to slow adoption but to bring agents into view, with a named owner, known permissions and a record of activity for each one.
Visibility is also what makes earned autonomy possible. Gravitee’s State of AI Agent Security 2026 survey of 750 leaders found that agent counts roughly doubled in about four months while monitoring coverage moved from about 47 to 52 per cent, and only 7.2 per cent could name an individual responsible for agent behaviour. An agent nobody owns tends to be checked by everyone. Give each one an owner and a place in a shared registry, an idea explored in One List Makes Your AI Agents Compound, and the review effort concentrates where it belongs.
Trusted agents lift CRM adoption too
To improve CRM adoption rates, make the CRM the place where work gets easier. That means fewer fields to fill in by hand, records that are already up to date when a rep opens them, and clear guidance on the next step. Well-governed agents help on every count: they log activities, summarise calls and keep opportunities current, so sellers spend less time on admin and trust the data they see. Pair that with training built around real workflows, visible support from sales leadership and quick fixes to the friction people report, and adoption tends to rise steadily rather than through mandates.
Oversight design shapes this directly. When reps are asked to approve every small agent action, the agent feels like extra work and enthusiasm fades. When approvals are reserved for the moments that genuinely need judgement, the agent feels like a capable colleague, and people lean in. Sharing the accuracy figures behind each autonomy step helps as well. Reps who can see that an agent has drafted 400 follow-ups with a 97 per cent approval rate are far more willing to let it send the next batch on its own.
The Sirocco perspective
We see the same pattern across Salesforce, HubSpot and Microsoft Dynamics 365: the agents work, and the organisations getting the most from them are the ones that have made trust measurable. As an independent CRM partner, we do not sell licences or models, so our advice starts from the workflow and the data rather than a product roadmap. That matters here, because the right autonomy thresholds depend on your processes, your risk appetite and the quality of your records, not on which vendor built the agent. An independent view also makes it easier to compare how each platform handles approvals, logging and limits, and to apply one consistent policy across a mixed estate.
Our suggestion is a focused 90-day programme. In the first month, measure oversight minutes for your three busiest agents and trace the top causes of correction. In the second, fix the data sources behind them and write explicit autonomy thresholds for each action. In the third, move the agents that qualify to sampled review and report the hours returned to the business. Most teams find that a meaningful share of their bot-sitting time comes back within a quarter, and that the agents they trust most are the ones they then scale fastest. That is a very good position to start 2027 from.
If you would like help measuring the oversight load on your agents and setting thresholds that let them earn more autonomy, schedule a consultation with our team.
Get in Touch
If your team is spending more hours checking agents than you planned, we would be glad to help you work out which CRM workflows are ready to earn more autonomy, on Salesforce, HubSpot or Dynamics 365.
