Within a week of launch, 127 of 150 account executives were using the AI agent I studied. A month later, more than half of them still had not fully given up their old dashboards.

Both facts are true, and together they say more about trust in AI than either one does alone. People did not hand over their judgment in one go. They extended trust task by task, and the tasks where they held back were telling.

What the agent did

The agent was built for an enterprise sales team to replace static reporting. An account executive could ask it, in plain language, where their book stood against quota, how much pipeline they had to cover the gap, and which accounts were most at risk or doing best. It ranked the top five of each, with figures like each account’s share of the portfolio and its Risk tier.

The Risk tier is the part I want to focus on, because it is the kind of score most sales teams now have some version of. It combined three things: declining engagement on the account, how close the account was to renewal, and how volatile its spending had been. It was recalculated at every daily data sync, so a tier reflected recent behaviour and not a label fixed when the contract was signed.

What showing the work meant

A score like that is easy to build and hard to trust. Nobody can argue with a bare “high risk” next to an account they believe is healthy, which means they can only accept it or ignore it.

This system was designed to give people a third option. Each metric the agent reported, the Risk tier included, was written as its own query feeding a central data table. The AI was not inventing those numbers when asked. It was reading figures that had already been calculated under rules the team could inspect. When the agent answered, it also surfaced how it got there: which data it pulled, how it filtered that data, how it combined it, and a confidence score alongside the result.

That design was tested before launch with account executives and sales managers across several regions. Their sessions specifically probed whether the agent’s logic was transparent and whether its recommendations were useful, and the feedback went back into the build.

What the data showed

Over the first month:

• 85 percent of the 150 account executives were active in week one, and 90 percent by week five, past the 70 percent target the team had set
• Tableau and Power BI use fell 68 percent on average, and 45 percent of users stopped using them for daily work
• 82 percent of sales intelligence queries went to the agent
• 78 percent of users opened it at least three times a week, and most of their questions concerned prioritizing opportunities and checking account health
• 75 percent of users said they felt they spent about 30 percent less time gathering insights
• on accuracy, the agent scored 0.89 on F1, a standard measure that balances precision and recall, for the insights it synthesized

The time-saved figure is perceived, not measured, and I would read it that way.

What I can and cannot claim

I cannot prove that visible reasoning caused this adoption. The rollout covered the whole organization at once, so there was no control group to compare against. The case study also flags a possible novelty effect: roughly 10 percent of early users interacted less after the first month.

What I can say is narrower. Interpretability has been linked in earlier research on opportunity scoring to better sales performance and forecasting accuracy, and in this deployment the agent’s logic was put in front of testers before launch, after which the wider team adopted it quickly. That is consistent with the idea that trust follows visibility. It is not proof.

Where the system fell short

The more useful part of the study may be the failures. About 7 percent of query volume produced a plausible figure with nothing grounded behind it, and those errors clustered in questions that combined several accounts whose fiscal periods did not line up. Around 15 percent of prompts were ambiguous enough to need rephrasing, and 5 percent were declined outright. Roughly three fifths of those declines, about 3 percent of all queries, were guardrails blocking requests that crossed client-confidentiality lines.

That last number deserves a careful reading. The refusals were a deliberate trade between flexibility and security, and the study did not measure whether users trusted the agent more because it sometimes said no. I find that plausible. A tool that never declines anything is a tool that will eventually answer something it should not. But it is my interpretation, not a finding.

The 55 percent who kept their dashboards were mostly doing work the agent was never built for: custom charts spanning several quarters and board-ready exports. Only 60 percent of account executives finished the advanced module on writing good prompts, and that gap lined up with the share of ambiguous questions. Adoption at this level still needs training and reinforcement.

When a black box is fine

Explainability has a price. Showing the working takes engineering effort, slows some answers down and adds text a busy rep has to read. There are decisions where that price is not worth paying. A team making a hundred quick outbound calls a day may be better served by a simple good-or-bad flag than a paragraph of reasoning, and nobody needs an audit trail to decide which lead to call first.

Account decisions are different. When a score influences how an executive spends a week on a renewal worth real money, the person acting on it needs to be able to ask why, and to hear an answer they can check. That is where I would insist on the working being shown.

What to ask before you buy

If you are evaluating an AI scoring tool, the vendor will quote you an accuracy number. That number answers one question. These answer others:

• Can it show how it reached a given score, in terms a sales rep can follow, not only a data scientist?
• Was each metric calculated under rules your team can inspect, or generated at the moment you asked?
• Can it decline to answer, and does it tell you why?
• Where is it wrong most often, and has anyone measured that?

A vendor who can only answer the first of those has told you how the model performs in a test. The rest tell you how it will behave on a Tuesday afternoon when a rep disagrees with it.

The adoption figures in this study are good, but the detail I keep returning to is the 55 percent. Those users trusted the agent enough to change how they work every day, and they still reached for a different tool when the job called for one. That is what healthy trust looks like. It has limits, it is earned task by task, and it starts with a system that shows its work.

Aakash Nagare currently works as a Senior Sales Operations Analyst at Uber. He has eight years of experience in sales and revenue analytics and recently completed a case study on the deployment of an explainable AI sales intelligence system at an enterprise GTM team.