Why Nobody Uses Your Agent After Launch: Five Approaches to Drive Adoption

Early this quarter we took on a struggling handoff. A client had launched an internal agent six months prior. It could pull contracts, calculate quotes, and auto-draft customer service replies. Everyone was optimistic during the board presentation.

I opened the backend to check the usage data: 22 weeks live, 340 targeted users across the company, 19 weekly actives.

The system worked fine. It ran smoothly and the answers were accurate. The problem was that nobody was actually using it.

This is the core issue we've seen most often the past two years. Demo passes. Acceptance gets signed. Budget gets spent. Then the agent sits unused. You can't calculate ROI without a denominator, without adoption, there's no payoff, whatever the business case looked like on the slide deck.

Why agent adoption gets stuck in single digits

Start with a stark number: most internal agent projects land at a natural adoption rate between 5% and 15% if nobody actively manages them in the first three months after launch. It's not that anyone's incompetent, it's the default outcome. And it's rarely because the model isn't smart enough. The root causes are more basic.

First, the agent isn't where users already work. They spend their day in CRM, ERP, Slack, and Outlook. Now they have to open another tab, remember another URL, learn another way to phrase their questions. Every added step costs users. Second, the agent got something wrong once, and it was at a critical moment. Trust breaks easily. A sales rep gets burned by an outdated price the agent provided, and afterward they'd rather flip through their own Excel spreadsheet. Third, nobody owns the system. Management doesn't use it either, so the team treats it as an IT experiment rather than a business tool.

In other words, low adoption is rarely a technical problem. It's a process, trust, and accountability problem. Retraining the model usually doesn't fix it.

Five approaches we've tested

Here are five specific things we've done that moved the numbers. They work in sequence, skip around at your own risk.

One: Embed the agent into the places users already go. Don't build a separate entry point. For that project with 19 weekly actives, we didn't retrain the model. We pulled it out of the standalone website and moved it into the CRM sidebar and Slack they were using every day. Users didn't have to decide whether to use the agent. It was just there. This single change brought weekly actives from 19 to 140 in six weeks.

Two: Pick one narrow scenario and get it right. Don't try to solve everything at once. Everyone wants to build a universal assistant that can answer anything. The result is something that answers everything poorly. We do the opposite: identify one high-frequency, high-impact scenario where being wrong is manageable, say, 'look up the latest price and stock for a product', and get the accuracy nearly perfect. Users build trust in that one use case first. They'll try the second thing only after the first one works. Trust gets built one win at a time.

Three: Share real adoption numbers publicly. Make adoption visible. Post a chart every week in the team channel: who's using it, how many times, how many hours saved. Not to shame anyone, to make 'everyone else is using this' a visible fact. We track manager usage separately because we've never seen this fail: if the manager doesn't use it, the team won't either.

Four: Make it safe to get things wrong. In high-stakes scenarios, keep a human in the loop. The agent drafts, then a person reviews before it goes out. Attach source links to every answer so users can verify in seconds. Allow mistakes, but make them obvious, adoption will be higher than with a black-box system that claims to never fail.

Five: Assign someone to watch the numbers and fix things. Launching the agent isn't the end of the project, it's the beginning of operations. Someone needs to look at usage data every week, collect friction points, and prioritize fixes. We call this role AIOps. They understand the actual work and can move the system forward. Without this ownership, adoption starts declining in week four.

Launch is just the start of the numerator

Back to ROI. When that project's weekly actives climbed to 210 and customer service drafts dropped from 9 minutes to 3.5 minutes, the business case worked, because people were using it and the time saved was real. Adoption isn't a side metric. It's the denominator. Without it, there's no ROI to calculate.

At Tenten, we hold one principle constant: a beautiful demo doesn't count. Real adoption counts. Our engineers go into the field, integrate the agent into your team's actual workflow, and stay with you as adoption grows week by week. They don't hand it off at sign-off. Whether a system gets used isn't decided at launch. It gets decided by who owns it afterward.