title: "Agents that earn their keep" date: "2026-07-06" excerpt: "Everyone's building AI agents. Almost none of them make it into real back-office work. The gap isn't intelligence... it's trust, guardrails, and knowing what to automate versus redesign."

There has never been a better time to build an impressive AI agent demo. There has also never been a wider gap between that demo and something a real team will actually let touch their work.

The number floating around the industry is that only about 5% of custom enterprise AI tools ever make it into production. That statistic isn't about the models. The models are extraordinary. It's about everything around them... the part nobody claps for in the demo.

I've now built several AI products that live in real workflows, from WaitWho, which does relationship intelligence and lead scoring, to the outbound engine inside GTM Hub, to the activation systems in Onboardly. So let me tell you what actually separates the 5% that ship from the 95% that die in a Notion doc.

Trust is the product, not the model

A demo agent gets to be brilliant once, under perfect conditions, for an audience that wants it to work.

A production agent has to be trusted by a skeptical reviewer who gets blamed if it's wrong. That reviewer doesn't care that it's right 95% of the time. They care about the 5% that lands on their desk with their name on it.

So the real design work isn't the prompt. It's the trust ladder. Start the agent in assisted mode: it proposes, a human approves, and every approval is a chance to catch an edge case before it becomes a customer-facing mistake. Let it act on its own only where the action is low-risk and reversible. Keep a human gatekeeper on anything that spends money, deletes data, or emails a client. You earn autonomy. You don't ship with it.

The build-versus-platform line

The second thing that kills agents is building the wrong layer.

If a capable off-the-shelf platform does 80% of what you need, and your edge is in the other 20%, do not spend six months rebuilding the 80%. Build the 20% that's actually yours and buy the rest. The teams that ship know exactly where their moat is and refuse to hand-craft the plumbing around it.

The question isn't "can we build this." With today's tools, you can build almost anything. The question is "should this be ours," and most of the agent should not be.

Automate the task, or redesign the job?

Here's the mistake I see most: teams point an agent at a broken process and automate it, faithfully, at speed. Now you have a fast broken process.

Before you automate anything, ask whether the task should exist at all. Half the back-office work an agent could take over is work that only exists because a system upstream was designed badly. Automating it cements the bad design. Sometimes the right agent isn't the one that does the task... it's the one that makes the task unnecessary.

Automate the tasks worth keeping. Redesign the ones that only exist because nobody stopped to ask.

What this actually takes

The unglamorous truth: a production agent needs the same seriousness as any core system. Evals, so you know when it degrades. Observability, so you can see what it did and why. Human checkpoints, so trust is earned in public. Feedback loops, so every failure becomes a test instead of a surprise.

The intelligence was the easy part. It's the guardrails, the trust, and the humans who have to sign their name under the output that decide whether your agent earns its keep... or joins the 95% that never leave the demo.

I give a version of this talk to teams putting AI into real operations. If that's the wall your company is hitting, put me in front of your team.