I Stopped Asking AI Questions and Started Handing It the Work

I'll admit the experiment started as a kind of dare to myself. I had spent months treating AI the way most owners do — as a very fast intern I'd ask a question and then double-check. So I decided to stop asking and start delegating. Not a prompt. A workflow. Something with a beginning, a middle, and an outcome I actually cared about. I wanted to see where the agent carried the job to the finish line, and where it quietly walked off the road.

The timing wasn't an accident. Gartner expects 40% of enterprise apps to ship a task-specific AI agent by the end of this year — up from under 5% in 2025. That's the kind of jump that makes you feel late even when you're early. But Gartner buried a second number in the same breath, and it's the one I kept thinking about: more than 40% of agentic AI projects could be scrapped by 2027, killed by runaway cost, fuzzy ROI, or governance no one set up. So I went in holding both numbers at once. Promise in one hand, body count in the other.

Where the handoff actually held

The first surprise was how unglamorous the wins were. The agents that earned their keep weren't writing strategy. They were closing loops.

I started with customer support, because that's where the loops are most painful. The pitch used to be "deflection" — keep people away from a human. That framing is dead. The agents worth deploying now aim for resolution: they sit on top of the CRM, work a question across email, chat, WhatsApp, and voice, book the appointment, issue the refund, and only tap a human when judgment is genuinely required. Watching Salesforce Agentforce, Sierra, and Decagon handle a ticket end to end — and others like HubSpot Breeze, Zoho Zia, and Tidio Lyro in the same lane — the thing that struck me wasn't intelligence. It was completion. The ticket was actually done.

Sales prospecting was the next place the handoff held, in a narrower way. I think of these as digital sales assistants rather than closers: Clay and Relevance AI will find leads, enrich them until a name becomes a context, write outreach that doesn't read like a mail-merge, and then — the part that matters — hand a warm lead to a human at the right moment instead of bludgeoning it to death. The agent did the part I hate. The human kept the part that needs a pulse.

The part nobody puts in the demo

Then there's the unglamorous middle of a business — the cross-app plumbing. For an owner without a dev team, this is where the real leverage hides, and where Lindy AI, Zapier Central, and Make let you stand up agents for legal, marketing, ops, and have them cooperate on a bigger workflow. I'll be honest: this is also where I watched the most quiet failures. Not dramatic ones. An agent would do nine steps perfectly and fumble the tenth because nobody had told it what "done" looked like. That's not a model problem. That's a context problem, and it's the exact thing the cancelled-project statistic is made of.

The internal-facing agents were steadier. The IT service desk is all alert fatigue and password resets and hardware tickets — repetitive enough that Glean, Moveworks, and Siit can clear the routine load and flag the small problems before they grow teeth. And the most genuinely startling category was autonomous software engineering. These have moved well past autocomplete: Cognition (Devin, now paired with Windsurf) and Cursor from Anysphere will pull a codebase, write the tests, debug, and open a pull request without a human touching the keyboard. The first time I saw a PR land that I hadn't asked a person to write, I felt the floor shift a little.

Why the boring answer is the right one

Here's the conclusion I didn't expect to reach. The general-purpose tools are converging — they're all becoming roughly the same competent assistant. The durable advantage is going vertical: agents built for one industry, with the domain knowledge, the compliance rules, and the integrations already baked in. The numbers back the instinct. Vertical AI agents are compounding at about 62.7% a year through 2030, outrunning the broader agent market's 46.3% — a market on its way from $10.9B this year to north of $50B by 2030.

But the spreadsheet isn't the lesson. The lesson is the same one the cancelled projects keep teaching, post after post. The agents that survived in my experiment were the ones pointed at a specific business outcome I could name before I started. The ones that died were pointed at "let's see what AI can do." The technology was never the variable. The clarity was.

So I've stopped asking what AI can answer. The better question — the one I'd put to any owner reading this — is simpler and a little uncomfortable: which loop in your business would you actually hand over first?