What's actually working in 2026
Three categories: inbox triage (replying to and tagging support emails), data-entry replacement (pulling fields out of PDFs/emails into your system), and outbound research (qualifying leads before a human ever sees them). All measurable, all paying back fast.
Where agents quietly fail
Anything open-ended: 'be our copywriter', 'run our marketing'. Without tight constraints, output drifts, quality varies, and the human ends up reviewing every message. The maths only works when agents have a narrow, repeatable job.
How to scope an agent build
Pick one job a junior would do. Define a clear input, a clear output, and a clear quality bar. Build the smallest version. Measure error rate against the human baseline. Ship when you beat it on cost + speed at acceptable quality.
What it costs to run
A well-scoped agent for a small business runs $20–200/month in inference, depending on volume. Compare to the $2–3k/month a part-time hire would cost for the same job. The ROI is usually a 90-day payback.
What we won't build
Agents that replace humans in customer-facing roles where mistakes carry reputational cost. Agents with write-access to production systems and no human approval step. Agents that 'learn' from production data without governance. The fun ones are also the dangerous ones.
Where to start
Pick the one task in your business done daily, by a person, on a screen, with predictable inputs. That's the agent worth building first. Save the autonomous-CEO stuff for someone with VC money to burn.