"AI agents" is the most oversold phrase in tech right now. Everyone has a deck; almost nobody has one running unsupervised on real work. We do — it's been cutting 80%+ of the manual research and follow-up load in a real-estate operation — so instead of another think-piece, here's what shipping one actually taught us.
The problem we started with
A real-estate team was drowning in the least human part of their job: monitoring listings, researching properties, qualifying leads and preparing follow-up — hours of repetitive digital work per agent, per day, before any actual selling happened. Classic agent territory: repetitive, rule-describable, all on-screen.
What the agent does
It runs a loop we design into every agent: perceive → plan → act → learn. It watches listing sources, extracts and enriches property data, qualifies leads against the team's actual criteria, drafts the outreach and follow-up, logs everything back into the CRM — and when it hits an edge case it wasn't built for, it escalates to a human instead of improvising. The result was an 80%+ reduction in the manual research and follow-up workload. The humans kept the judgment calls and the relationships; the agent took the grind. You can read the full build in the case study.
Five lessons from getting it into production
1. Automate one workflow, not "the business." Every failed agent project we've rescued started with a mandate the size of a mission statement. Pick the single workflow that hurts most, get it to boring reliability, then expand. Narrow is fast; broad is a demo.
2. Approval comes before autonomy. For the first stretch, the agent drafted and a human clicked send. Only after weeks of accepted drafts did it earn autonomy on the routine cases. Trust is built on a paper trail, not a promise.
3. Evals before autonomy, always. We scored the agent's decisions against what the team would have done — hundreds of cases — before letting it act alone. If a vendor can't show you this step, they're proposing to test in production on your customers.
4. Track cost per task, not cost per month. The metric that matters is "what does one completed follow-up cost?" — model calls included. Engineering that number down (caching, small models for simple steps) is what makes the ROI arithmetic undeniable.
5. Boring beats clever. The version that shipped is less impressive in a demo than the version that didn't. Retries, timeouts, idempotent writes, escalation paths — reliability engineering is 60% of a production agent and 0% of a keynote.
Is your workflow agent-ready?
Score it: the work is repetitive and digital · the inputs are readable by software · "done correctly" is definable · an occasional error is catchable through review · it happens at real volume. Four out of five and you have a candidate.
The fastest next step: tell our homepage AI which workflow you'd hand over — it'll tell you honestly whether an agent can own it, and a senior engineer can map it in a free 20-minute call.