What we've learned building production agents inside small and mid-sized companies. Practical, opinionated, no AI-influencer hype.
"Agentic AI" became a buzzword sometime in late 2024. Most people using the term mean something between a chatbot and a Zapier flow. The actual definition is more useful — and more demanding. Here's how we draw the line, and what it means for SMEs deciding whether to build one.
An agent perceives, reasons, acts, and reflects in a loop. The hard parts are the same hard parts they've been for forty years: knowing when to stop, knowing when you're wrong, and knowing when to ask a human.
Why we wrap every tool call in a typed contract — and the eval suite we run before letting an agent write to a real system.
A walkthrough of why support and lead triage have the right shape for early agentic deployment — and the three you should NOT start with.
Five things every "AI transformation" framework gets wrong when applied to a company under 500 people, with worked examples.
Week-by-week: discovery, baseline, build, eval, soft-launch, ramp. The version we run with every new SME engagement.
Agents accumulate beliefs. Most of them are wrong. Here's the audit pattern we use to keep agent memory grounded and revisable.
A teardown of three real engagements with the actual labor cost displaced — and the costs no one mentions in the brochure.
The phased access matrix we use across every engagement, and why "shadow mode" is non-negotiable for the first two weeks.
A field guide to running open-source models on your own hardware: the break-even calculation, the maintenance reality, and the compliance wins.
Why we treat eval suites as the contract for an agent's behavior, and the four eval categories every production agent needs from day one.