Field Notes on AI in the Wild

Every founder we talk to has the same question: 'Where do we actually start with AI?' After shipping dozens of agent and automation projects across legal, healthcare, e-commerce, and service businesses, the pattern is clear. AI pays off fastest in three places: response time, repetitive judgment work, and revenue recovery.

Response time is the unsung killer of conversion. If a lead waits more than 5 minutes for a response, your close rate falls off a cliff. AI receptionists, voice agents, and chat agents close that gap to under 10 seconds — 24/7. We've seen this single change lift booked appointments by 30–60% with no change to the offer.

Repetitive judgment work is the second wedge. Triaging support tickets, categorizing leads, drafting first-pass responses, qualifying inbound — all the 'thinking work' that doesn't really require thinking. AI agents handle the 80% so your team can focus on the 20% that actually needs human judgment.

Revenue recovery is the third — and most underrated. Every business has dead leads, abandoned carts, lapsed customers, and dormant accounts sitting in their CRM. A reactivation agent that runs in the background can recover 3–8% of that pile in the first 90 days. It's pure margin.

The build-vs-buy question used to be simple. Now it's not — because the line between 'tool' and 'agent' has blurred, and most SaaS vendors are bolting on AI features that may or may not do what you need.

Buy when the problem is generic, the workflow is standard, and your competitive advantage doesn't live there. Email marketing, calendar booking, transcription, basic CRM — buy.

Build (or commission a custom agent) when the workflow encodes your unique operating knowledge, when integrations matter more than features, or when you need control over data, prompts, and behavior over time. Reactivation, qualification, internal ops, custom support flows — usually build.

The deciding questions: Does this workflow change as we grow? Does our edge depend on how we do this? Will we outgrow a tool's defaults in 12 months? If any answer is yes, build.

We get the cost question constantly: 'Won't running AI agents 24/7 burn through tokens?' Almost never, in practice. The mental model people use — pricing per request like a SaaS seat — is wrong for agents.

A typical reactivation agent runs a few hundred conversations per month. Even with a premium model, that's $20–$80/month in API costs. The infrastructure, integrations, and monitoring around it cost more than the model usage itself.

Compare that to the alternative: a part-time SDR at $3,000/month who can't work nights, weekends, or holidays, and who needs onboarding, management, and benefits. The agent is 1–3% the cost and works around the clock.

The real cost of AI agents isn't tokens. It's the engineering required to make them reliable, integrated, and observable — which is exactly the work we do.

Roughly 70% of AI pilots we've seen at other companies never make it to production. The technology works. The problem is upstream of the model.

Cause #1: No clear owner. The pilot lives in 'innovation' or a side team, never integrates with the actual operators, and dies the moment attention shifts.

Cause #2: No baseline metrics. Without a 'before' number, you can't prove a 'better' number. Pilots that don't pre-commit to measurement get killed by anecdotes.

Cause #3: Wrong workflow chosen first. Teams pick something flashy instead of something painful. The right first project is boring, repetitive, and high-volume — that's where the ROI lives.

Cause #4: Built but not integrated. The agent works in a demo but isn't wired into the CRM, calendar, or comms stack. Without integration, it's a science project.

Avoid all four and your pilot moves to production. Skip any one of them and it doesn't matter how good the model is.