AI agent development

The four rungs, and what each one buys you

• 1 — It watches and tells you. No actions at all. It reads, notices, and raises the things that matter. Boring, cheap, and it already recovers most of the value on a job nobody has time to review daily.
• 2 — It drafts and you sign. It prepares the order, the reply, the report, and a person approves. This is where most business value actually sits, because the slow part was never the clicking.
• 3 — It acts inside a limit. Routine work under a spend cap, a margin floor, a volume ceiling. Anything outside stops and names the reason. This rung needs the audit trail to exist first.
• 4 — It acts and adjusts its own limits. Rarely worth it, and the point at which the guardrail becomes decoration. I have not needed rung four for anything, and I would argue hard before building it.

How you prove one before it touches money

The test is not a demo. It is a week of the agent running beside the person who does the job now.

• It proposes, the person decides, and both answers are written down. Agreement rate is the number that matters, not accuracy in the abstract.
• Every disagreement gets read by a human once. Half of them will be the agent being right about something nobody had noticed.
• Feed it a deliberately abnormal week — a promotion, an outage, a spike — and watch whether it stops. If it never stops, the guardrail is not wired to anything.
• Only then does it get a limit and permission to act.

What makes an agent expensive, and it is never the model

• Data that has to agree first. If two systems hold different numbers for the same record, the agent will pick one and act confidently. Joining them is usually the longer half of the project.
• The audit trail. Every decision with the inputs it used, kept long enough to settle an argument in three months. Skipped in demos, needed in production.
• The stop list. The explicit set of things it must never decide alone. Writing it takes real meetings with the people who own the risk, and it is the actual deliverable.
• The monthly bill. Model calls, storage, the machine it runs on. A number that grows with your volume and belongs on the table before the build, not after.

Build an AI agent, or buy an agent platform

• The platform sells the plumbing. An AI agent platform gives you the orchestration, the retries, the tool wiring and a screen to watch it on. That is real work you no longer have to write. It is also the cheap half.
• The expensive half stays yours. Data that agrees with itself, the audit trail, the stop list. No platform writes those for you, because they are decisions about your business wearing the clothes of software.
• Buy when the job is standard. Common systems, an ordinary shape of work, no unusual boundary — the platform gets you there faster and cheaper than I can, and I will tell you so in the first hour.
• Build when the boundary is specific. A custom AI agent earns its cost when what it must never do is particular to you, or the systems it has to reach are your own. That is exactly where a platform starts charging you to work around itself.
• Price both at your real volume. Platforms charge per seat, per run or per action. The number that looks obvious at ten users is a different decision at three hundred, and that arithmetic belongs before the pilot rather than after it.

Too early for you if

• The job happens a few times a month. An agent earns its keep on repetition.
• Two people in your team would judge the same case differently. Agree the rule first — that alone is worth the meeting.
• Nobody owns the exceptions queue. An agent that stops and names a reason still needs someone reading it the same day.