What an AI automation agency actually delivers

The AI automation agency category grew faster than its standards did. Two years ago there were a handful of people who could actually ship this work. Now there are thousands of listings, and a large share of them sell a slide deck with a roadmap in it.

I have inherited enough of those engagements to know the pattern. So here is the honest version: what a real engagement delivers, what it costs, and the five questions that separate a builder from a broker.

The deliverable list, item by item

A finished AI automation engagement should hand over these things. If any one is missing, you own a prototype.

1. Working software in your accounts. Your cloud, your repository, your API keys, your database. Not the vendor's workspace with your logo on it. If cancelling the retainer turns the system off, you did not buy an asset.
2. A written runbook. What the system does, what it costs per run, what alerts mean, how to pause it, who to call. Two pages beats forty. A successor should be able to keep it alive without the builder.
3. An eval or test suite. A set of real cases with known correct outcomes, runnable on demand. Without this, nobody can tell whether next month's model update improved things or quietly broke them.
4. Monitoring and alerts. Failure rate, latency, spend per day, and a quality signal. Routed to a channel a human reads.
5. Cost controls. Per-run token caps, daily spend ceilings, and a kill switch. AI systems fail expensively when they fail in a loop.
6. A recorded handoff. One session, screen shared, covering how to change the parts you will want to change.

What gets sold instead

Four substitutes show up constantly, and each one is a tell.

• The discovery phase that produces a document. Discovery is real work, but it should take days and end in a scoped build, not weeks and end in a maturity model.
• The dashboard. A screen showing how much time you theoretically saved. It is a marketing artifact for the retainer, not an operational tool.
• The demo that never leaves the demo account. Works beautifully on the three sample records it was built against. Has never seen a scanned fax, a duplicated customer, or a 400-page contract.
• The seat-license wrapper. A thin interface over a model API, priced per user, hosted by them, with your data inside it. Sometimes a fine purchase. It is not automation of your process.

Real ranges, so you can calibrate

Numbers vary by market and complexity, but these are the bands I see holding across North America and Europe.

• One production workflow, no retrieval. Two to four weeks. Roughly 3,000 to 12,000 dollars. Example: intake, classify, extract, route, notify, log.
• Agent system with retrieval and human review. Four to ten weeks. Roughly 15,000 to 60,000 dollars. Example: document pipeline with citations, approval gates, and an audit trail.
• Running cost. Usually 100 to 800 dollars a month at moderate volume once caching and model routing are in place, which I break down in what AI workflow automation really costs.

Quotes far below those bands usually mean a prototype. Quotes far above usually mean layers of people between you and the person writing the code.

Five questions to ask before signing

1. Show me a system you built that is still running, and tell me what broke in month three. Anyone who has shipped has an answer here. It is the fastest filter I know.
2. Whose accounts will this live in? The only good answer is yours.
3. How will we know if quality drops after launch? If the answer is not a concrete signal or eval, there is no answer.
4. What is the per-run cost, and what caps it? Silence here is expensive later.
5. Who writes the code, and will I talk to them? Not the account manager. The engineer.

We paid for six months of AI strategy and got a Notion page. The first thing that actually worked was two hundred lines of code.

Why so many engagements produce nothing

It is not usually incompetence. It is that the hard parts of this work are invisible in a sales conversation. Nobody buys retry policy. Nobody gets excited about an eval suite. Nobody asks whether the extraction step handles a rotated scan. Those are precisely the things that decide whether the system is alive in six months, so the market underprices them and buyers do not know to ask.

The counter-move is simple: judge proposals by their operational content, not their vision content. A proposal that mentions failure handling, evals, cost caps, and ownership was written by someone who has been paged at 2am. That is the person you want.

Where I land

I build the systems and hand them over. If you want a ranked, honest read on what you already have, the AI production readiness audit gives you a prioritized action plan in 48 hours, and the fee credits toward the build. If you know what you need built, the services overview lays out how I scope it, and the work shows systems currently in production.

Frequently asked