The $500K deck trap why most mid-market AI projects die before shipping.

The anatomy of the deck trap

Every quarter, about forty mid-market companies in Europe and the US pay someone between $300K and $700K to produce a document. The document is called an “AI strategy” or an “agentic transformation roadmap.” It is, without exception, a PDF. It circulates for three to six months. At the end of those months, no agent is running in production.

The deck itself isn’t the problem. The problem is what the deck replaces: decisions. A sixty-page strategy document is almost always a symptom of an organization that hasn’t decided who owns the first workflow, which system the agent will read from, and who gets woken up at 2 AM when the agent does something unexpected.

I’ve watched this cycle from both sides — as the operator hired to ship the thing after the deck fails, and as the investor reading the post-mortem. The pattern is remarkably consistent. The failure isn’t technical. It’s organizational.

That quote isn’t cherry-picked. It’s the median case. The companies that ship agents fast don’t skip strategy — they compress it into a different format. A two-page scope. A named owner. A deadline that means something. And critically, a team that has permission to fail fast on a narrow wedge before they’re asked to transform the whole department.

Four decisions that separate shippers from stalls

After twenty-plus engagements, the pattern is clear. The teams who ship production agents in weeks (not months) make four decisions early that the stalled teams never make at all.

1. Name the workflow, not the department

A good agent scope is “automate the invoice-to-PO matching workflow in Procurement” — not “bring AI to Procurement.” The teams that ship always name a specific workflow in the first week. The teams that stall produce a department-wide assessment and call it progress.

2. Assign a production owner before writing a line of code

An agent without a production owner is a demo. Someone inside the client organization needs to be on the hook for what happens when the agent runs at 2 AM on a Saturday and does something wrong.

3. Set a six-week shipping window

Six weeks isn’t a magic number. It’s a forcing function. A six-week timeline makes it impossible to run a four-month discovery phase. It requires the team to pick a narrow enough scope that they can actually deliver, test, and iterate within a single sprint cycle. Every engagement where we’ve shipped successfully has used this constraint.

4. Kill the staging environment mindset

Agents don’t work in staging. They work in production, with real data, real users, and real consequences. The companies that ship fast design for production from day one — monitoring, guardrails, human-in-the-loop fallbacks. The companies that stall spend months in a sandbox and then discover their agent can’t handle the one data format their ERP actually produces.

What this means for your next pilot

If you’re about to start an agentic AI pilot, here’s the short version: don’t start with a deck. Start with a decision. Name the workflow. Name the owner. Set the clock. And build for production from the first commit.

The $500K deck trap is real, and it’s not malicious — the consultancies producing those decks believe in them. But the distance between a document and a deployed agent is measured in decisions, not pages. The teams that ship understand this. The teams that don’t, produce beautiful PDFs.

If this resonates and you want to talk through your specific situation, I take a limited number of 30-minute calls each week. No pitch, no deck — just the conversation this essay opened.