On the Seam: Working Drafts · The Last Mile, Part 1 of 7

The Demo Is Not the Deliverable

AI projects don't fail at the demo. They fail at the handoff to daily work — and the more autonomy AI gets, the more that handoff is worth.

By Dr. Trey Harper

Last week I gave a talk and workshop at Morgan State University's Center for Equitable AI and Machine Learning Systems. The subject was a pattern I've watched for years — from the seller's side, the buyer's side, and now the research side.

This is the first of seven editions working through it in public.

It starts with a scene most of you have lived.

Someone runs an AI demo. It works. The summary is crisp, the forecast lands, the agent books the meeting without a hitch. People clap. Someone senior says the words every vendor waits for: let's roll it out.

Six months later the licenses are paid for. The tool is installed. The training happened. And the work looks exactly the way it did before.

Nothing broke. Nothing changed.

That gap — between the demo and daily use — is the last mile. My argument is simple:

The demo isn't dishonest — it's incomplete

The demo works because the room is friendlier than the business.

In the demo room there is one clean version of the process. The data behaves. The audience wants it to work. And nobody in the room has to live with the result on Monday.

The business is different. The process has a version for every region, each with its own workarounds. The data has gaps and history. The people whose work is about to change are skeptical — often for good reason. And somebody has to answer when it goes wrong.

None of that shows up in a demo. It can't. A demo can only show the process it was given.

This isn't a fringe problem

In August 2025, MIT's NANDA initiative reported that about 95 percent of the enterprise generative AI pilots it studied had produced no measurable impact on profit and loss. The cause wasn't the models. It was a learning gap — tools that never adapted to how the organization works, and organizations that never adapted to the tools.

In July, a Grant Thornton survey of 100 manufacturing leaders found that none reported significant cost savings from AI. Nearly half were still in pilots.

The models are good. The demos are good. The value goes missing at the handoff.

Why the stakes keep rising

Picture a simple curve. Along the bottom, autonomy — a person doing the work, then GenAI advising a person, then an agent acting on its own. Up the side, stakes.

As autonomy rises, the value AI can deliver climbs steeply. That's why the money is flowing. But the cost of failure climbs right alongside it. A flawed process run by a person hurts one case at a time. Run by an agent, the same flaw executes across systems before anyone notices.

More autonomy puts more money on the table — in both directions. The last mile decides which side you collect.

Where the last mile actually lives

After enough projects, the failures start to rhyme. They come down to four questions every project should answer before it goes live:

• Owner. Who is accountable after launch? Everyone loved the demo. Nobody answers when it breaks.
• Process. What about the work actually changes? If the answer is nothing, the tool is bolted onto the old way.
• Trust. Why will people rely on the output? If they don't, they re-check everything — or quietly ignore it.
• Measure. Which number moves, and from what baseline? Without one, nobody can say it worked.

Skip one, and that's where the project dies.

Last week I posted about one of them — trust — and why trust and distrust aren't opposites. We'll come back to it in Edition 6.

What's coming

Over the next six editions I'll take these apart one at a time:

• Why usage is not value — and why the number has to come before the tool
• How an average can hide the regions where a pilot failed
• Why the pilot everyone celebrates often just breaks even — and the return lives in the rollout
• How AI amplifies whatever process you hand it
• What my research on swift trust and distrust suggests about agents — and the hypothesis I want to test next
• How to run a premortem on your own AI project, and what a room full of researchers found when we tried it

Arguments offered, not finished. Push back where I'm wrong.

Your turn: think about the last AI demo that genuinely impressed you. What happened at the handoff?

Dr. Trey Harper writes on trust, legitimacy, and the architecture of the agentic enterprise at treyharper.com and in the LinkedIn newsletter On the Seam: Working Drafts. Views are my own and do not represent any current or former employer or client. Examples are composites.