The last mile of AI is human

On October 19, 2026, I spoke for Publicis Sapient at the E-ZPass Interagency Group's 2026 Industry Day Forum, held at the Port Authority of New York & New Jersey in Jersey City. The forum is for IAG and its member agencies. My talk was called The Last Mile of AI Is Human.

One question, seven kinds of information

I opened with my wife's toll notice. Her card was stolen and her toll payments declined. Then a notice arrived with fines. The calls went nowhere. Agents couldn't find her information, and one finally told her there was no solution in the system. Now she updates her account by hand every month rather than call again.

To her it was one question. Behind it sat the account, the vehicle, the transponder, the transaction, the payment, the policy and the case history.

A public agency went first

Ahead of the World Cup and America250, the Port Authority's Aviation and Technology teams set out to make four complex airports feel simpler before a traveler reaches a terminal. In July 2026 they launched rebuilt websites for JFK, Newark Liberty, LaGuardia and New York Stewart International, each with its own AI assistant. We built them alongside their teams.

One traveler question there draws on live flight information, security waits, parking, accessibility and more. Her notice needs the same thing: a shared context layer, with AI working through the systems of record instead of becoming another one.

The work tested the institution: procurement, security, data access, legacy systems, governance. Design went by committee, and each round added time. The team was wary of committing quickly to GraphQL, and seeing the benefit is what moved it. Both were reasonable instincts. Neither was a technology problem.

The hard part moved to people

The premise of the talk: "AI has made it possible for almost anyone to create. It has not made it easier for humans to decide."

Before AI, making the thing was roughly a quarter of the work. AI shrank ideating the most and making and integrating a lot. Operating shrank some. Persuading and approving take as long as they did. I said on stage that those proportions are illustrative, not measured.

It happened to us too. On one large program, we put calls, documents and work tracking into one searchable graph with a language model on top. It worked. Then people started finding reasons not to like it, because once everything was easy to find, it also showed where work was waiting on people.

Keep the caution, and still move

Public agencies get told they're slow. I said their caution is a strength. The way through is a first step small enough to govern, real enough to show measurable value and visible enough to build confidence. Start with one customer journey, connect only the context it needs and expand from evidence.

Don't start with autonomy. Earn each level, from assist to automate to orchestrate, with evidence, evaluation, observability, security and a clear path to a person.

The Port Authority is already moving. To support its new assistants, it moved on Cursor licenses and trained its own people, and a group in its Technology team now maintains and improves the assistant. A good partner becomes less necessary over time.