Every AI pilot works. The trouble starts when it becomes operations.

I have seen plenty of AI pilots succeed. I have never seen one fail in the presentation.

That should bother more people than it does. A pilot is an environment where everything works in your favor: the data was cleaned for the test, the questions were chosen, the team is watching, and the person running it is whoever knows the subject best in the company. Under lab conditions the agent answers well, the customer is happy, the slide looks great.

Operations are a different place. In operations nobody chooses the question. The lead arrives at eleven on a Sunday night, with typos, asking about an apartment in a development that already sold out. The record came from three different sources with the same phone number. And the sales agent who will receive that lead never asked for an AI agent, doesn't trust it, and will judge the whole technology by the first conversation that goes wrong in front of them.

I run AI agents in customer service and sales inside CRMs and WhatsApp. Between pilot and operations, three things almost always break.

The first is the data

The model is the part that stalls least. In almost every project I have followed, the bottleneck came earlier: an outdated price list, a development with two names, inventory that only exists in a spreadsheet someone updates on Fridays, a source field filled in at the whim of whoever created the record.

The agent doesn't fix that. It exposes it. An operation that used to live with bad data because a human compensated by eye starts making mistakes that are visible, written and traceable. AI didn't make things worse. It's just that now it's on the record.

Those who start with the data take longer to get the agent live and go much further afterwards.

The second is the rule nobody wrote down

Every operation has rules that aren't written anywhere. When a lead asks for a discount, the sales agent knows how far they can go. When a customer disappears for ten days, someone decides whether it's still worth calling. When an investor shows up buying four units, the tone of the conversation changes.

None of this is documented. It lives in people's heads.

Putting an agent into operations forces you to write those rules down for the first time. That is the real work, and it is business work, not technology work. When a rollout runs late, most of the time it isn't the model being tuned. It's the company finally deciding what the policy is.

The third is the handoff

This is where the most expensive mistake lives, and it's the one I see repeated most.

The agent is measured by the conversation. How many leads it handled, how long it took to respond, how many questions it solved on its own. Everyone celebrates those numbers.

But the sale doesn't happen in the conversation. It happens in the handoff. The moment the agent decides a lead is ready and passes it to a human is where the result is won or lost. Hand off too early and the sales agent gets junk and stops trusting it. Hand off too late and the customer goes cold or talks to a competitor. Hand off without context and the salesperson starts from scratch while the customer repeats everything, which is more irritating than not having been served at all.

An agent that answers in two seconds and hands off badly delivers less than a salesperson working alone. I have seen it happen.

What I would do differently

If I were starting over in an operation, I would do it in this order.

First, define what a sales-ready lead means, in writing, with the sales team in the room. Before any prompt.

Run the pilot on dirty data on purpose, at the bad hours, with the questions nobody picked. A pilot under ideal conditions only proves the model works, and we already know that.

Have the sales team rate the handoff from the first week, not at the end. If the people receiving the leads don't trust them, the project dies from within before it dies in the report.

And measure only one thing at the start: what happened after the handoff. Everything else is interesting. Only that is results.

The takeaway

AI pilots are easy. The difficulty isn't technical, it's organizational, and it shows up exactly where the technology meets people who didn't ask for it.

Treat the rollout as a technology project and you deliver an agent. Treat it as an operations project and you deliver a different operation. Only the second one changes the number at the end of the month.