Ideas and evidence for people who run things.

It is almost never the model. In the stalled pilots we have examined, the failure sits in one of three places. First, the data: the tool was asked to read labor, sales or inventory data that was never clean because no human needed it to be. Second, the integration: the pilot ran on a manual export or a vendor's one-off connector that nobody could extend. Third, the manager: the tool changed a Tuesday routine without anyone redesigning the routine, so the manager worked around it and the dashboard quietly filled with stale numbers.

All three are visible in advance. That is the entire argument for a three-week Discover phase before anyone buys a tool. Map the systems, inspect the data, watch the managers. Then write down the baseline the tool is supposed to move. A pilot without a baseline cannot fail, which is why so many of them never end.

The practical fix for a stalled pilot is not a better vendor. Go back to location three. Fix the export. Redesign the routine with the manager. Measure one number for four weeks. Then decide whether to roll to location four or stop. Most of the time the answer is roll, but it is the measurement that makes the decision safe, and the measurement is what the original pilot skipped.

1. What number does this move, and what is it today? If nobody in the room can state the baseline, stop. You are about to buy something you cannot evaluate.

2. Who owns the data, and may you train on it? Read the clause. Many hospitality tools reserve the right to use customer data to improve the product. Sometimes that is acceptable. It should never be a surprise at renewal.

3. What happens when it is wrong? Every AI tool is wrong sometimes. Ask for the escalation path, the human review step and who is liable when a guest or employee is affected.

4. What does it cost at 50 locations, not 5? Per-location pricing often steps up with usage, API calls or "premium" features. Model the run rate at the size you intend to be.

5. What does it need from my other systems? If the honest answer involves a nightly export someone has to remember, the tool's reliability is now that person's reliability.

6. Who on my team owns it after go-live? Not the vendor's support queue. A name, in your company or your partner's, with a monthly review on the calendar.

7. What would make us turn it off? Agree the exit criteria before the start. A tool that cannot be turned off is not a pilot, it is a commitment with extra steps.

None of these questions is hostile. A good vendor will have crisp answers and will respect you more for asking. A vendor who changes the subject has told you what you need to know.

This is why we built Managed AI Operations as a practice, not an afterthought. Here is what the desk actually does in a month. It watches every agent's output for drift: a labor forecast that starts running hot, a review-response agent that gets more generic, an invoice-coding agent whose exception rate creeps up. It triages failures, which are usually upstream data or vendor API changes rather than the agent itself. It tests model and prompt changes against your baseline before any store sees them. And it writes the monthly review: one page per use case, the result against baseline, what changed, what we recommend.

The economics work because the desk is shared and the tooling is ours. One operations team watching a hundred client locations costs a fraction of what a hundred clients would spend watching one. Per-location pricing passes that economy to you, and it keeps our incentive aligned: we are paid to keep the thing working, not to rebuild it.

The alternative is an internal hire, which for most operators under 200 locations does not pencil, or a vendor's support queue, which is paid to close tickets. Neither owns the result. Someone has to.

Hourly billing puts the incentive on hours. Every efficiency the firm finds is revenue it loses. A fixed fee reverses that: the firm is rewarded for knowing how to do the work quickly and well, and punished for scoping it badly. It is harder for the firm, which is why most firms avoid it, and better for the buyer, which is why you should ask for it.

Retainers are right for work that is ongoing and senior: a fractional executive, a standing advisory relationship. The test is whether the value is in the person's judgment over time rather than in a deliverable. If it is, a retainer with a named person and a cadence is the honest structure. If the firm is trying to put project work on a retainer, that is a rate card with a different name.

Gainshare is the structure everyone asks about and few firms offer, because it requires a baseline both sides trust and a verification method that cannot be gamed. Done properly, it is the strongest alignment there is. Done casually, it is an argument waiting to happen. Insist on the baseline first.

Our own prices are on this site for the reason above: we would rather compete on the work than on the opacity.