Your AI dashboard can keep going up while use falls

Tesla’s July 22 update says its Robotaxi service is live in seven major metro areas. The same page shows paid Robotaxi miles as a cumulative line rising toward 2.5 million through June. Tesla also says it began Cybercab production, expanded unsupervised operation in Austin, and launched unsupervised rides in three Florida cities in July.

Those are real milestones. They do not answer how much the service was used during the latest quarter. For that, you have to subtract earlier cumulative values from later ones. TechCrunch did that and found a quarter-over-quarter decline. Electrek’s estimate came out flatter rather than lower, but it also found that the rate slowed after April.

Neither outside estimate should be treated like an audited monthly table. Both were reconstructed from the shape of Tesla’s graph. The useful point is simpler: readers should not need to reverse-engineer a cumulative curve to learn whether current use rose, fell, or held steady.

Cumulative miles are not a fake metric. They can show total operating exposure, experience gathered, or distance covered since launch. They are simply a poor answer to a momentum question. The same distinction applies to AI adoption inside a company.

Suppose a 20-person company rolls out an AI assistant for customer follow-up. In month one, the team generates 400 drafts. In month two, it generates 250. In month three, 90. A lifetime total of 740 drafts still points sharply upward. The weekly habit is dying.

There may be a good reason. The assistant could have cleared a one-time backlog. A seasonal rush could have ended. The team may have learned that only a narrow set of messages benefits from a draft. A falling rate does not automatically mean failure.

It does mean someone should ask why. Lifetime totals suppress that question. They combine yesterday’s enthusiasm with today’s behavior, then present the sum as current health.

The same problem shows up in “hours saved.” If a vendor asks users how long a task used to take and multiplies that estimate by every AI run, the total can grow forever. It may not subtract prompting, checking, corrections, duplicate work, abandoned outputs, or time spent explaining the tool to the next person. The number measures claimed gross time, not time returned to anyone’s day.

Start with the rate. Show completed AI-assisted tasks by week, not only since launch. Put the number beside the count of people eligible to use the tool. Ten weekly tasks from a three-person finance team can mean more than 100 from a 500-person company.

Then measure repeat use around a named job. A person opening an assistant twice is not necessarily adoption. A team using it for the same invoice check every Friday, without also doing the old check, is closer. Track how many people return to the same useful task after the novelty period.

Finally, count repair. How many outputs were corrected, rerun, abandoned, escalated, or redone outside the tool? How many support questions did the rollout create? How much reviewer time did a completed task consume? These numbers are less flattering than a lifetime run count. They are also the numbers a manager needs before renewing the contract.

A practical dashboard can fit on one screen: weekly accepted outcomes, repeat users for those outcomes, old work actually retired, median checking time, correction rate, and unresolved failures. Keep the lifetime total if it helps with capacity planning. Do not let it sit alone at the top.

Mara Vale’s objection is not that cumulative numbers are false. It is that they are socially convenient. The buyer gets a line that still rises after enthusiasm cools. The vendor gets a success slide. The employee who returned to the old process becomes invisible unless someone asks for the weekly rate and the repair load.

Noah Park is less interested in prosecuting the chart than replacing it with a quick test. Pick one recurring task, compare the last four ordinary weeks, and count accepted outputs, repeats, corrections, and the old steps that disappeared. If the task is not quieter by week four, a larger lifetime number will not rescue it.

Both views lead to the same buying rule: a useful AI metric should make it easier to stop, narrow, or improve a rollout. If the dashboard can only congratulate the rollout, it is marketing furniture.

What happened in the latest complete week or month? Ask for a period rate beside the lifetime total. If usage is seasonal, compare it with the same period or the team’s actual workload rather than forcing every dip into a failure story.

How many people came back to the same task? New-user counts and one-off experiments are useful during launch. After that, repeat behavior around a real job matters more.

What stopped happening? If the assistant drafts a report but the old manual report still gets made “just in case,” the company added activity without removing work.

What had to be repaired? Count review minutes, corrections, retries, reopened cases, support requests, and abandoned outputs. A completed run is not the same as an accepted result.

Who got time back? Faster output can become more assignments, a denser queue, or an evening spent checking. Name the person and the pressure that actually dropped.

Tesla’s chart does not settle whether its Robotaxi rollout is succeeding. It does show why one rising line cannot settle the question. The same is true for AI assistants at work. Totals tell you how much has happened. Rates, repeat use, retired work, and repair tell you whether the tool still deserves a place in the week.