An AI assistant finishes a task in thirty seconds. Everyone is impressed. Then somebody spends twenty minutes checking what it did.
The demo clock, mysteriously, has stopped.
That does not make AI useless. It means we need to count the right work. A first draft can be valuable. A sensible summary can make a difficult morning easier. A tool that helps a new employee find the right information can be worth paying for.
But “faster” leaves several questions unanswered. Faster at what? With which mistakes? After whose review? And if the tool really does free up time, what happens to it?
Those are business questions. They are also questions about what kind of workplace we want to build.
There are real benefits worth taking seriously
In a study published in The Quarterly Journal of Economics in 2025, Erik Brynjolfsson, Danielle Li and Lindsey Raymond examined an AI assistant used by 5,172 customer-support agents. Issues resolved per hour increased by 15% on average. Benefits varied: less experienced workers improved more, while the most experienced workers saw small speed gains and a small decline in quality. The agents retained control over what they said to customers. The study concerned one firm's setting; it did not establish economy-wide effects on jobs or wages. Read the published study.
That is a reason to investigate a useful task. It is not a coupon for 15% more productivity at every business with an internet connection.
A separate field experiment across 66 firms and 7,137 knowledge workers examined AI inside familiar work applications. Its November 2025 revision reports that, in the latter half of the six-month experiment, the 80% of treated workers who used the tool spent about two fewer hours a week on email and worked less outside regular hours. The researchers did not detect changes in the quantity or composition of tasks from providing AI to individuals. Read Shifting Work Patterns with Generative AI, version 4.
Less time on email can matter even when it does not produce an obvious new output. Getting home with less unfinished work is a result people can care about. Neither study promises the same outcome for a restoration office, a chauffeur operator or a UGC creator. Each business still has to test its own task.
Count the whole task
Suppose an office employee normally takes twelve minutes to review a job update and prepare the next internal action. An AI-assisted version takes two minutes to prepare, four to review and two to correct.
That is eight minutes of work, not two. In this invented example, the saving is four minutes. Repeat it twenty times and you have eighty minutes of potential capacity, assuming the cases and quality are comparable. You have not automatically created eighty minutes of profit.
Could that time help someone clear a backlog? Return a customer call? Train a colleague? Leave on time? It depends on what the business chooses to do with it.
Track the output too. A shorter review that misses a changed visit date is not a win just because the stopwatch looks better. Compare usable results, including preparation, review, correction and rework. If the task is too risky to get wrong, a time saving does not excuse the mistake.
The people doing the work should help define that test. They know which exceptions are ordinary and which apparently tiny detail creates tomorrow's problem.
A clean summary can hide a messy decision
Consider an invented water damage restoration job. The schedule says the crew will return Tuesday. A field message says access is still unconfirmed. The office is preparing an update for the property owner.
An AI tool could help gather the messages and draft a summary. It should not turn “Tuesday, if access is confirmed” into “Your crew is booked for Tuesday.” A smoother sentence has changed the promise.
The useful output identifies the conflict, points to the records and gives a person a clear question to resolve. The office confirms access and the responsible person approves the customer message. Technical restoration decisions stay with the qualified team.
A workflow is that sequence: what starts the task, which information is needed, who acts and what happens next. When the sequence is unclear, an impressive assistant may simply move the uncertainty into a nicer-looking paragraph.
Our restoration handoff example illustrates this kind of operating question using invented records. It is not evidence of a deployed restoration client system.
Skepticism is useful when it changes the test
“I don't trust it” is a reasonable starting point. The next sentence matters.
If the concern is accuracy, require sources and test missing or conflicting information. If it is data access, start with invented records and decide which information the real task would require. If it is an action taken in your name, keep sending, booking or changing records outside the trial.
Our Muse article looks at this distinction through early users' experiences: finding a message and making a commitment about it are different jobs.
Treat the skeptical employee as a useful reviewer. Ask them to bring the awkward case. Reward finding an error before it reaches a customer. Otherwise the business risks learning more about what people feel safe reporting than about how well the tool works.
Refusing to examine a useful result also has a cost: you may miss a way to remove work nobody enjoys. Curiosity does not require surrendering judgment. Put the tool through a test it can fail.
Decide who benefits before changing the target
If the trial works, there are several legitimate uses for the capacity. A business might serve more customers, improve response quality, reduce overtime, train its staff or protect time for difficult work. An owner may need some of the gain to cover the investment and keep the company healthy.
The decision deserves to be explicit. If every improvement immediately produces a higher quota, do not be surprised when the next improvement meets a chilly reception.
For Ordo Mentis, conscious capitalism means bringing customers, workers and the business's durability into the same decision. Who receives the benefit? Who checks the mistakes? Can the person using the system challenge it? Has the improvement reduced a burden or moved it onto someone less visible?
That is our point of view, not a claim that adopting software automatically creates a fairer workplace. Values have to survive the operating decision.
Start with one task and a decision date
Choose a repeated task with a clear input and a result someone can judge. Use invented or appropriately authorized examples. Include ordinary cases, missing details and contradictions. Write down the expected result before running the tool.
Record the full effort for the current method and the assisted method, with comparable cases. Check correctness separately from speed. Include subscription, setup, training and ongoing supervision when deciding whether the improvement is worth its cost.
Agree in advance what would make you continue, revise or stop. A small trial helps reveal failure modes; it does not prove a low failure rate or establish that the tool can act without supervision.
Then ask the question the demo skipped: if this works, what should the people involved get back?
Want to test one task in your business? Tell us what the team does today, which tools it uses and what a useful improvement would look like. We can discuss a bounded practical AI evaluation before deciding whether implementation makes sense. Describe the task.