Ai tools fail at work because users treat them like magic vending machines
Thousands of corporate teams have shelved generative AI after the first week. The culprit is not buggy code or hallucinations—it is the instruction manual written inside the human brain.
Tom Hewitson has watched the crash-and-burn cycle repeat in conference rooms from London to Lisbon. The corporate trainer, who has taught more than 12,000 professionals how to coax value out of large-language models, says 70 % abandon the experiment before the first invoice is paid. The reason is humbler than any algorithmic flaw: users expect a finished answer instead of a sparring partner.
The mirage of the instant oracle
Hewitson’s data set is intimate. He sits shoulder-to-shoulder with lawyers, nurses and marketing executives as they type their maiden prompt. Almost universally they frame a single, immaculate question, then wait for the machine to vomit perfection. When the reply arrives—verbose, laced with caveats, occasionally off-topic—two camps emerge. The first group copy-pastes the slop and forwards it to the boss, triggering a later disaster. The second group smashes the „dislike“ button and declare the tool „not ready“. Both miss the lever that actually works: iterative friction.
„They think ChatGPT is a vending machine,“ Hewitson told TechFlux. „But it behaves like a jazz trio. You throw out a riff, it answers, you correct the tempo, and only after three or four cycles does the solo emerge.“
Prompt engineering is attitude, not syntax
Surprisingly, the employees who end up extracting 40 % productivity gains are rarely the Python-savvy engineers. They are the administrative assistants who treat the chat window like a chaotic colleague. They rephrase, poke, paste half-finished paragraphs back in and demand tighter analogies. Over weeks they build a private recipe book: when to ask for bullet points, when to switch personas, when to demand counter-arguments. The company sees the payoff in shorter report-drafting cycles; the worker pockets a new line on the CV.
The medical profession offers a stark counter-example. Dr. Mieses Malchuk, an emergency physician in Barcelona, unplugged the hospital’s pilot AI scribe after three shifts. „It turned triage into a circus of copy-editing while patients piled up,“ he said. The tool was not faulty; the deployment model was. Staff had zero minutes for the iterative dance.
Corporate licences rot in the drawer
Enterprises bought 4.3 million ChatGPT Plus seats in 2023, yet internal logs show median active use at 5.8 days per licence, according to data aggregator UsageMeters. The pattern is consistent: aggressive C-suite rollout, two-week honeymoon, radio silence. CIOs blame „data privacy anxiety“ or „workflow integration issues“. Hewitson blames folklore. „We keep telling workers the AI is smart. We never tell them it is still a mirror that only polishes what they feed it.“
His remedy is brutally low-tech. Every workshop begins with a paper exercise: write the same request five times, each version narrower than the last. Participants watch the digital responses tighten in real time. The tactile loop rewires expectations faster than any slide deck.
The salary premium is already here
LinkedIn’s July workforce report lists „AI interaction strategist“ as the fastest-growing job title in Europe, median salary €78,000. The description demands no coding, only „documented prompt refinement cycles“. Translation: companies will pay extra for humans willing to argue with a machine until it behaves.
Hewitson’s prediction is colder. Within two years, annual reviews will stop asking „Do you use AI?“ and start asking „How many iterations do you run before you ship?“ Employees who answer „one“ will find themselves on the outfield bench, vending machine coins scattered at their feet.