Artificial intelligence has a rare office skill: it can be so confidently wrong that you feel you should apologise. Ask for five sources and you may get five beautifully formatted links. You still need to open them. That is sometimes where the magic ends and the work begins.
We build AI products at White Studio, so declaring AI the answer to everything would be convenient. Businesses already have enough subscriptions to a brighter future, though. Let’s look at what is worth using, where people remain essential and why “make it look good” is still a rather incomplete brief.
The useful work starts with the blank page
Summarising a long email, structuring a presentation, grouping customer questions, drafting alternative scripts or turning a meeting transcript into actions are sensible starting points. You have the source, can check the output and haven’t given a machine control of the bank account. A mistake in a draft needs an edit. A mistake in an automatically sent quotation needs an awkward conversation.
In Noy and Zhang’s 2023 Science experiment, participants using ChatGPT completed professional writing tasks 40% faster on average, with assessed quality rising by 18%. Those were particular tasks, not a promise to halve every working day.
A useful question is: “Which repetitive operation can we remove?” Asking “Where could we put some AI?” tends to produce an attractive widget that answers everything except the customer’s question.
Faster sometimes. No universal multiplier
The 2023 working paper Generative AI at Work reported roughly 14% more issues resolved per hour among 5,179 support agents. Gains were uneven: newer workers benefited more than experienced colleagues.
Meanwhile, a 2025 METR experiment involving 16 experienced developers and 246 tasks found that access to the AI tools studied increased completion time by 19%. It concerned large repositories familiar to participants, not all software development.
These numbers describe different settings and generations of tools. Treating either as a universal forecast for 2026 would be exactly the confident oversimplification we are discussing. Measure your own process: time including review, correction costs and the proportion of outputs actually accepted. Paragraphs generated are a poor measure of value. Printers are productive by that standard too.
A convincing answer is not yet a fact
NIST’s 2024 generative AI risk profile identifies confabulation: systems can present incorrect content confidently. Privacy and security risks also need assessment.
“Don’t invent anything” is a useful instruction, but it is not verification. A price needs a current price list; a delivery date needs confirmation; a product claim needs a specification. When a chatbot cannot find an answer in approved sources, handing the question to a person is a valid outcome. “Let me check with the team” can be more sophisticated than three paragraphs of imaginary stock availability.
Documents raise a similar issue. Check that all pages and tables were extracted before discussing conclusions. A model can explain half a contract brilliantly while knowing nothing about the other half. Eloquence does not compensate for missing inputs.
A familiar voice with an unfamiliar request
The FTC warns about scams using cloned voices. If a “relative” urgently requests money, call a known number or verify through another trusted person. A recognisable voice alone is not proof of identity.
At work, that becomes a straightforward rule: a voice message from the director does not replace the payment approval process. Especially when the “director” demands haste, forbids a callback and has developed a sudden enthusiasm for cryptocurrency.
What we would keep off autopilot
Do not paste passwords, API keys, payment details or other people’s personal data into a random chat for convenience. Before using work documents, establish who receives them, where they are stored and who can access them. An “AI” label is not a confidentiality agreement.
Do not run suggested code on a production server without review and a backup. Do not give a bot unrestricted authority to delete records, transfer money or publish statements. Do not impersonate people to create reviews, evidence or messages. And do not make an AI answer the sole basis of a medical, legal or financial decision.
A practical sequence is draft, review, act. The more costly the mistake, the stronger the check. Preparing a purchase request is one permission. Approving a million-pound order is quite another.
Start without a miniature digital revolution
Choose one manageable task: routine enquiries, incoming requests or script options. Gather real examples, define an acceptable result and test difficult cases: an empty file, conflicting prices, an unknown product or an attempt to obtain someone else’s data. Name the person responsible for unresolved decisions.
A pilot reveals what you actually save. Perhaps three hours a week. That is less dramatic than “replace an entire department”, but three real hours can be spent on real work — or a lunch eaten at a reasonable speed.
Those are the results we like. AI helps us build video, chatbots and useful tools when the task has boundaries, reliable inputs and an accountable person. Keep human intelligence switched on. It is already included in the project.
