What AI is good at in everyday work
Language models are useful for the parts of a job that involve reading and writing: drafting a reply from your own notes, summarising a long thread, sorting incoming requests by topic, pulling fields out of documents, or collecting material for a decision.
They are also sometimes confidently wrong. A model can state a figure that is not in your data, or answer a question it did not understand. That is manageable when its output is a draft that a person reads, and costly when its output goes straight to a customer.
Where it should stop
A simple rule: the model may prepare, a person decides. Keep a human step before anything that is hard to undo:
- Sending anything to a customer, a supplier or the public.
- Moving money, issuing invoices or changing prices.
- Legal, medical, financial or employment decisions.
- Deleting or overwriting records.
Five design choices that keep a person in control
- Give the model only the context it needs for the task, not access to everything.
- Ask it to show its sources, so a reviewer can check a claim in seconds instead of redoing the work.
- Put a review step with a named owner before the result is used, and make approving a deliberate action.
- Keep a record of what the model produced, what the reviewer changed and what was approved.
- Start with drafts. Only automate an action once the drafts have been reliably right for a while.
Privacy: which data goes where
With a hosted model, the text you send leaves your computer and is processed by the provider. Before building anything, decide which data may be sent at all, and read the provider's terms for how it is handled.
When the data must stay on your own machine, a local model can do the work. Local models are usually slower and less capable than the largest hosted ones, so they suit narrower tasks such as sorting, extracting or summarising your own documents.
Running costs
Hosted models are usually billed by use, so the cost grows with the amount of text processed. Estimate it on a sample of real work before you commit, and include the cost of the review step: a person's time is often the larger part.
What this looks like in practice
Signal Drafts keeps the research sources attached to every draft and records the exact text that was approved. Shift Brief checks each fact in a generated brief against the records it came from. Control Room gives agent work a separate verifier and a trail of what happened.
Our AI integration with human controls page describes how we build tools like these for a specific task.