Choose a Useful AI Task

A useful AI task has a bounded job. You can state what goes in, what should come out, and what a person will do next. 'Turn these three comments into themes' is bounded because the source comments and requested output are visible. 'Fix customer satisfaction' is not a job; it mixes diagnosis, decisions, communication, and follow-through. A bounded task gives the assistant a narrow contribution and gives a reviewer something concrete to inspect.

Think of an assistant as a junior preparer at a large table. You may ask it to group sticky notes or draft a first paragraph, but you do not ask it to sign the purchase order sitting beside the notes. The analogy has a limit: a person can notice context, responsibility, and consequences in ways a text generator does not. Use the analogy to decide what preparation is useful, not to pretend the system has an employee's judgment.

Preparation changes the shape of information: summarize, sort, translate, propose headings, or draft language from supplied facts. Authority changes the real world: approve a refund, set a price, choose a candidate, promise a delivery date, or send a message. The same draft can be helpful before authority is exercised. A draft refund reply can explain the documented policy; a person still decides whether this customer qualifies and presses send.

This distinction protects quality as well as accountability. A concise, confident answer may contain an invented detail or may miss an exception in the record. NIST's generative-AI profile treats confabulation and over-reliance as risks to manage. The practical response is not a slogan about being careful. It is a visible review point tied to evidence: the contract, policy, ledger, or application that supports the next action.

Put each request in one of three lanes. Use means the output is a low-consequence preparation aid and can be checked quickly. Use with review means the output may influence a customer, payment, public claim, or internal decision, so a named person compares it with a source before acting. Do not use means the task asks the assistant to make a decision, invent a missing fact, or handle information that should not enter the exercise.

Lanes are not labels for entire professions or tools. The same assistant can help with a harmless subject line and be unsuitable for an unsupported contract price. The lane changes when the evidence changes. If a pricing table arrives and a pricing manager owns the decision, the assistant might format options for review. It still does not become the pricing authority.

'A human will check it' is too vague to guide work. Name the reviewer, the evidence, and the decision that follows. For a comment summary, the customer-success lead checks that every theme appears in the three comments and that no complaint was added. For a refund draft, the finance lead checks the transaction record and policy before deciding approve, deny, or ask for more information.

A good matrix also records a stop condition. Stop when the source is missing, when two sources disagree, when a request asks for a price or approval, or when the output contains a claim nobody can trace. Stopping is productive: it moves the work to the person who can resolve the uncertainty instead of producing a polished guess.

Brightlane has only the six requests below; no private customer documents are used.

Brightlane receives: summarize comments, invent a contract price, draft an email, approve a refund, classify repeated questions, and decide a hire.

A fictional flower shop wants help with return emails.

Classify these jobs: group ten return questions, draft a reply from a supplied policy, approve a refund above $100, and promise a replacement date when inventory is unknown.