Where does AI help, and where should a person decide?

Help versus decide

There is a difference between helping with a decision and making it.

Helping looks like this: summarizing twenty customer reviews into the five most common complaints, drafting three versions of a price-change notice, sorting incoming emails by topic, or listing questions to ask a supplier.

Deciding looks like this: choosing to raise prices, approving a refund, hiring or letting someone go, signing a contract, or telling a customer what they are legally entitled to.

NIST, the U.S. standards agency, describes a range of arrangements between people and AI, from fully automatic to fully manual. An AI system can make a decision on its own, hand the decision to a human expert, or be used by a person as an additional opinion [1]. NIST also notes that some AI uses need little human oversight while others specifically require it [1]. Your job as an owner is to decide, for each task, which arrangement fits.

Why people stay in charge of the big calls

AI can be confidently wrong. Lesson 3 covered how AI can state false information as if it were certain. NIST notes this risk deserves special attention when AI is used in decisions with serious consequences [2].

People tend to over-trust machines. NIST describes "automation bias," the tendency to defer too much to automated systems or to see AI output as higher quality than it is [2]. The more polished an answer looks, the easier it is to approve without thinking. Knowing this bias exists is the first defense against it.

Accountability stays with you. Customers, employees and regulators hold your business responsible for what it does, whether a person or a tool produced the first draft. One Federal Trade Commission case shows how this plays out. In February 2025, the FTC announced a final order against a company that sold an online "AI lawyer" service. According to the FTC's complaint, the company did not test whether its "AI lawyer" performed at the level of a human lawyer, and the order bars it from advertising that the service performs like a real lawyer unless it has sufficient evidence [3]. The lesson for any business: do not let a tool stand in for professional judgment it has not been shown to match.

Context matters. You know the long-time customer who is having a hard month, the employee who just had a family emergency, the supplier who always comes through. AI does not, unless you tell it, and even then it does not carry the relationship.

A simple sorting test

For any task, ask three questions:

1. Can the mistake be undone easily? A typo in a draft is easy to fix. A wrongly denied refund or a rejected job applicant is not.
2. Does it affect a person's money, job or rights? Decisions about hiring, pay, credit, refunds or legal commitments belong with a person. NIST names employment and lending as examples of consequential decision-making settings [2].
3. Would I be comfortable explaining it to the person affected? "Our software decided" is not an explanation most people will accept.

If the answer to the first is "no," or to the second is "yes," a person decides. AI can still help prepare.

Tasks that usually suit AI

• First drafts of emails, posts, job descriptions and policies
• Summaries of long documents, reviews or meeting notes
• Sorting and tagging incoming messages for a person to handle
• Turning notes into checklists and step-by-step instructions
• Brainstorming names, ideas and questions

Tasks where a person should decide

• Hiring, firing, discipline and pay
• Approving or denying refunds, credit or exceptions
• Pricing changes and discounts
• Anything you sign or promise in writing
• Legal, tax or financial positions
• Public statements in a sensitive situation

Common mistakes

Rubber-stamping. Approving AI output without reading it turns "a person decides" into "the tool decides." Review means actually reading.

Letting AI make a promise. A chat tool that tells a customer "you'll get a full refund" may have just committed your business to something you never agreed to. Keep promises and policies with people.

Using AI to avoid hard conversations. Difficult news about a job, a price or a dispute deserves a person's voice.

All or nothing thinking. It is not "AI does everything" or "AI does nothing." Most tasks split cleanly into AI preparation and human decision.

Worked example

A car dealership's sales manager receives 40 online inquiries over a weekend.

What AI does: reads each inquiry, tags it by vehicle and urgency, and drafts a short reply that thanks the person and asks one or two questions about what they are looking for. It also produces a one-page summary of the weekend: which models people asked about most and which inquiries mention a trade-in.

What people do: the sales manager reviews the summary and decides who follows up with which leads. Each salesperson reads the draft reply, adjusts it, and sends it. Any question about pricing, financing terms or trade-in value is answered by a person, because those answers are commitments.

AI saved the morning sorting work. Every decision that could affect a customer's money or the dealership's word stayed with a person.

Checklist

• For each AI task, I have decided whether AI helps or decides.
• Decisions about people's jobs, money or rights stay with a person.
• AI does not make promises or quote terms to customers on its own.
• Reviewers actually read AI output before approving it.
• I can explain any decision to the person it affects.

Try it

List ten tasks you or your staff did last week. Next to each, write "AI drafts, person decides," "AI can handle with spot checks," or "person only." Use the three sorting questions. You now have a starting map for where AI fits in your business.

Tips

• Use AI to prepare decisions, not to make the important ones.
• Watch for automation bias: polished output still needs a real read.
• Keep anything that commits your business in writing with a person.
• Revisit your list every few months as your experience grows.

Sources

1. National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework (AI RMF 1.0)," NIST AI 100-1, January 2023, https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
2. National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile," NIST AI 600-1, July 2024, https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
3. Federal Trade Commission, "FTC Finalizes Order with DoNotPay That Prohibits Deceptive 'AI Lawyer' Claims, Imposes Monetary Relief, and Requires Notice to Past Subscribers," press release, February 11, 2025, https://www.ftc.gov/news-events/news/press-releases/2025/02/ftc-finalizes-order-donotpay-prohibits-deceptive-ai-lawyer-claims-imposes-monetary-relief-requires