Do not fully automate a task when a wrong answer would be costly or hard to undo, when it creates a legal or contractual commitment, or when nobody on your team can check whether the output is correct. AI can still help with those tasks as a drafting or analysis aid, but a named person should decide and take responsibility. Everything else is a candidate for automation, starting with the low-risk and easy-to-check work.
The three-part test before you automate
Ask three questions about any task. If the answer to any of them is a clear "yes", keep a person in charge of the final step.
- What is the cost of a wrong output? A typo in an internal summary is cheap. A wrong figure sent to a customer, regulator or board is not.
- Can the result be reversed? A draft email can be deleted; a payment, a signed contract or a message already sent cannot.
- Can someone verify the output quickly? If checking the result takes as long as doing the work, or needs expertise nobody on the team has, automation adds risk instead of removing effort.
High-stakes decisions that need a human owner
AI is good at preparing the ground for a decision: summarising documents, laying out options, listing risks. It should not be the one that decides when the outcome affects a person's livelihood, safety, money or reputation. Typical examples are hiring and firing, disciplinary action, credit or pricing exceptions, medical or safety judgements, and anything involving vulnerable customers.
The reason is accountability, not capability. A model cannot be held responsible, and "the AI said so" is not an answer your company can give to a customer or an auditor. Keep a named decision-maker, and let AI shorten their preparation time. This idea is often called human-in-the-loop.
Legal and contractual commitments
Do not let an AI system send, sign or accept anything that binds your company without review: contracts, quotations with terms, settlement offers, compliance declarations, or replies to a legal notice. A model can produce fluent text that sounds right while quietly changing an obligation, a date or a liability clause.
Using AI to draft, compare or summarise legal documents can save time, but a qualified person should review them before use. This article is general information and not legal advice; confirm requirements with your legal or compliance team. If you are drawing up internal rules, see what a company AI usage policy should include.
Work where the output cannot be verified
Language models can state wrong things confidently, an effect described in what an AI hallucination is. That is manageable when you can check the result against a source. It is dangerous when you cannot. Be cautious with:
- Citations, case references and regulatory references nobody will look up.
- Numbers pulled from long documents with no cross-check against the original.
- Technical or specialist advice outside your team's own expertise.
- Anything where the only reviewer is the same AI that produced it.
A useful rule: automate only what you would be comfortable spot-checking. Our guide on how to check if AI output is accurate covers practical checks.
Sensitive data and confidential situations
Some tasks are poor candidates because of what flows through them rather than what comes out: personal data, customer records, unreleased financials and privileged communications. Whether and how such data may be used with an AI tool depends on your contracts, your tool settings and your obligations, so decide this with your compliance team first. See can you put customer data into Claude for the questions to ask.
Tasks that need empathy and judgement
Delivering bad news, handling a complaint from an upset customer, negotiating a sensitive deal or managing a team member's performance rely on trust and reading the room. AI can help you prepare, for example by drafting talking points or rehearsing scenarios, but the conversation itself should be human.
Good candidates to automate first
The flip side of the test is a list of safer starting points. Tasks that are repetitive, low-stakes, reversible and easy to check are where automation pays off fastest: formatting, first drafts, meeting notes, internal summaries and routine categorisation. Start there, measure the result, then widen scope as your team builds checking habits. Our guide on how to run an AI pilot shows how to do this in a controlled way.
A simple way to keep control as you scale
- Draft, then approve. AI prepares; a person approves anything that leaves the company.
- Escalate by risk. Define which cases always go to a person, such as high value, complaints or legal wording.
- Log and sample. Review a sample of automated outputs regularly, not just when something goes wrong.
- Name an owner. Every automated workflow has one person accountable for it.
These ideas map to workflow integration and governance topics in the CCAO-F workflow integration domain and the governance and risk domain. For senior leaders deciding where AI belongs in the business, the AI Forward CXO Program covers this kind of decision-making.