We're teaching AI to be human. Are we teaching businesses to be less human?

These systems produce plausible language and useful drafts at a speed that changes what a team can produce and how much review has to be designed around it. That is a real capability, and overstating it is a mistake as well. What matters operationally is not a claim about the system's inner life, which this article does not settle, but that nothing is at stake for it: it carries no accountability, cannot be called to account by a customer, and has no standing to weigh what an honest answer costs the relationship. A named person or business remains answerable for what goes out.

Plausibility is also not correctness. A confident paragraph and an accurate one look identical on screen, which is precisely why the review step cannot be optional in work that carries consequence.

Most companies are not deciding to become less human. They are removing friction, one task at a time, and friction is often where judgment lived. A complaint arrives, a draft reply appears, it sends in seconds. The customer is answered. Nobody read it closely enough to notice the case is the third time this month, or that the promise inside it cannot be kept.

The risk is not that machines take over. It is that attention gets delegated quietly, and the signal that used to reach a person stops arriving.

Four questions sort most tasks. Is the work repeatable enough that the same input should produce the same output? Is the input well formed, or does it require reading a situation? Is the cost of an error visible quickly and recoverable? Does the customer need a person on the other end?

• Usually delegable with review: summarising a call, drafting an internal note, formatting a proposal, preparing a first-pass agenda, tagging and routing.
• Delegable only with a named reviewer: customer-facing drafts, pricing and scope summaries, marketing claims, and anything that will be quoted publicly.
• Keep with a person: committing the business to something, apologising, refusing, handling an exception, delivering difficult feedback, and deciding what a relationship needs next.

A single review policy applied to everything is either too slow for low-stakes work or too thin for high-stakes work. Four levels cover most situations: sample review for internal drafts nobody outside sees; named review before send for anything a customer reads; human decision with tool-assisted drafting when the output creates an obligation; and human-accountable decisions, with a named owner, for irreversible commitments and for consequential personal situations such as employment, credit, health, or disciplinary outcomes. Everywhere else the test is proportionate: the higher the stakes and the harder an error is to reverse, the more human attention the work needs.

The value of writing the levels down is not compliance. It is that the choice gets made once, deliberately, instead of being decided by whoever is busiest that afternoon.

A worked example, clearly hypothetical. A complaint arrives on a Friday and a tool drafts a reply in seconds; it reads well. The reviewer's job is not tone but three questions: does this commit us to anything, is this the third time this month, and does this customer need a person. If any answer is yes, the draft is a starting point rather than the response.

If a business rewards volume and speed, tools multiply volume and speed, including the parts that should never have scaled. If leaders stop reading customer messages because a summary arrives each morning, they lose the raw language that made them good at their jobs. If a team is measured on output, drafting assistance will produce more output and less thought, and the measures will call it improvement.

This is the human side of the question. Technology and the choices around its deployment can reinforce the habits a culture already rewards, or reshape them; which happens depends on decisions leaders make about scope, review, and measurement. That is why the decision belongs with leadership rather than with whoever buys the licences.

• Which tasks may be drafted with a tool, and which may not.
• Which output requires a named human reviewer, and who holds that role.
• Where the business will say plainly that a person reviewed something, and where it will not claim to.
• What happens when an error reaches a customer: who contacts them, what is disclosed, what changes.
• Who owns the policy, and when it is reviewed.

None of this requires a technical opinion about model architecture. It requires the same judgment a business applies to any other delegation: what is the task, what is the risk, who answers for it.

Answer-engine work is a good example of the distinction. Structuring content so a search or answer system can find and quote it is legitimate supporting work; it does not replace the model of how a market is served, and it does not guarantee visibility. Volume without editorial judgment produces material nobody trusts, which is the opposite of an asset.

Where this fits: the delegation decisions are strategy questions, handled in Strategy & Transformation; the customer-facing and content applications sit in Marketing, Media & Digital; and the process and system side belongs in Operations & Systems. Two practical companion reads are integration before automation and AEO content planning.

This is an argued perspective, not a forecast, and it promises no outcome. It is simply the position we would defend in a leadership meeting: teach the business to be deliberate about what it hands over, and the tools can make it more capable without making it less human.