Say “outsource the accountability” in a boardroom and the room will usually react before it thinks.
That is useful. The phrase exposes a tension that calmer language can hide.
Getting serious value from AI requires businesses to hand over consequential work. If every task must remain close enough to a person for that person to perform it again, the work has not really been delegated. It has acquired an expensive audience.
But the accountability does not travel to the AI with the task.
An AI system can be responsible for producing an output in the ordinary operational sense: this is the system that reviewed the contract, reconciled the figures or answered the customer. It cannot become accountable in the institutional sense. It cannot hold an office, accept a sanction, repair a relationship or explain itself as a moral and legal participant in the business.
Responsibility can travel further down the working chain than accountability. AI makes the distance between them unusually visible.
This is not entirely new. Businesses have long separated the person doing the work from the person answerable for it. A junior analyst prepares the model; a partner signs the advice. A contractor installs the system; the operator remains responsible for the service. A board delegates authority without ceasing to govern.
AI extends the chain into an actor that can perform sophisticated work without becoming an accountable address.
The default response is to leave accountability where it was. The work moves; the title does not. A finance system may prepare the forecast, question assumptions and draft the board commentary, while the CFO remains answerable for a process they increasingly supervise rather than perform.
This creates the accountable ghost: a person still attached to the consequences of work that has moved beyond their direct production.
The ghost is not necessarily a mistake. Accountability often should remain with the person or organisation that chose the system, supplied its context and benefited from its work. The mistake is allowing the arrangement to emerge without deciding what supervision, evidence or authority the remaining accountable person needs.
An accountable name without the means to govern is not governance. It is a label left behind by the work.
There is a worse outcome. Accountability can become difficult to locate at all.
The operator says the model produced the answer. The model provider says the customer controlled the use. The implementation partner says it configured what it was given. The manager says a person was technically in the loop. Everybody can describe their part; nobody can give the name of the address that finally answers for the result.
This is the accountability void.
The void is rarely created by declaring that nobody is accountable. It is produced by a chain of locally plausible handoffs. Each participant transfers a task, a decision or a risk, and the organisation never tests whether the transfers still resolve into someone with both authority and obligation.
The Air Canada chatbot case provided a compact version of the problem. The airline’s chatbot gave a customer inaccurate information about a bereavement fare. Air Canada argued that the chatbot was responsible for its own actions. The British Columbia Civil Resolution Tribunal rejected the separation and held the airline liable for the negligent misrepresentation on its website.
The decision did not settle AI liability in general. It demonstrated something narrower and more useful: putting an automated actor between a business and its customer does not, by itself, create a new accountable party.
Regulation is also making the allocation more explicit rather than allowing responsibility to dissolve. Under the European Union’s AI Act, providers and deployers of high-risk systems have different obligations. Deployers must assign competent human oversight and monitor operation; providers retain duties concerning the system, its information and its compliance. Accountability is distributed, but it is not handed to the model.
A business therefore has choices, although “the AI is accountable” is not one of them.
It can deliberately retain accountability. This is often the right posture where the work is close to the organisation’s identity, promise or relationship with the customer. A business may automate much of a decision and still say, plainly, that it stands behind the result.
That can be a positive commercial position rather than a defensive concession. When customers care who will answer if something goes wrong, retained accountability is part of the product.
A business can also buy accountability from another organisation. This already happens in familiar professional services. Clients do not merely buy accounting software, legal templates or engineering calculations. They buy the participation of a firm that accepts defined professional duties and carries some of the consequence when its work fails.
AI will make this distinction more valuable. Many providers currently sell capability while leaving the customer answerable for its use. A more consequential category will sell a result together with a credible accountable address.
Call it accountability as a service.
That does not mean a provider can absorb every legal, regulatory or moral obligation a customer holds. Accountability is constrained by law, contract, professional duty and the facts of who controlled what. The useful product is not a magical transfer of liability. It is a deliberate allocation: named duties, evidence, escalation, remedies and someone capable of answering when the system reaches its limit.
The market for this is likely to appear first where the work is important but not distinctive to the buyer. If payroll, compliance checking or routine contract review is not the reason the organisation exists, paying a specialist to provide both capability and a meaningful share of accountability may be more attractive than operating another tool internally.
Closer to the organisation’s core, the opposite may be true. The business may automate the labour while keeping the accountability because standing behind the decision is part of what customers are buying.
These are tendencies, not a clean map. Generic work can carry serious consequences. Distinctive work can still depend on outside accountable specialists. Accountability is its own dimension: it asks who answers for the work, not merely who or what performs it.
That question should be designed into the handoff.
Who can approve the system’s use? What evidence must it retain? Which result requires human judgment? Who can stop the work? Who explains the decision to the affected person? What remedy exists? Which organisation remains standing when every supplier points to its terms?
Without those answers, “human in the loop” is often ceremonial. The person may be present without time, information or authority to change the outcome.
Outsourcing accountability is therefore both the wrong phrase and the right provocation.
It is wrong because accountability cannot simply be deposited in an AI system. It is right because serious delegation forces a business to reconsider where accountability should live, what it requires and whether another organisation is prepared to carry some of it.
The work will move whether or not that conversation happens.
The responsible choice is to make sure accountability arrives somewhere real.