The Autonomous Enterprise Needs Clear Decision Rights

The interface is moving ahead of the organisation

At SAP Connect on 6 October, SAP described the customer rollout of Joule Work and assistants that execute work within business processes. The announcement includes customer pilots and early deployments; it does not establish uniform maturity across the installed base.

My reading is that natural-language access makes an old organisational problem harder to ignore. An employee can express an objective in seconds, while the authority to fulfil it may be distributed across several teams. A conversational interface can hide that complexity from the requester. It cannot make the underlying decision rights disappear.

References: [1]

Separate advice, preparation and commitment

A sourcing assistant can propose a supplier, assemble a comparison and draft a purchase request. Committing spend is a different action. The same distinction applies to credit adjustments, employee records and contractual promises. These activities should not inherit a common level of autonomy simply because one agent can perform them.

I would classify actions by the consequence of error and the ease of recovery. Reversible preparation can often proceed with lightweight review. External commitments, sensitive disclosures and difficult-to-reverse changes need stronger approval and evidence. This classification should be owned by the business process owner, with security and technology teams translating it into enforceable controls.

Document the real process before delegating it

Accenture’s 23 September investment in Within focuses attention on discovering how work actually happens, including informal handoffs and exceptions. That commercial move supports a useful question: what would an agent learn if it observed the real process rather than the approved process diagram?

Some workarounds preserve customer service when systems are incomplete. Others bypass controls. Automating both indiscriminately would reproduce the organisation’s unresolved compromises. Process discovery therefore needs judgement: retain, redesign or prohibit each exception before treating it as an instruction for an agent.

References: [2]

Measure the whole decision cycle

Time saved in producing a recommendation is only one part of value. Include the time spent reviewing it, correcting downstream errors and resolving disputes about responsibility. Otherwise a faster front end may simply transfer work to finance, operations or compliance.

A useful pilot compares an entire completed case with the existing process: elapsed time, human intervention, unresolved exceptions and the quality of the recorded decision. Set the baseline before deployment. If the assistant helps staff make better decisions but does not safely execute them, that is a legitimate operating model and should be measured on its own terms.

Where this judgement could be wrong

Decision rights will not be the binding constraint everywhere. A narrow, low-risk workflow may have clear ownership already and still fail because the model cannot interpret its inputs or the application has no reliable interface. In that case, improving capability or integration deserves priority.

The practical test is whether a team can name who may authorise each material action and show how that authority is enforced. Where it cannot, adding autonomy expands an unresolved responsibility gap. Where it can, the organisation has a sounder basis for testing whether greater automation is worth the effort.