The New Standard for Business Processes in the AI Era

A broader unit of design

A business process includes the work, decisions, knowledge, systems, and responsibilities that connect an initiating need to an intended outcome. Its boundaries should reflect the business promise being made. In the billing example, that promise includes resolving the customer's concern, correcting any verified error, and addressing a confirmed cause of recurrence.

Many organizations already pursue these aims through process management, service design, quality disciplines, and automation. AI creates additional ways to interpret information, retrieve relevant knowledge, and support or execute selected actions. The opportunity is to reconsider the complete process as these capabilities develop.

Evidence supports that opportunity without establishing a universal result. In a study of 5,172 customer-support agents, Brynjolfsson, Li, and Raymond found that access to an AI assistant increased issues resolved per hour by 15% on average. Benefits varied substantially with experience and skill. That is evidence from a particular deployment, not a forecast for every process. Generative AI at Work, published in The Quarterly Journal of Economics, 2025; accessible author manuscript.

The standard proposed here is a set of executive design expectations. Together, they describe what a process should increasingly be capable of doing and what leaders should require evidence for. They leave room for different technologies, different degrees of autonomy, and deliberate human control.

Seven expectations for a modern business process

1. Reliable outcomes with flexible execution

A well-designed process should maintain a clear outcome while allowing the route to reflect the circumstances. Some work is predictable and benefits from fixed rules. Other work involves incomplete information, competing obligations, or exceptions that require judgment. The design must distinguish these conditions and provide a valid path through each.

In the billing example, a straightforward duplicate charge and a disputed interpretation of a contract should not receive identical treatment. The first may qualify for a controlled correction. The second may require an authorized person to interpret the agreement and communicate a decision. Both need an Outcome owner and a clear definition of completion.

That definition should extend beyond a closed ticket. Was the adjustment accurate? Did the customer understand the resolution? Was the next bill checked where recurrence was plausible? Where an upstream correction takes longer, the customer case and the underlying problem need linked ownership so that closing one does not erase the other.

Reliability also requires attention to decision quality. A sound decision can encounter an adverse outcome; a weak decision can benefit from luck. Leaders should assess whether the process used appropriate evidence and authority as well as whether it produced the intended result.

2. Context travels with the work

The next person or system should receive the relevant history, current facts, unresolved questions, and reasons for prior decisions. Employees should not have to reconstruct the same situation every time responsibility changes.

For the recurring charge, this means connecting the earlier investigation, the customer's agreement, the credit already issued, the applicable billing rule, and the status of any corrective action. It also means checking what has changed. Last month's conclusion may no longer apply if the contract, product, or customer's circumstances have changed.

Context needs sources, dates, ownership, and access appropriate to the role. A useful summary should make it possible to examine the underlying evidence and distinguish confirmed facts from assumptions. NIST's Generative AI Profile identifies the risk of confidently presented false content, including explanations and citations. A generated rationale cannot substitute for records of what actually informed a decision. NIST, Generative AI Profile, 2024.

The organizational implication is an Enterprise Knowledge Backbone: maintained knowledge connected to work, with responsibility for its quality. The aim is to reduce dependence on individual memory and repeated searching while preserving the expertise needed to challenge incomplete or misleading information.

3. Human and machine roles are deliberate

Work should be allocated according to the nature of the task, the consequences of error, available evidence, and the authority required. AI might assemble the billing history and identify a likely discrepancy. Rules might calculate an adjustment. A person might interpret an ambiguous commitment, consider unusual circumstances, or authorize an exception.

Human participation needs to be meaningful. A reviewer requires enough time, information, competence, and authority to question a recommendation. Routing every difficult case to an already overloaded expert does not create a sustainable operating model. Neither does asking someone to approve an output they cannot reasonably verify.

A 2024 meta-analysis of 106 experiments found that human–AI combinations improved on humans alone on average, but performed worse than the better of humans or AI alone on average. The studies covered work published from 2020 to mid-2023. Their lesson for design is to test the particular arrangement; combining people and AI does not by itself establish better performance. Vaccaro, Almaatouq, and Malone, When combinations of humans and AI are useful.

Contextual guidance can help more employees handle work competently and reduce reliance on a few individuals. It should support training, practice, and professional development. People doing the work should help shape the design, with clarity about changing responsibilities, workload, and how concerns will be addressed.

4. Governance operates within execution

Authority should be expressed in the way work proceeds: who can access information, recommend an action, approve it, execute it, and review the result. Conditions for escalation and intervention need to be enforceable in the systems performing the work.

In the billing example, permission to prepare a credit should be distinct from permission to issue it. An adjustment should require the relevant evidence, confirmation that it has not already been made, and authorization within agreed limits. A change to a billing rule may affect many customers and deserves a different approval and validation path from a single-case correction.

This is the practical role of Bounded Autonomy. The scope of delegated action can grow as evidence supports it, with business accountability remaining explicit. A model's confidence is one input to evaluate; it cannot grant permission.

NIST's AI Risk Management Framework calls for defined responsibilities, clarity about human–AI roles, and continuing monitoring and risk management. These expectations support designing governance into operations. They also require owners who can review effectiveness and change the boundaries when conditions warrant it. NIST, AI RMF Core, 2023.

5. Performance is visible and recovery is bounded

A process should make its state and consequences visible enough for people to intervene. Technical availability is only part of this picture. Leaders need to know whether cases are progressing, obligations are being met, errors are accumulating, and work is moving into unowned queues.

For the billing process, relevant measures include total resolution time, accuracy, repeat contacts, recurrence, customer effort, and the effort absorbed across participating teams. Results should be examined across different case types so that a favorable average does not conceal a recurring problem for a particular group.

Recovery must also respect the business consequences of an action. If a credit request times out, an automatic retry could create a second credit. The process should first establish whether the initial action took effect. Where that cannot be determined, it should preserve the case state and escalate to an owner with a workable recovery path.

Some failures can be corrected automatically within tested limits. Others require stopping, restoring a prior state where feasible, or moving to staffed fallback arrangements. NIST includes monitoring, override, incident response, recovery, and change management in its post-deployment expectations. These are capabilities to demonstrate under realistic conditions. NIST, AI RMF Core.

6. Learning becomes shared capability

Experience should improve the process beyond the person who handled the latest case. That requires a deliberate path from an operational signal to investigation, a verified conclusion, an approved change, and evidence that the change worked.

Suppose the billing investigation confirms that a discount rule is being applied incorrectly to a defined customer group. The organization can assess the affected population, correct the rule through its change process, update guidance, and check subsequent bills. The original customer receives follow-through, and other customers may avoid the same problem.

The lesson is broader than storing a case summary. Knowledge, instructions, controls, training, or the process itself may need to change. A workaround that closed one case should not automatically become a rule for future cases. Conflicting evidence should remain visible until someone with the relevant responsibility resolves it.

The Continuous Evolution Loop describes this connection between operational signals and approved improvement. Learning becomes an enterprise capability when people can reuse a validated correction and when its effects remain open to review.

7. The process can absorb change

A process should be designed to adapt as products, customer expectations, policies, technology, and organizational responsibilities change. This depends on knowing which rules, knowledge, interfaces, and roles support the outcome, and how a change in one may affect the others.

If the business introduces a new subscription plan, the billing process should have a practical way to update eligibility rules, customer communications, employee guidance, and test cases together. A change should be versioned, evaluated against representative work, and introduced with a way to detect and address unintended effects.

Over time, the ambition could extend to identifying a likely recurring error before the next bill is issued and initiating an authorized preventive action. That is a possibility to establish through evidence. It depends on reliable signals, permitted access, validated decisions, and clear responsibility for the customer consequence.

Adaptability has an economic purpose. It should reduce the cost and disruption of worthwhile change, and make new service promises feasible. The organization still needs to decide which changes merit investment and which dependencies it is prepared to maintain.

Connect the destination to operating reality

These expectations describe an ambitious destination. Progress begins with an honest account of the present process: its outcomes, variations, informal work, knowledge gaps, costs, and dependence on particular people. The Digital Operating Model places those process choices within the wider design of capabilities, roles, governance, and technology.

Leaders should consider three questions together. What could available technology make possible, and where does its performance remain uncertain? What can the organization operate responsibly, given its skills, knowledge, systems, authority, and resources? What would create enough customer or enterprise value to justify the full cost and fit the organization's mission?

The destination should include opportunities that the current process cannot readily support. In the billing example, this might be a more dependable customer promise: identifying and resolving certain errors before a customer needs to contact the business. Its value would depend on customer experience, preventable loss, cost, and the consequences of incorrect intervention.

Economics should include integration, evaluation, knowledge maintenance, human review, exceptions, recovery, and continuing improvement. Time released in one team becomes business value through an explicit use, such as absorbing demand or improving service. Additional review elsewhere can consume that benefit. The relevant measure is the cost and value of acceptable outcomes across the complete process.

The destination can remain ambitious while the next increment is modest. If the immediate constraint is fragmented case history, begin by making reliable context available. If responsibility is unclear, resolve ownership. If a stable rule can remove a known defect, implement and verify that correction. The choice of AI, conventional automation, or human judgment should follow the work and the evidence.

Progress through evidence from real work

Choose a bounded part of the process and establish a baseline before changing it. Include people who perform the work and those who receive its outputs. Agree on the expected benefit, quality requirements, permitted authority, and conditions that would lead to expansion, revision, or withdrawal.

For the billing case, an initial change might give employees a source-linked history and a proposed next action while retaining existing approval responsibilities. Evaluate whether it improves resolution and reduces repeated investigation after accounting for checking and correction. Follow cases long enough to see whether the problem returns.

Then test the design under three conditions. These tests examine the same process from different angles; they are not levels of maturity.

Condition	Question to test	Evidence to seek
Ordinary work	Can a verified billing error be resolved accurately with appropriate authority?	Correct adjustments, lower total effort, clear customer communication, and fewer repeat contacts.
An exception	What happens when the agreement is ambiguous or the payment system is unavailable?	Preserved context, a named owner, no duplicate action, and a workable escalation or recovery path.
A material change	Can a new subscription plan be introduced without recreating known errors?	Updated rules and guidance, representative evaluation, controlled release, and follow-up on subsequent bills.

Expand the scope or authority only when the results support it and the organization can sustain the capability. Some work may appropriately remain assisted or under human control. As evidence accumulates, revisit the original ambition: better knowledge, stronger ownership, or improved technology may make a more valuable redesign practical.