Articles

How to Decide What to Redesign with AI

Follow a process redesign using the SCOPE framework, including evidence to justify proceeding or holding.

November 4, 2026
15 minutes
‌

Article 2 - System-outcome redesign for the AI era by Jafar Shayan

A transformation team can explain its technology, demonstrate a working assistant, and still struggle to answer a more consequential question: which business change should be authorized next?

Consider Meridian Components, the fictional industrial supplier introduced in the first article. Its entry assistant exposed a conflict between release capacity and inventory reservation. The team now has to choose which operating rules to change and whether more AI is worth adding.

The next decision is not obvious. Meridian could control order release, change reservation rules, clarify exception authority, add an AI assistant to exception handling, or redesign how customer promises are made. Those options require different investments and expose different assumptions.

SCOPE is my framework for making that decision. It uses five recurring stages: Set, Characterize, Originate, Prove, and Embed. Meridian is a teaching example throughout, including its observations, tests, and decisions.

Keep a traceable argument from the intended outcome to the chosen change, including who can authorize and fund it. If evidence changes the argument, change the decision.

A lifecycle organized around decisions

SCOPE asks teams to keep four requirements visible. Local gains must meet agreed system-value thresholds without breaching hard guardrails. AI must face a credible non-AI comparator. Each option must anticipate where work, risk, or a constraint may move. Feedback, correction, and knowledge maintenance must have owners and capacity.

The five stages organize the work; the four requirements govern the quality of the decision. Supporting methods help answer particular questions and are selected as needed. The stages can repeat when evidence changes an assumption.

These requirements have substantial prior art. Business process management (BPM) supplies an established lifecycle of process analysis, redesign, implementation, and monitoring. Lean and Theory of Constraints focus attention on value and flow. Service design supports reframing and alternatives; sociotechnical design examines organization and technology together. AI risk management adds contextual evaluation and ongoing governance. [1]

SCOPE brings these requirements into one usable decision process. It draws on established methods; the integrated framework awaits field validation.

‌

Figure 1. The stages organize the work; they do not require a one-way project sequence. A failed test can reopen an option, a diagnosis, or the original outcome.

The central record contains the outcome, causal hypothesis, alternatives, evidence, and current decision. It can begin as a short document. Supporting maps and tests are added when they resolve a consequential uncertainty.

Set: agree on the result and the terms of the decision

Meridian’s initial objective was faster order entry. The redesigned objective is that customers receive complete, correct orders on the date originally promised, within agreed economic, service, and risk limits.

That statement changes both the boundary and the authority required. The measurement begins with the customer’s request and continues through delivery and recovery. It includes work waiting before entry, unresolved reservations, exception decisions, and interactions with fulfillment. Restricting the boundary to the assistant would hide the mechanism under investigation.

The team chooses delivery against the original promise as its primary outcome measure. It also tracks requested lead time, promise-date changes, customer contacts, exception workload, and cost. Recording the original promise prevents an apparent improvement achieved by quietly moving delivery dates. Segmenting the measures helps reveal whether one customer group absorbs the delay.

The criteria distinguish three judgments. The primary measure states the intended improvement. Countermetrics reveal its costs and distribution. Hard guardrails identify actions the team cannot authorize, such as decisions outside delegated credit authority.

A modest cost increase may be acceptable for a sufficient improvement in service. A prohibited credit decision is not made acceptable by a good average delivery result. Any temporary burden has a time limit, an owner, and a recovery plan.

Set also identifies the benefit owner. If the proposal is supposed to reduce overtime, who can actually change staffing or scheduling? If it promises better service, who commits the capacity? An expected benefit without an operating decision is an incomplete hypothesis.

For Meridian, the outcome owner needs a mandate spanning release, reservation, and exception priorities. Finance retains its credit-policy authority. A sponsor must resolve trade-offs that exceed either role’s delegation.

The Set decision is whether this outcome, boundary, and set of criteria justify further work. Its minimum output is an agreed problem statement and decision contract. A team that cannot agree may need to narrow the proposal or resolve the conflict before spending on design.

Characterize: establish what is actually causing the problem

A process diagram is a starting point. Meridian must connect the reported delivery misses to the reservation mechanism.

The team follows affected orders from customer request to dispatch. For each delayed shipment, it checks whether stock existed, whether unresolved orders held it, whether releasing those reservations would have enabled shipment, and whether another constraint would still have prevented delivery. It compares this with ready orders that shipped and similar orders before the entry assistant was introduced.

This evidence can disprove the preferred explanation. A supplier shortage might account for the missed deliveries. The customer mix might have changed. The reservation rule might have little practical effect because most exceptions clear before fulfillment. In those circumstances, changing the rule would not address the main problem.

The queue figures provide a useful check. In the baseline, 100 daily releases at a 28% exception rate produce 28 exceptions against nominal resolution capacity of 30. At 140 releases, arrivals rise to 39.2 while resolution capacity stays unchanged. The accumulating queue is consistent with overload. The link to delivery still depends on the reservation evidence.

Observation should include the people doing the work. A planner may explain that ostensibly identical exceptions differ greatly in difficulty. A credit specialist may identify why an undocumented check remains necessary. A warehouse colleague may show that stock displayed as available cannot be picked. These accounts should be checked against operating records where possible, rather than accepted or dismissed because of role or seniority.

Where handovers or ownership boundaries appear to prolong a case, compare those boundaries with the information and authority the work needs. At Meridian, a shared exception record will help only if finance and operations can resolve the decisions it exposes.

The team then writes a causal hypothesis: increased release creates more unresolved reservations, which block ready orders; fragmented evidence and unclear authority prolong those reservations. It records competing explanations and its confidence.

Constraint analysis is useful here, but SCOPE should not force every opportunity into a throughput story. A different proposal might target avoidable cost, accessibility, or recovery from disruption. In that case, Characterize should establish the corresponding mechanism and dependencies.

The Characterize decision is whether the evidence supports a specific point to redesign and whether the boundary remains adequate. Meridian may need to include inventory allocation and sales promises even though the project began in order entry. It can also stop if the suspected problem is not supported.

Originate: compare designs that could change the result

Develop alternative design hypotheses, compare their likely value and consequences, and choose which options warrant testing in Prove and in what order. Eliminate options that breach guardrails or fail essential requirements before investing in a test. This comparison can happen on paper; parallel pilots are not required. At Meridian, separate the operating rules that need changing from the technology carrying them out. Replacing an AI component would not necessarily fix a reservation policy. Adding another assistant might clarify a case while leaving nobody authorized to decide it.

Established redesign heuristics include eliminating work, simplifying it, combining activities, and changing their sequence. [2]  Here, the team uses those ideas to construct alternatives rather than merely compare AI vendors.

  1. Current release design
    • Operating change: Continue the existing release and reservation rules.
    • What it must establish: Whether the suspected harm is real. If confirmed, this is not an acceptable scale design.
  2. Conventional redesign
    • Operating change: Control release, prevent unresolved orders from locking scarce stock without approved priority, share exception evidence, and clarify decision rights. No new AI in exception handling.
    • What it must establish: Whether these operating changes are sufficient at acceptable cost.
  3. AI-assisted exceptions
    • Operating change: Keep the conventional option’s rules and authority; use AI to assemble source-linked evidence for exception handlers. Conventional rules handle authorized routine cases.
    • What it must establish: Whether AI adds useful value beyond the conventional option after checking, correction, and maintenance costs.
  4. Promise redesign
    • Operating change: Establish feasible options before acceptance, using reliable capacity and substitution data, conventional optimization, and bounded AI assistance.
    • What it must establish: Whether data, integration, authority, and customer choices support the larger change.

For this comparison, both exception designs retain the already deployed entry assistant at a controlled release rate. This holds the background condition constant while the team tests additional AI in exception handling. The conventional workflow also remains usable with manual entry if fallback is needed.

The conventional option must be credible. It receives the same attention to workflow, training, and authority as the AI-assisted option. Comparing a well-supported AI design with an intentionally weak manual process would tell the team very little.

Each option also needs a prediction about displaced work. The conventional option might move effort into maintaining shared information. The AI-assisted option might reduce evidence gathering while adding verification and correction. The promise redesign might prevent downstream exceptions but increase negotiation before order acceptance. The team names a leading signal for each and an alternative path under changed demand or incomplete data.

Use these hypotheses to identify what to monitor as the design changes. Meridian’s larger flow may eventually be limited by production changeovers or carrier capacity. If the proposed design does not materially change arrivals, the team can record that expectation and test it rather than inventing migration.

The Originate decision records the test sequence and the evidence that would make each option preferable. Keeping a conventional change and declining additional AI remains a valid option.

Prove: match the evidence to the decision

Prove establishes sufficient evidence for the next bounded decision. The strength of that evidence must match the consequences and uncertainty of the proposed action.

Meridian separates three questions. Does the conventional redesign repair the release and reservation mechanism? Does the AI-assisted option add value beyond the conventional option? Is the promise redesign ready for a larger change in promise creation? One before-and-after pilot cannot answer all three cleanly.

The team can examine historical order traces to estimate which reservations the conventional design would have changed. Replay is useful for identifying affected cases; it cannot show exactly how staff or customers would have behaved under the alternative. A queue model can explore demand variation and capacity limits, but its results remain conditional on its assumptions.

A bounded operational test of the conventional option then observes whether fewer ready orders are blocked and whether delivery improves without moving unacceptable cost or delay elsewhere. Outcome measurement follows the original customer-request cohort, not only cases conveniently completed during the test.

To assess the AI-assisted option, the comparison keeps the conventional option’s rules, authority, and staffing assumptions in place. The added element is AI evidence assembly. The team measures total handling effort, including verification, correction, escalation, and maintenance. It also checks whether a polished summary causes handlers to miss contradictory source information.

The comparison unit must preserve the mechanism being tested. Two independent stock pools can each contain ready and unresolved orders competing for inventory. A comparison between pools can therefore retain the reservation mechanism while limiting spillovers between groups. Product-family labels alone do not establish independence; shared stock and handlers must be checked. A staged change across suitable pools or a comparison over time can help, with differences in demand, supply, and timing examined explicitly.

In Meridian, the additional-AI comparison follows establishment of the conventional design so the team can isolate AI’s contribution. This is a case-specific sequence, not a required order for every SCOPE application. Another team can test its selected design directly against the current method or a credible alternative, with evidence proportionate to the decision.

‌

Figure 2. Meridian’s evidence path. Evidence for the operating change does not automatically justify additional AI. Each decision has its own comparison and approval boundary.

The promise redesign faces a separate readiness test. Shadow-mode promise generation exposes incomplete capacity and substitution data. Planners must correct too many proposed options. That result supports holding the promise redesign. Shadow mode does not establish how customers would respond to a changed promise process, so readiness would still be only one part of its evidence.

In this walkthrough, the conventional change has sufficient supporting evidence for the tested scope. The additional-AI comparison remains inconclusive: its uncertainty spans effort reductions that would and would not justify the additional cost, and the service evidence cannot yet clear the required segment limits. Meridian therefore adopts the conventional redesign and holds expansion of AI-assisted exceptions. Promise redesign also remains on Hold because of its separate data-readiness failure.

These are stipulated teaching findings. Their purpose is to demonstrate that different parts of a proposal can receive different decisions. The practitioner guide shows the contract, the feasibility check, and what evidence would reopen the held option.

The Prove decision is Stop, Adapt, Hold, or Scale within an explicit boundary. A small test supports only the expansion its evidence can justify. Broader rollout may need another learning step.

Embed: make the change operable after the project

Embedding the conventional redesign means making its release rules, reservation controls, and shared evidence usable in normal work. Holding additional AI does not remove the need for ownership, maintained knowledge, or recovery.

Meridian assigns ownership of release limits, stock-reservation policy, exception priorities, source data, and the existing entry assistant. It gives handlers time for source correction and knowledge maintenance. It records the fallback to conventional evidence gathering and tests whether staff can use it at the expected load.

The benefit owner reviews whether the saved effort is actually realized. If the plan is to reduce overtime, the operating schedule must reflect that choice. If capacity is retained for resilience, the budget and staffing agreement must protect it.

Controls need owners too. A handler must know how to challenge a source, escalate an ambiguous case, and pause an unsafe output path. A threshold change should have a reason and an approver. A change to a model, retrieval source, supplier service, or price can invalidate the tested economics or behavior. Record the version and dependencies used in the test, identify the changes that require reassessment, and maintain an operable fallback. Recheck the affected assumptions before expanding exposure after a material change.

Meridian reviews whether the operating change continues to improve reliable delivery within its agreed trade-offs. The practitioner guide provides the full monitoring record.

The team also looks beyond this flow. If another business unit wants the same reviewers or knowledge owners, the outcome owner needs a route to resolve shared-capacity conflicts. Local success does not grant a right to consume resources already committed elsewhere.

The Embed decision is whether the organization can operate, fund, monitor, and change the design. If the answer depends on a project team continuing unpaid aftercare, the operating design is unfinished.

Keep the method proportionate

For a low-consequence change, the five decisions can fit in a short record and a normal operating review. A cross-functional redesign with uncertain data, significant employee effects, or difficult recovery deserves deeper investigation. The stage names remain the same; the amount of work changes.

Do not require every map, workshop, or analytical method. Ask what uncertainty the next artifact resolves and whether it could alter the decision. If the answer is unclear, it may be unnecessary.

Evaluate SCOPE through the decisions it changes and the effort it requires. In an initial application, record the decisions it changed, assumptions it exposed, options it stopped, staff effort it consumed, and time it added. Follow the eventual outcomes and compare with existing practice where feasible. Record where the existing method would have reached the same decision with less effort. Retrospective applications can expose usability problems, but hindsight limits what they establish. Prospective applications should record predictions before outcomes and include unsuccessful cases.

For Meridian, the useful result is a specific choice: correct the operating rules, hold additional AI pending sufficient evidence, keep the data-dependent promise redesign on Hold, and assign authority for the next review. Another context may justify a simple tool or no change.

The next article examines the authority and funding behind those choices. The implementation guide turns the same logic into a worked record and a method-selection path. Before either, a team can begin with one page: the decision it faces, the mechanism it believes, and the evidence that would change its mind.

Decide where AI can make a difference

If you’re deciding what to redesign or where to invest in AI, at Xebia we can help you define the business outcome, compare AI and conventional approaches, and identify the evidence needed to choose your next step.

References

[1]  Established sources include Dumas et al., Fundamentals of Business Process Management; the Design Council’s Double Diamond; and NIST’s AI RMF 1.0. These methods supply established foundations; they do not establish SCOPE’s novelty or superiority.

[2]  Hajo A. Reijers and Selma Limam Mansar, “Best Practices in Business Process Redesign: An Overview and Qualitative Evaluation of Successful Redesign Heuristics,” Omega 33, no. 4 (2005): 283-306.

Written by

Contact

Let’s discuss how we can support your journey.

‌
‌
‌
‌
‌
‌
‌
‌
‌