August 2026
Governance Is a First-Class Requirement in Supply Chain Planning
Written by
The first question about an AI planning system should not be:
What can the AI do?
It should be:
What must be true before the organization acts?
A forecast can be accurate and still support an impossible plan. A recommendation can be explainable and still exceed the system’s authority. An automated action can follow its algorithm exactly while relying on unaccepted data, ignoring a binding constraint, or violating an organizational commitment.
These are not failures of intelligence.
They are failures of governance.
Governance determines which evidence is accepted, which calculations are authoritative, what the system may recommend or execute, who can approve a trade-off, what becomes the official plan, what risk has been accepted, and when changing conditions require a new decision.
Without that foundation, an AI planning system can produce outputs.
It cannot produce decisions an organization can safely stand behind.
A recommendation is not a plan
Supply chain software often collapses a forecast, a recommendation, a scenario, and a plan into one screen. They are not the same. A forecast estimates what may happen. A recommendation proposes a response. A scenario shows what could happen under a set of assumptions. A plan records what the organization has decided to do.
That distinction matters because a supply plan changes the real world. It can reserve inventory, prioritize one location over another, commit budget, change production, alter a customer promise, or expose someone else to risk.
A recommendation becomes a plan only when the organization can establish:
which source data and assumptions were accepted;
which rules, constraints, and calculation versions were applied;
whether the proposed action is feasible;
whether the action is within the system’s or user’s authority;
which alternatives and consequences were considered;
who approved the decision;
what residual risk was accepted;
what was released as the plan of record;
and whether the decision still holds as conditions change.
That chain is governance.
It is not administrative paperwork surrounding the planning system. It is part of the planning system itself.
The industry has been solving the wrong first problem
The o9 Solutions white paper Building Trust in Touchless Planning is useful because it makes a widely held industry logic explicit.
Begin with automation. Improve the forecast. Explain the output. Build planner trust. Reduce human intervention.
The paper defines touchless planning as planners guiding the system to decide on their behalf. Its target model reduces active planner involvement from nearly 100% of decisions to less than 25%.
Its six foundations are high-fidelity data, the right granularity, an advanced AI engine, accurate results, explainable forecasts, and feedback mechanisms.
These are sensible capabilities. The paper is also right that planners should not review every product-location-time combination manually or make overrides that worsen the forecast. Their attention should move to data, exceptions, scenarios, and information the system does not possess.
But the sequence has the order backwards.
The paper begins by deciding that the system should touch more decisions and the planner should touch fewer. Trust, transparency, guardrails, audits, and process standardization are then used to make that transition acceptable.
That treats governance-adjacent controls as the price of achieving automation.
A governance-first system asks a different question:
Which decisions may be delegated, under what evidence, rules, authority, and accountability?
Automation is then a result of the answer. It is not the starting objective.
This is a critique of the white paper’s argument, not a claim about every capability in the broader o9 platform.
Trust is not authority
Trust is psychological. Authority is institutional. A planner may trust a model because it has performed well historically, understand why its forecast increased, and see a high confidence score. None of that establishes whether the resulting action followed policy, relied on usable supply, or remained within an approved tolerance.
Trust can influence whether someone feels comfortable acting.
Governance determines whether the action is legitimate.
This matters most when the person held accountable did not control the decision. The o9 paper identifies the problem directly: planners may remain accountable for forecasting outcomes while losing control over the process.
But accountability cannot be repaired by asking the planner to trust the system more. It must be matched with evidence, authority, review rights, exception rules, the ability to hold or escalate a decision, and a durable record.
Where the system has authority, that authority must be explicit and bounded.
Where a person remains accountable, that person must have a real role in the decision.
Accuracy is not permission to act
Consider a hospital managing a critical medication shortage.
The system may correctly predict that available inventory will not cover expected treatments, then recommend a transfer, substitution, expedite, or reservation of stock for particular patients.
Before the hospital acts, it still needs to know:
whether the inventory is usable, quarantined, expired, or already reserved;
whether the substitute is clinically acceptable;
whether a transfer would expose another facility;
whether an expected delivery is actually confirmed;
whether the lead time makes replenishment too late;
whether the action follows clinical and allocation policy;
who has authority to approve the trade-off;
and what patient risk remains after the decision.
None of those questions is answered by forecast accuracy.
The model can be accurate and the action can still be infeasible, unauthorized, or unsafe.
Accuracy is evidence about a prediction.
Yet, the decision built from it can still be wrong.
Explainability is not accountability
Explainability is necessary, but it stops too early.
An organization may need to answer a much broader set of questions:
Why did the model produce this forecast? What signals, drivers, or patterns explain the prediction?
What evidence was accepted? Which data, assumptions, confirmations, and source records were considered reliable enough to use?
How did that evidence become this planning result? What calculations, rules, policies, and model versions were applied?
Is the proposed plan actually feasible? Were inventory, lead times, capacity, materials, budget, commitments, and other constraints respected?
What alternatives were considered, and why was this one chosen? What trade-offs were made across service, cost, risk, expiry, capacity, or other objectives?
What uncertainty or residual risk remains? What could still go wrong, and which assumptions matter most?
Who had authority to approve and release the decision? Was it automated within approved limits, accepted by a planner, or escalated for additional approval?
Does the decision still hold? What has changed since approval, and has anything changed enough to require intervention or replanning?
Can we reproduce the decision later? Can we reconstruct what the organization knew, what the system calculated, what people considered, and why the final action was taken?
Explaining why a model produced a number answers only the first of these questions.
Governed planning has to answer all of them.
That is the difference between model explainability and decision accountability.
Most model explainability addresses the first question.
A system may explain that demand increased because of a promotion, price change, seasonal pattern, or external signal. That still does not explain why the organization chose to expedite rather than transfer inventory, change a commitment, substitute a product, or accept a service risk.
Model explainability describes an output. Decision accountability explains how that output became action. It is the difference between inspecting a model and defending a plan.
Governance must exist throughout the planning cycle
Governance cannot be added only at the approval screen. By then, consequential choices have already been made. It must govern the full planning cycle:
1. Accepted evidence. Which source data, mappings, assumptions, and external signals are fit to enter the plan?
2. Reproducible calculation. Which deterministic rules and calculation versions establish the planning quantities?
3. AI-assisted interpretation. What may AI investigate, infer, explain, or recommend—and what may it not invent or change?
4. Decision authority. Which actions may be automated, which require review, and which require escalation?
5. Approval and release. Who can turn a recommendation into an organizational commitment?
6. The Planning Record. What did the organization know, calculate, consider, approve, and change?
7. Monitoring. Does the released decision still hold when supply, demand, constraints, or commitments change?
These are not controls around the real work.
They are the architecture of the real work.
A “human in the loop” does not provide this architecture by itself. The phrase says that a person appears somewhere in the process. It does not say which person, with what evidence, exercising what authority, against which policy, or leaving what record.
Human presence is not the same thing as governed human judgment.
Automation is delegated authority
“Percent touchless” can be a useful efficiency measure. It is not a sufficient measure of planning maturity.
A high touchless percentage says nothing by itself about whether decisions used accepted evidence, respected constraints, stayed within delegated authority, or were monitored after release.
The goal should not be maximum automation. It should be appropriate delegation.
Routine and reversible decisions may be fully automated. Bounded decisions may be executed within approved tolerances. Material exceptions may require planner review. Decisions with clinical, regulatory, financial, or public consequences may require cross-functional approval.
A 95% touchless rate may be responsible in one portfolio and reckless in another. Touchless planning is not a maturity level every decision should reach. It is a property that some governed classes of decisions can safely earn.
Governance makes responsible speed possible
Governance is sometimes treated as friction. Weak governance does create it: when evidence is uncertain, calculations cannot be reproduced, authority is ambiguous, and decisions leave no record, planners compensate by checking everything themselves.
Strong governance has the opposite effect.
When routine decisions operate within accepted evidence, defined rules, approved tolerances, and explicit authority, planners do not need to reopen every calculation simply because they distrust the process.
Their attention can move to the exceptions where judgment matters.
Governance is not anti-automation. It is what makes appropriate automation defensible.
Vista’s model is:
Accepted evidence → deterministic planning → AI-assisted interpretation → authorized decision → release → monitoring
Deterministic engines own reproducible calculations. AI retrieves evidence, investigates changes, compares scenarios, and explains results. Humans retain authority where judgment and accountability require it. The Planning Record preserves the chain from source evidence to released decision.
We do not add governance so that we can pursue touchless planning.
We govern the planning system so well that appropriate decisions can safely become touchless.
AI does not earn the right to act merely by being accurate, explainable, or trusted.
It earns that right by operating inside a governed planning system that can prove the evidence, logic, authority, and accountability behind every consequential decision.
Vista was architected with governance as a first-class requirement—not as a control layer added after a recommendation is produced. That governance trail is carried in a first-class object: the Planning Record, which connects accepted evidence, assumptions, reproducible calculations, AI contributions, authority, approvals, released actions, and monitoring across every planning cycle. If your organization is deciding how much planning authority to give AI, we would be glad to show you how this works in practice.
You can read the full o9 Solutions paper here.