Responsible AI

Where AI shows up in the product

DemandLabs is a planning platform with machine learning in four places. It is worth being precise about them, because the obligations we take on differ by surface.

• Forecasting. Machine learning fused with causal drivers, so price, promotion, holiday, weather and macro effects are modelled explicitly rather than buried in a trend line.
• Scenario solving. Options are solved against the decision graph so revenue, margin, service and cash move together, and each option carries its own numbers.
• Copilot. A natural language surface over your own plan, used to ask why a number moved and to reach the screen that answers it.
• DemandLabs AI agents. Long-running agents that sense, explain, simulate and act: the Demand Sensor, Supply Rebalancer, Inventory Governor and S&OP Orchestrator, plus any custom agents you configure.

None of these is a general-purpose chatbot pointed at your company. Each is scoped to a planning job, and each operates on the same versioned model of your network that the rest of the platform uses.

Human in command

Planning decisions belong to planners. That is a design constraint in DemandLabs, not a preference. Agents watch signals, explain what changed, and price the options. A person decides what to do.

• Publishing a consensus plan is a human act. An agent can recommend a scenario and show what it closes, but approving and publishing is done by a named user.
• Alerts arrive staged, not executed. A disruption in the control tower comes with a costed response prepared and waiting for a decision.
• Agents escalate when judgment is required rather than guessing. The design principle is simple: act inside the guardrails, or hand the planner the pen.
• Every action an agent takes is logged, explained and reversible, with the same version lineage as a human edit.

We do not build features whose purpose is to remove a human from a consequential decision. Where autonomy exists, it exists because a customer switched it on inside a boundary they set.

Explainability is not optional

A forecast nobody can interrogate is useless at a consensus table, because the first challenge kills it. So every prediction ships with its reasoning attached.

• Driver attribution: the contribution of promotion calendar, price move, holiday shift, weather, macro index and residual, ranked, for the move you are looking at.
• Confidence bands drawn on the forecast, so the range is visible rather than implied by a single line.
• Forecast value add: every human touch measured against the untouched baseline, so an override that helps and an override that hurts are told apart with evidence.
• Plain-language root cause from the agents, rather than a score with no story behind it.
• Full lineage on the consensus grid: every touch logged, attributed to its owner, and scored next cycle.

The same transparency applies to us. When a model is retrained or a method changes materially, that change is recorded and available to the customer, not slipped in quietly.

Guardrails on autonomous agents

Autonomy in DemandLabs is a permission, granted deliberately and scoped narrowly. Customers define the boundary before an agent may act inside it, along axes such as:

• Which agents are enabled at all, and for which planning surfaces
• Which parts of the network, product hierarchy and geography they may touch
• Value and quantity thresholds above which an action becomes a recommendation instead
• Time windows where autonomous action is suspended, such as a close period or a peak season freeze
• Who is notified, who approves, and who is the named owner of each agent

Anything outside the boundary is staged rather than executed, and waits for a person. Guardrails are set and changed by authorised users, changes to them are audited, and an agent can be paused or switched off at any time. Because actions are versioned, an agent action can be rolled back like any other plan change.

Agents also inherit the permission model rather than sitting above it. A user cannot use Copilot or an agent to reach data their role does not entitle them to see.

Your data is not our training set

We do not use customer data to train models that are shared with other customers. Models are trained and backtested on the history of the customer they serve, inside that customer’s tenant and permission boundary.

Nothing a planner types into Copilot, and nothing an agent observes in your environment, becomes training material for anyone else’s forecast. If we ever want to use customer data for a broader purpose, we will ask that customer first, in writing, for that specific purpose, and a refusal will cost them nothing.

Where the platform uses a third-party model provider for language capabilities, we contract for the same restriction, so prompts and content are not retained for that provider’s own training.

Tested before go-live, monitored after

Accuracy is a claim, and claims should be checked against your data rather than ours. During deployment, models are trained and backtested on your own history, and performance is compared against your incumbent method before anything goes live. If the model does not beat what you already run, that is a finding, and you should hear it from us first.

After go-live, forecast accuracy and forecast value add are tracked continuously and visible in the product rather than assembled for a quarterly review. Degradation, drift and data quality problems upstream are surfaced as issues to fix, not absorbed silently by the model.

The illustrative figures shown on this website are examples of the shape of results, not a promise of what your numbers will do.

What we will not build

DemandLabs plans supply chains. Some uses are outside that purpose and we will not support them:

• Automated decisions about individual people, including hiring, promotion, discipline, credit or any other decision with a legal or similarly significant effect on a person
• Surveillance or productivity scoring of individual employees dressed up as planning analytics
• Any use that requires us to infer protected characteristics from customer data

The value-add scoring in the consensus workspace measures forecast touches, and is designed to improve the plan and retire overrides that do not help. It is not a performance management tool, and we ask customers not to use it as one.

Limitations, stated plainly

Forecasts are probabilistic. A confidence band is an honest statement that the future has a range, and the range is sometimes wrong. Models learn from patterns, so they are weakest exactly when the world stops repeating itself: a first-of-its-kind disruption, a structural market shift, a product with no comparable history.

Model output is decision support. It should be read alongside human judgment and the context a planner has that the data does not contain, and it should not be the sole basis for a decision with safety, legal, financial or employment consequences.

Data quality sets the ceiling. The platform scores incoming data quality for exactly this reason, and a forecast built on a broken feed will say so rather than pretend.

Raising a concern

If a model output looks wrong, unfair or unexplainable, we want to hear about it while it is still a small problem. Inside the product, planners can flag a forecast or an agent action for review and the escalation path routes to the named agent owner in your organisation.

From outside, or where the concern is about our practice rather than a single number, write to [email protected] and it reaches a person. We aim to acknowledge within a few working days, and to tell you what we found and what we changed, including when the answer is that we were wrong.

Nobody at a customer organisation should face any consequence from us for raising a concern about how our models behave.

Changes to this statement

This statement describes the platform as designed and as we intend to operate it. It will be revised as the product ships and as regulation such as the EU AI Act settles into practice. The last updated date at the top of this page reflects the current version, and material changes are notified to customers through their account contacts.