“We need twelve percent growth.”
- What it means
- Commercial target
- What RabbitHawk checks
- Keep it as a goal, not expected demand
RabbitHawk Forecast Interventions
The forecast is saved. Then someone knows more.
A store manager notices an access problem. A buyer hears a supplier warning. A planner sees a pattern the baseline did not represent. RabbitHawk lets them speak or type what they know, asks only the questions that matter and produces a probabilistic candidate for human review without overwriting the plan.
Modelled planning walkthrough
See what happens when roadworks threaten store demand
RabbitHawk captures the notice, asks for missing facts, tests whether the forecast and a still-open decision should change, and keeps the saved forecast as the reference until a planner approves an update. Need the connected action plan? See Forecasting & Optimization.
The language layer structures evidence and questions. Registered numerical models calculate the uncertain candidate. Named people approve forecast publication and any operational action separately.
Why this is hard today
The first guess becomes the forecast. The reason gets lost.
A store, region and planner may each know something useful. Asking them for a new number can discard the source, mechanism, assumptions and uncertainty that made the evidence useful in the first place.
- 01
The first guess becomes the forecast
A useful frontline observation is compressed into a percentage before its assumptions, mechanism and uncertainty are understood.
- 02
One story is counted several times
Buying, stores and digital may all repeat the same source. Apparent corroboration can be an echo unless dependence is traced.
- 03
Useful context never reached the model
Information can be known before the forecast and still be omitted, stale, unstructured, misencoded or filtered from the baseline.
- 04
The forecast changed. The plan did not.
Pack sizes, timing, stock and other limits can leave the same order or transfer in place. Check before triggering another replan.
Not every message means more or less demand
Change the right thing
Expected demand, observed sales, targets, open-to-buy, commitments, service policies and constraints are different objects. RabbitHawk decides which object the evidence can change before any forecast or plan is recalculated.
“Our open-to-buy is tight.”
- What it means
- Buying budget / OTB
- What RabbitHawk checks
- Constrain feasible buys, not customer demand
“Sales fell while shelves were empty.”
- What it means
- Censored sales observation
- What RabbitHawk checks
- Estimate what availability hid
“Production can make only 5,000.”
- What it means
- Supply constraint
- What RabbitHawk checks
- Change the feasible plan, not demand
“The promotion has changed.”
- What it means
- Commitment or demand driver
- What RabbitHawk checks
- Check approval, timing and baseline representation
“We have too much inventory.”
- What it means
- Inventory state
- What RabbitHawk checks
- Test transfer, markdown or buy options
How it works
Ask people what changed, not what number to type
Contribute evidence in seconds
Speak, type or attach what you noticed. RabbitHawk keeps the original contribution, its source, timing and affected business scope.
- Private-first capture before group influence
- First-hand, relayed and shared-source evidence kept distinct
- Facts, assumptions, unknowns and corrections preserved
Turn intuition into testable evidence
Reflect what was heard, then ask neutral follow-up questions only while the answer could change scope, uncertainty or a still-open decision.
- Check what the baseline already represented
- Expose mechanism, duration, mitigation and counterevidence
- Allow ‘unknown’, no effect and abstention
Produce a governed probabilistic candidate
Language AI structures the case. A registered numerical model calculates a bounded distribution for the affected scope and reconciles it with the planning hierarchy.
- Reference and candidate stay side by side
- No change and insufficient evidence remain valid
- People approve forecast publication and action separately
Learn from the whole case
Keep every forecast, evidence set, decision, action and suitable actual. Match similar past cases before opening their outcomes, then propose reviewed improvements.
- Forecast, decision, execution and commercial results scored separately
- Private calibration feedback across repeated comparable cases
- No person leaderboard or one-event source score
- No silent retraining or automatic policy change
Keep the record
Learn from the change, not by blaming the person
Keep the saved reference, original evidence, guidance shown, probabilistic candidate, approval, action and suitable actual. Compare forecast versions on the same eligible cases, then measure decision, operating and commercial outcomes separately.
- 01
Reference
Exact saved distribution, hierarchy and information manifest
- 02
Evidence
Original contribution, questions, assumptions and representation check
- 03
Candidate
Versioned probabilistic alternative from a registered numerical model
- 04
Approved
Named human decision on forecast publication
- 05
Action
Separate operational decision, exposure and execution record
- 06
Actual
Suitable outcome, version scoring and governed learning proposal
Evaluate the intervention, not the person
Targets, incentives and authority are recorded as decision context, not evidence of motive. RabbitHawk looks for repeatable process patterns across comparable cases. A miss on one case does not prove bias, gaming or poor judgment.
Learning proposals
- Forecast calibration
- Evidence usefulness
- Question value and burden
- Source dependence
- Timing and duration
- Decision and operating outcome
Evaluate before you expand
Keep the planning system you already use
- 01
Write the intervention and evaluation contract
Choose one recurring decision, a representative product-location-cycle subset, named owners, suitable actuals, guardrails and stop rules.
- 02
Build the minimum data foundation
Allow at least four weeks to connect approved read-only sources, align versions and hierarchies, and validate as-of lineage, security and latency.
- 03
Run shadow, gate readiness, then evaluate
Keep official planning unchanged during shadow. Influence decisions only after readiness, through a powered concurrent test or a clearly labelled quasi-experiment.
A serious first evaluation
Start with a minimum four-week data foundation
Choose a representative subset of products, locations and forecast cycles. Connect and validate approved read-only data, then run in shadow while the official plan stays unchanged. Influence begins only after a pre-agreed readiness gate.
Scope one governed evaluation
Define one recurring decision, a representative subset, the minimum four-week data foundation, shadow protocol, readiness gate and controlled evaluation before data access.
Scope the evaluation