Net benefit: forecasting model cuts food waste and raises IT energy
FreshRoute, a grocery retailer, has deployed a demand-forecasting model across 900 stores. Pilot results show an 18% reduction in food waste in participating stores - roughly 3,100 tonnes a year at full rollout - but the model requires 40% more compute, adding an estimated 2.4 GWh annually, and store staff now use handheld devices about 30 minutes more per shift. The pilot ran in one region on a grid emitting 210 gCO2e/kWh; the full rollout includes regions at 520 gCO2e/kWh. Food waste emissions sit in the retailer's scope 3, while the additional electricity is scope 2. The CFO is enthusiastic about the waste saving; the sustainability manager is concerned the project will be reported as a win without checking the net effect, and that pilot performance will not hold at scale.
1. How should the net sustainability effect be assessed, and what boundary decisions must be made first?
Read the worked answer
Decide the boundary before computing anything. The assessable effects are: avoided food production and disposal upstream and downstream of the store (scope 3), the added compute and network electricity (scope 2, and scope 3 at the provider), and the additional handheld charging and replacement (scope 3). If the boundary is drawn narrowly - only the pilot region and only the compute - the answer is predetermined. Draw it to include the regions that will actually be served and the device fleet.
Use rollout conditions, not pilot conditions. The pilot's 210 gCO2e/kWh is less than half the intensity of some rollout regions. At 520 gCO2e/kWh, 2.4 GWh is roughly 1,250 tCO2e; at pilot intensity it is about 500 t. The waste saving is real, but the electricity penalty is two and a half times larger than the pilot implies.
Do not net across scopes in the reported figures. Report the scope 2 increase and the scope 3 reduction separately, then present the net effect as an analytical conclusion. Netting them into a single number inside the inventory misrepresents both and will not survive assurance.
Stress-test the 18%. Pilot stores were presumably selected, staff were observed, and the period may have been atypical. Model a conservative case - say half the pilot effect holds at scale - and check whether the conclusion survives. If the decision flips at 9%, the case depends on an assumption nobody has tested.
Price the uncertainty. A staged rollout with a control group costs less than a national deployment that turns out not to deliver.
2. The compute load is largely batch forecasting that runs overnight. What can be done to reduce the electricity impact, and what should not be claimed?
Read the worked answer
Exploit temporal flexibility. Overnight batch work has slack. Running it when grid carbon intensity is lowest in each region - rather than at a fixed time - reduces emissions without reducing work. The saving depends on how much the local grid varies, so measure it in a couple of regions before claiming it nationally.
Exploit spatial flexibility where latency allows. Central forecasting can be placed in lower-carbon regions because store staff do not wait on it. Store-level inference is different: it is latency-bound, so its placement follows the store, not the grid.
Reduce the work itself. The 40% compute increase should be interrogated: retraining frequency, model size, redundant feature computation, and duplicate pipelines all inflate load. A model that is 40% more compute-intensive by design may be 20% more intensive with better engineering.
Address the device side. Thirty extra minutes of handheld use per shift implies more charging cycles and earlier battery degradation - an embodied-carbon and cost effect that the compute analysis misses. Check the charging infrastructure and the replacement cycle.
Be careful what is claimed. Carbon-aware scheduling reduces the emissions attributable to the workload; it does not make the workload free of impact. Claiming hourly matching based on annual certificates, or describing the model as 'carbon neutral', would overstate the position. State the method, the region, and the hours covered.
Keep the systems boundary honest. If the scheduling tool itself consumes resources and requires new infrastructure, that belongs in the assessment.
3. What conditions should the approval decision carry, and how should the project be reported to avoid greenwashing?
Read the worked answer
Approve conditionally, not outright. The case is plausibly net positive, but the evidence base is a single-region pilot and the electricity penalty is understated. Conditions:
- A staged rollout with a matched control group, so the waste effect is measured rather than assumed at scale.
- A stop rule: if the realised waste reduction falls below an agreed threshold, or the compute penalty exceeds a stated ceiling, expansion pauses pending review.
- A named owner for the net-benefit claim, accountable for the quarterly figure.
- A sustainability assessment recorded in the business case alongside the financial case, so the trade-off is visible to whoever approves the next phase.
Report the two effects separately and in the correct scopes. The scope 2 increase from compute and devices goes in the inventory as an increase. The avoided waste emissions are reported as the reduction they are, with the calculation method and the boundary stated. If a headline net figure is used in communications, it should be presented as an analytical estimate, not as an inventory entry.
Constrain the marketing. '18% less food waste in pilot stores' is supportable once verified. 'Carbon neutral forecasting' or 'climate-positive AI' is not, and would expose the retailer to a credibility problem far more expensive than the project's savings.
Set a review gate at twelve months covering waste tonnage, compute consumption per store, grid intensity by region and device replacement rates - and a commitment to publish the result whether it is favourable or not.