Did you know that the cruel thing about ETAs is that guests remember the misses? Predictive windows only build loyalty when they are calibrated—honest about kitchen load, distance, and the messy middle between packed and picked up. This article is about precision as trust: fewer heroic best-case quotes, more promises the pass can keep without turning dispatch into a lie machine.
Predictive ETAs are a trust surface: the guest bets their evening on a number you showed at checkout. Precision is not about boasting the shortest minute count—it is about calibration: when you say nineteen minutes, the food is hot and the doorbell rings close to that promise often enough that people come back.
Forecasts should combine make-time models (queue depth, item mix, historical station load) with dispatch reality (courier assignment, traffic priors, batching rules). The output must update when the kitchen slips—static promises are how brands accumulate one-star “waited an hour” stories. All of it hangs off the same order object we describe in checkout architecture.
ETAs should align with how you drew delivery zones—promising Brooklyn timing with a Manhattan kitchen model guarantees drift.
Under-promise beats over-promise
A small buffer that preserves on-time rate beats a heroic number that fails weekly. Communicate ranges or live updates after fire if your data supports it; guests forgive movement they can see more than silence.
Separate pickup and delivery forecasts—courier variance dominates the latter. If you reuse pickup math for delivery, you will systematically lie in the direction guests care about most.
Transparency without drowning the guest
Show why briefly when delays happen: high demand, driver en route, kitchen remaking a ticket. Ops scripts should match what the app displays so support is not improvising. Trust is consistency between channels.
Avoid notification fatigue: one meaningful update beats three pings that say “still working on it.”
Measure calibration, not vanity
Score MAE or quantile error on ready and delivered times, segment by zone and daypart, and correlate with refunds and CSAT. Improve the model where error clusters—not only where averages look fine.
Review worst decile errors weekly; tail risk drives reviews, not mean error.
Human overrides with discipline
Managers should be able to extend ETAs globally when a station goes down—but those overrides should log reasons for postmortems and training.
ETAs on the Menuella spine
Menuella keeps ordering, kitchen signals, and guest messaging coherent in the Menuella ecosystem and delivery flows—so the ETA at pay is the same one the pass and courier see. Precision is operational discipline expressed as a timestamp.
When dispatch and kitchen share telemetry, guests stop receiving physics-defying promises.
