Skip to content

Instacart Expert Feedback - 2025-07-01

TL;DR

Strong conceptual validation but critical UX feedback: settings page is "too academic" as standalone - should be accessed as drill-down from forecast numbers when investigating anomalies. Also wants assumption tracking for variance attribution.

Overall Sentiment
Direction validated Partially
Stage shown Early prototype
Ready for next stage Needs UX rethink
Blocker for them? No (expert interview)

Source

  • Transcript: 2025-07-01-instacart-forecasting-discovery.md
  • Date: 2025-07-01
  • Participants: David Watt (Treasury, Instacart - formerly Sonder)
  • Context: Expert discovery interview - David left Sonder, now at Instacart (uses Kyriba)

This Iteration: Validation

What they validated about the current design:

  • Per-category control concept - "I just want to manually upload this because I get these numbers from my colleague over there and they're more reliable anyway than the machine learning"
  • Turn off ML for specific categories - Validated use case: exclude big client from ML, get manual input from relationship team
  • Mixed approach (ML + manual) - "You could have a mix... that's one option"

Important Nuances

  • The settings PAGE itself is not the right entry point - it feels "academic" without context
  • Settings should be a DRILL-DOWN from forecast numbers when you spot something odd
  • "I don't know that I'd come in here first without seeing the numbers"

Future Iterations: Suggestions

High Value (UX Direction Change)

  • Settings as drill-down, not destination - "If I double clicked and then it popped up something like this of like, well, there's two rules generating this balance... this would be helpful as like that second layer"
  • Show settings in context of numbers - Only show rules that are MATERIAL to the forecast, not all possible rules
  • Re-categorize from account page - Fix category errors while investigating variance, not navigate away

Medium Value

  • Assumption/version tracking - Save "July 1st forecast" version, track assumption changes over time
  • Variance attribution - System should show: variance = actuals variance + assumption changes
  • Actuals overlay - When changing settings/assumptions, show recent actuals for sanity check

Longer Term

  • Forecast composition view - Stacked bar showing ML vs manual vs ERP contribution per category
  • Big client exception workflow - Separate category with ML off, prompted for manual input

Questions Asked

  • How would you prefer to access settings? → "Start with numbers, drill in when something looks odd"
  • Would configuring from account page work? → Yes, but after seeing the numbers first

Raw Feedback Quotes

"I don't know that I'd come in here first. Right without seeing the numbers. It's very academic, just to kind of scroll through and see all the rules. Right? Like well, okay, but if there's 10 rules but they're all times zero and it has no impact on the forecast. I don't... that rule is going to result in zero." - David Watt

"If I double clicked and then it popped up something like this of like, well, there's two rules that are generating this balance and one is X and the other is Y. This would be helpful as like that second layer." - David Watt

"How would I know which one of these rules is actually material to the forecast? I'm going to focus in on those five big items right to start with at least." - David Watt

"I find a number that looks odd to me. And then I, you know, that's where I want to double click in. How did they arrive at this number?" - David Watt

Key Use Case: Big Client Exception

David shared a concrete use case from Instacart: - One client pays $20M/month in unpredictable spikes - Baseline collections for everyone else: $200k/day - Need: Separate category for this client, ML turned OFF, manual input from relationship team - Quote: "You can't assume it's gonna be a million dollars a day, it's gonna be 200k a day. And then you have someone else tell you like, okay, they're gonna pay us on the 21st"

This validates the need for: 1. Per-category ML control 2. Integration with manual/external inputs 3. Separate visibility in variance analysis

Action Items

  • [ ] Redesign settings as drill-down from forecast view, not standalone page
  • [ ] Show only material rules in drill-down context
  • [ ] Consider assumption tracking for variance attribution