For AI labs Talk to us

The dashboard said next-day delivery was a hit. It was losing money.

It started with a question.
Nikhil Raoruns supply at Simcorp Homein Veloccy
Did the 857-product next-day pilot actually pay off, and should we expand it?
AI with access
Access
Experience
none
+277.4% sales, pilot products against the rest

With access to the data, the pilot is a clear win.

Growth’s board compares products in the pilot with products outside it. Anyone reading that board, a person or an AI, would say expand, and fill every slot the warehouse has.

AI with experience
Access
Experience
Pilot list frozen · 857 productsBengaluru city campaignand more
+15.5% real lift, against 30.3% to break even

With experience of the company, it is a loss.

Veloccy knew how the pilot had been picked and what growth was running over the same months. With both taken out, next-day delivery lifted sales by 15.5%. It needed 30.3% to cover what the warehouse costs.

How it happened in Veloccy.

A simulated home marketplace with a next-day warehouse in Bengaluru. We built it and planted the cause, so we know what the right answer is. These are a few moments from the work.

app.veloccy.com/simcorp-home/arcs/cat-1
Simcorp HomeArcsBLR-1 second-floor Diwali assortment
AK RM NR
CAT-1

BLR-1 second-floor Diwali assortment

Properties
Readout Learned incremental_contribution RMRyan Mercer opened 29 Sep review 9 Dec
Continues
SUP-1 · BLR-1 next-day pilot payoff
In the Brainwhat this bet read, and what it led toOpen in Brain
What it read
SUP-1Next-day pilot payoff Return rate Fragile products 9 tables
CAT-1 · BLR-1 second-floor Diwali assortmentArc · learned · 8 findings
What it led to
CAT-1Stock 678 products Second-floor Diwali test Gifts leads the next expansion GRO-1Gift badge discount
The result

The second floor paid about $23,919, inside the range Veloccy wrote down before anything shipped.

The next-day badge lifted add-to-cart by 26.9%. The payoff range still crosses zero, so it probably paid and that is not proven yet.

The full answer
Net contributionexpected against measured +$23,919inside the expected range
Expectedwritten 29 Sep
+$18,885−$33,791 to +$70,092
Measuredread 10 Dec
+$23,919−$47,745 to +$94,337

How it got there

8 checks
  1. 01 Found Diwali demand sits in gifts, lighting and decorBefore Diwali last year gifts sold 2.8 times their usual level, lighting 4.1 times and decor 1.7 times.
  2. 02 Found Forecasts from recent sales run hotReplayed, they came in at 46.1% of forecast, and 78.0% of the products they picked sold nothing.
  3. 03 Found Stock 678 products, hold 170 back, and leave 822 slots emptyThose 678 clear break-even at the best estimate. The low end of the range loses money, so this is a measured bet.
  4. 04 Found Add-to-cart can be read by 9 Nov; whether it paid waits for returnsThe test can detect a 10.7% lift. Payoff is read on 9 Dec against a 22.1% break-even lift.
  5. 05 Checked The bet ran the way it was approvedAll 678 products went in on 6 Oct, and none of the 170 held back got the next-day badge.
  6. 06 Found The next-day badge lifted add-to-cart by 26.9%The range is 10.0% to 40.5%, clear of the 13.1% line written down before the test.
  7. 07 Found It paid about $23,919, and the range still crosses zeroGifts made $32,472. Lighting & Festive lost $11,748.
  8. 08 Checked Returns came in lower than for the products held back4.12% of stocked orders were returned, against 6.67%.

Decision

Iterate. Do not make the 678-product assortment permanent. Gifts leads next season, and Lighting & Festive returns only for a festive window, with its own products held back.

RMSigned by Ryan Mercer10 Dec, 10:18

What came out of it

  • Change · AppliedStock 678 measured Diwali products at BLR-1 and keep 170 ranked products as holdout
  • DashboardBLR-1 second-floor Diwali test
  • LearnedGifts leads the next expansion

    Lighting & Festive comes back only for a festive window, with its own products held back.

  • Continued · Change proposedGRO-1 · Next-day Gift badge discount pullback

    Owned by Anna Keller.

01Diwali demand sits in gifts, lighting and decor
02Forecasts from recent sales run hot

03Stock 678 products, hold 170 back, and leave 822 slots empty

Do not fill 1,500 slots before Diwali. Stock 678 products of the kind the first floor proved, with 1,396 opening units, keep 170 like them supplier-fulfilled to compare against, and leave 822 slots empty. The best estimate clears the 22.1% break-even lift, but the low end of the range loses money.

categorystockedheld backopening unitspredicted net
Gifts5391351,087$11,556
Lighting & Festive9825225$5,470
Decor411084$1,859
Total6781701,396$18,885
Warehouse via sql, through 28 Sep. Category effects come from SUP-1.QueryRun now
04Add-to-cart can be read by 9 Nov; whether it paid waits for returns
05The bet ran the way it was approved

06The next-day badge lifted add-to-cart by 26.9%

It worked. The overall lift of 26.9% is above the 13.1% line written down before the test, and even the low end of its range is above zero. Gifts and Lighting & Festive are each positive on their own. Decor points the same way but is not conclusive.

add-to-cart lift, % stocked against held back, by category
categorystocked, before → afterheld backlift
Total4.77% → 6.50%4.86% → 5.22%+26.9%
Decor4.33% → 5.51%4.31% → 4.56%+20.4%
Gifts4.34% → 5.92%4.60% → 4.77%+31.6%
Lighting6.23% → 7.90%6.15% → 6.51%+19.7%
Warehouse via sql, product pages, 6 Oct to 8 Nov against 1 to 28 Sep.QueryRun now

07It paid about $23,919, and the range still crosses zero

The best estimate says the test paid about $23.9k, driven by Gifts. Lighting & Festive lost $11.7k. The range, −$47.7k to +$94.3k, crosses zero, so the payoff is not proven.

categoryunits soldwithout the badgenet payoffrange
Gifts674635.8$32,472−$29,937 to $97,533
Decor150116.8$4,359−$7,953 to $16,454
Lighting & Festive539494.8−$11,748−$42,229 to $13,050
Total1,3631,225.6$23,919−$47,745 to $94,337

CaveatsHandling is a floor, because receipts are not logged. Stock left at the close is charged at what it costs to return.

Warehouse via sql, orders 6 Oct to 8 Nov, returns through 9 Dec.QueryRun now
08Returns came in lower than for the products held back

Stock 678 measured Diwali products at BLR-1 and keep 170 ranked products as holdout

Appliedfbs · done by NRNikhil Raoproposed 29 Sepapproved by NRNikhil Rao, 29 Sep

Work order

wherewhatfromexpected cost
Second floor · GiftsStock 539 products, open with 1,087 units6 Oct$84,733
Second floor · Lighting & FestiveStock 98 products, open with 225 units6 Oct$30,520
Second floor · DecorStock 41 products, open with 84 units6 Oct$8,725
Held backKeep 135 Gifts, 25 Lighting & Festive and 10 Decor products supplier-fulfilled, without the next-day badge, until 8 Nov6 Oct$0

Applied when the data shows it

All 678 stocked products are on hand at BLR-1 with the next-day badge live, and none of the 170 held back has one. Nikhil’s message that it was done is kept separately; the change stayed approved until the warehouse data agreed.

stocked · 678 of 678 liveheld back · 0 of 170 live

Predicted effect

Add-to-cart for stocked against held back rises 13.1% to 18.3% by 8 Nov. Net contribution is +$18,885 (−$33,791 to +$70,092). It worked if add-to-cart is up at least 13.1% and the final net contribution is positive on 9 Dec.

Rollback

Stop replenishing a category if its add-to-cart falls behind the products held back, or its returns run more than 0.5 points above them, two weeks in a row. Remove the badge and return the remaining stock.

Why, exposure, evidence and checks
Add-to-cart lift+26.9%range 10.0% to 40.5%
Net contribution+$23,919expected +$18,885
Stocked products live678 of 678from 6 Oct
Held back with a badge0 of 170none leaked
Add-to-cart, stocked against held back, by day
0.81.01.21.41.61.8 worked, 1.13 held back, 1.0 6 Oct13 Oct20 Oct27 Oct3 Nov change 1 applied
Product allocation
categorystockedheld backopening units
Gifts5391351,087
Lighting & Festive9825225
Decor411084
Total6781701,396
Predicted payoff range
scenariodemand liftnet contribution
lower7.0%−$33,791
point30.5%$18,885
upper58.6%$70,092
Decision thresholds
metricthresholdunit
Add-to-cart worked13.1% lift
Smallest lift the test can see10.7% lift
Demand break-even22.1% lift
Returns and failed deliveries, after 30 days
groupreturnedfailed delivery
stocked4.12%1.83%
held back6.67%3.17%

The bet this work led to, as the team sees it. Click through its tabs.

01It starts from what it knows

Veloccy began with how the pilot was picked.

That is in no table. Supply made the decision in April and growth launched a campaign in June, and Veloccy knew about both. When growth asked how +277% could be 15.5%, Veloccy walked their number down in the same thread, so two teams ended up on one number.

Simcorp HomeArcsSUP-1
NR
Nikhil Rao

Did the 857-product next-day pilot actually pay off? Should we expand?

Started from
Pilot list frozen · 857 productsBengaluru city campaignApp 4.8 · product-view break
The answer

The next-day promise lifted demand 15.5%, but the 857-product mix lost $294,403; do not expand it unchanged.

Gifts paid. Bedding and kitchen were confidently loss-making. Keep gifts, and test the next allocation with a random holdout.

Every number opens the query behind it
02It writes down what it expects

The answer became a bet, with its result predicted before anything shipped.

The warehouse was opening a second floor with 1,500 slots. Veloccy proposed filling fewer than half, held some products back to compare against, said what it expected to make, and took it to the person who owns the warehouse.

Simcorp HomeChangesCAT-1

Stock 678 products on the second floor, leave 822 slots empty

CAT-1
Predicted effect

+$18,885 (−$33,791 to +$70,092)

Held back

170 ranked products, at random

Rollback

Stop replenishing any category that falls behind the products held back

NRApproved by Nikhil RaoApproved
03It comes back

After Diwali, nobody had to ask.

Veloccy first checked the bet had run the way it was approved, then read the result against what it had written down. It also said what was not proven yet, and what to do next.

Simcorp HomeArcsCAT-1Readout
Veloccy came back with the result
Readout · CAT-1

The second floor paid about $23.9k, inside the predicted range.

Expected+$18,885
Measured+$23,919

Range −$33,791 to +$70,092. The measured range still crosses zero.

  • All 678 products went in as approved
  • None of the 170 held back leaked in
  • Add-to-cart up 26.9%
DecisionIterate. Gifts leads the next batch.
04It keeps all of it

The bet is now part of what Veloccy knows.

What it read, who approved it, what it expected, what it made, and what it led to. The next question about that warehouse starts from here.

The second-floor bet in the Brain: what it read on the left, what it led to on the right, and its expected and measured result.

Two months later, the same question went to Claude and to Veloccy.

Nikhil came back with a new question about the same warehouse. Claude had the data. Veloccy also knew why the floor was half empty.

Nikhil Raoruns supply at Simcorp Home10 Dec
More than half of the second floor is still sitting empty, and wedding season is starting. What should we fill the empty slots with?
ClaudeOpus 5.5, connected to the same warehouse
ClaudeFileEditViewWindowHelpConnected to warehouse
BLR-1 · December stocking

More than half of the second floor is still sitting empty, and wedding season is starting. What should we fill the empty slots with?

Queried the connected warehouse63 queries

Nikhil, short answer: fill it with SKUs that already sell fast, not with wedding stock. I could not find a wedding season in our sales, and the second floor is half empty mainly because the October batch put stock on the wrong SKUs.

What I would put in, in this order

OrderWhatSKUsUnits a month
1Fast movers that are not at BLR-1 yet (non-fragile)1187.4
2Restock the October SKUs that do sell1342.2
3Next tier of non-fragile sellers not at BLR-11492.2
4Fragile dining and decor, as a pilot1195.6

Orders 1 to 3 at a month of cover come to roughly 1,500 to 1,700 units across about 400 SKUs.

Reply…Opus 5.5

A careful list of 844 products, ranked on units sold. It fills slots left empty on purpose, mostly with the kind of product that already lost money here.

Veloccythe same warehouse, and the Brain
Simcorp Home›Arcs›December stocking
OverviewThreadFindingsChanges
NR
Nikhil Rao10 Dec

What should we fill the empty slots with?

Veloccy

Don’t fill it.

Those slots were left empty on purpose.

Stock 678 products, leave 822 slots emptyCAT-1

ChangeNRApproved by Nikhil Rao29 Sep

The products that would go in already lost money here.

Next-day pilot payoffSUP-1

ArcGifts paid. Bedding and kitchen lost.−$294,403

NR
Reply in the thread

Veloccy knows the slots were left empty on purpose, and that bedding and kitchen lost money on the first floor. All of it is in the Brain.

Then we ran both plans forward to the end of February.

Claude’s list went onto the shelves on 14 December. Each week shows what the plan had made or lost by then, after the cost of stocking it.

Claude’s plan−$270,340by 28 Feb
Veloccy’s plan$0822 slots left empty
14 Dec27 Dec10 Jan24 Jan7 Feb28 Feb
  1. 14 DecClaude’s 844 products go onto the shelves−$285,502
  2. 7 FebIts best week−$193,431
  3. 28 Feb4,762 units still on the shelf−$270,340
What the stock brought in+$548,563
What it cost to stock−$818,903
Where it lost the most

Bedding−$87,889Bath−$67,924Kitchen−$44,961

Claude did careful work with what the warehouse could tell it. The decision behind the empty slots and the first floor’s result are in the Brain.

Claude Opus 5.5 with read-only SQL on the same warehouse, one run. A demonstration on a simulated company, not a benchmark. The Claude window is a recreation; the words in it are Claude’s.

Why access to the data was not enough.

01The data alone says expand.

How the products were picked is a decision, not a column. No query returns it.

02Two teams had two numbers.

An answer inside one person’s chat stays there. Here supply and growth ended on the same page.

03Somebody has to come back.

An answer nobody checks adds nothing. A result read against a written prediction is experience.

What stayed in the Brain.

Every one of these is experience the next question starts from, in every team. See the Brain

HSimcorp Home
  1. ArcSUP-1Next-day pilot payoff

    The pilot’s real lift and what it cost.

  2. ChangeCAT-1Stock 678, leave 822 empty

    Why 822 slots were left empty, and who approved it.

  3. LearningGifts lead the next expansion

    What was learned: gifts lead the next expansion.

Next use case · Operations Everyone blamed the new kitchens. The cause was a lunch menu.

Which bet is on your desk this month?

Show us, and we will show you what Veloccy does with it.

Talk to us

Become a design partner

We’re working with a small number of teams before a wider release. Tell us a little about yours and we’ll reply personally, usually the same day.