Project B · Policy Analysis · March 2026

Shopper Earnings at Risk: Modeling the Fulfillment Cost of Agentic Commerce

Jessica Batbayar · Instacart APM Projects · Kaggle Instacart Dataset (2017, 3.2M orders)

48.5%

Lower hourly earnings on high-complexity orders (base pay only)

−12.6%

Projected shopper earnings at 25% Cart Assistant adoption

3.2M

Orders analyzed

The side of the marketplace that wasn’t the primary audience

Cart Assistant was built to serve retailers and their customers. Retailers get a competitive, AI-powered shopping experience. Consumers get a smarter, more personalized way to build their grocery baskets. Advertisers benefit indirectly: higher-intent consumers building larger, more intentional baskets means better placement opportunities and stronger conversion on sponsored products.

What is less examined is what happens to the fourth side: shoppers.

When a consumer checks out through a Cart Assistant-powered retailer experience, the order is passed to Instacart’s fulfillment engine and assigned to a shopper. The shopper receives a standard SKU list. The recipe Cart Assistant built the cart around, the dietary constraints the consumer mentioned, the preference hierarchy it optimized for: all of that stays in the session and goes no further.

Shoppers are not a direct intended beneficiary of Cart Assistant. They do not interact with it, they do not configure it, and they do not receive its session context. They just walk into a store with a flat item list and a flat batch rate, regardless of how complex the order is.

Building a complexity score

Since Cart Assistant order data is not publicly available, I defined a “high-complexity order proxy” using five structural features from the Kaggle dataset. Each feature was normalized to a 0–1 scale and combined with equal weights into a composite score per order. Orders were then split into Low, Medium, and High tiers using quantile-based segmentation.

Feature	Rationale
Order size (item count)	More items = more decisions and substitution risk
Unique aisles visited	More aisles = longer physical pick path
Department diversity	Crossing departments signals a multi-meal or multi-purpose cart
Reorder rate (inverted)	Low reorder rate = more unfamiliar items = more decision time per item
Days since prior order	Longer gaps signal a full shop, not a top-up

Equal weights are the defensible default when no ground-truth outcome variable is available to learn weights from (like actual observed pick time from the Shopper app). A production model would calibrate these weights against real telemetry.

What the earnings model found

Hourly earnings by complexity tier, base pay vs. with proportional tip included — Hourly earnings by complexity tier: base pay only vs. proportional tip included

Tier	Avg hourly (base only)	Avg hourly (with tip)	Avg fulfillment time
🟢 Low	$16.33/hr	$22.07/hr	30.4 min
🟡 Medium	$11.94/hr	$18.26/hr	45.1 min
🔴 High	$8.41/hr	$14.46/hr	67.8 min

What happens as Cart Assistant adoption grows

Projected average hourly earnings under 10%, 25%, and 50% Cart Assistant adoption — Projected average hourly earnings under three Cart Assistant adoption scenarios

Scenario	Projected avg hourly	Change from baseline
Baseline (today)	$12.23/hr	–
10% shift to High complexity	$11.61/hr	−5.1%
25% shift to High complexity	$10.69/hr	−12.6%
50% shift to High complexity	$9.14/hr	−25.3%

The base-case 1.5x complexity multiplier for High-tier orders was stress-tested at 1.3x and 1.8x. The earnings gap ranged from 43.6% to 54.3% across all three settings. The conclusion holds across all parameter choices tested.

Why each side of the marketplace has a stake in this

Shopper

The earnings gap is invisible to the people experiencing it. They just notice their hourly yield feels lower on certain orders, and over time they either start avoiding complex batches or leave the platform altogether.

Consumer

Cart Assistant’s highest-intent orders get deprioritized or fulfilled with errors by time-pressured shoppers. The consumer-facing improvement degrades at the fulfillment layer.

Retailer

Retailers evaluate Cart Assistant on outcomes. When shoppers make errors on complex orders, retailers may lose confidence in the product even though the underlying issue is about how shoppers are compensated for complex work.

Advertiser

Cart Assistant tends to generate larger, more intentional baskets built around meals and dietary goals. That is exactly the kind of shopping context where sponsored product placement performs well. If fulfillment quality on those orders degrades, the consumer experience that makes them worth targeting gets worse too.

Three options and a recommendation

The earnings gap exists because high-complexity orders take longer to fulfill, and they take longer because shoppers navigate unfamiliar items without the context Cart Assistant already has. The three options differ in whether they address that time cost directly or compensate for it after the fact. I recommend Option B long-term, with Option A as a bridge while it is being built.

Pay shoppers a higher flat rate for orders that exceed a complexity threshold. This directly compensates for the extra time and can be deployed as a market pilot quickly. The order still takes just as long to fulfill. Shoppers earn more for the added effort while the longer fix is being built.

Deployable without a new data pipeline. Does not reduce complexity or give shoppers any additional context about the order.

Give shoppers a summary of the Cart Assistant session before they start picking: the recipe, the dietary constraint, the substitution boundaries. This shortens the time cost at the source by giving shoppers the context they need to make faster, more confident decisions. It also connects directly to Project A, which requires the same Cart Assistant context to flow into the substitution engine. One pipeline, two surfaces.

Gets at the root cause directly. Requires a new data pipeline from Cart Assistant to the Shopper app, which makes this a longer build than Option A.

Let shoppers opt into high-complexity batches in exchange for a pay premium. Market-driven and autonomy-preserving. It also generates real data about what complexity is worth based on what shoppers actually choose. A shopper who opts in still arrives at the first aisle without any context about why the order is built the way it is, so the time cost does not go away.

Useful as a data-collection mechanism and could run alongside Option A. Does not reduce the time cost or close the context gap.

The full tradeoff analysis and success metrics for each option are in the policy brief.

Connection to Project A

Project A asks: when a Cart Assistant-generated order hits an out-of-stock item, is the substitute safe for the consumer? Project B asks: what is the cost to the shopper of navigating these decisions without context, and does that cost scale with order complexity?

Both projects trace back to the same missing piece. Cart Assistant session context needs to flow downstream: to the substitution engine in Project A, and to the shopper before they start picking in Project B. Building that pipeline once and routing it to both surfaces is more efficient than building it twice, and it makes the combined case for investing in it much harder to ignore.

Back to Project A

Catalog-Powered Smart Substitution via Store View →