As a trusted Shopify Tax Platform partner, Avalara automates U.S. and global tax, from real-time calculation to streamlined tax returns, all with one solution.

Most AI projects fail before the first line of code gets written. Anne-Claire Baschet, Chief Data and AI Officer at Mirakl, runs the exception. Three things from this conversation are worth your time.

Lesson one: The problem you pick matters more than the technology you choose.

MIT says 95% of AI projects produce no ROI. Anne-Claire's explanation is blunt: they solved the wrong problem. Companies ran toward the long tail — low-stakes experiments, edge cases, things that wouldn't disrupt anything. Safe bets that produced nothing.

Mirakl went the other direction. They identified the three things that actually drive commerce value — catalog onboarding, retail media, and, software usability — and went all-in on those. Not the periphery. The core.

That's not a technology decision. That's a product decision. The companies still debating which LLM to use haven't answered the prior question: what problem are we solving, and does it matter? Get that wrong, and the model is irrelevant.

Lesson two: The adoption unlock isn't training. It's eliminating the copy-paste game.

Mirakl started at 50% Generative AI adoption in June 2024. By the time this conversation happened, they were at 95%. The breakthrough wasn't a better onboarding program. It was agents that operated within systems rather than generating text for humans to manually move elsewhere.

When people still had to copy output from an AI tool into Google Sheets, a ticket, a doc — adoption stalled. The moment the agent connected directly to the workflow, the value was immediate and obvious.

The RFP example is the cleanest proof point. The security and legal team was drowning in questions they didn't have time to answer. An agent connected to their SOC 2 certifications and previous responses now drafts answers directly in Google Sheets with a confidence score attached. Humans review, adapt, and ship. They went from choosing which RFPs to skip to answering all of them.

That's not efficiency. That's a structural shift in what's possible.

Lesson three: Shipping AI-native B2B SaaS is a data quality problem first.

Mirakl's Catalog Transformer has been in production for 18 months. The numbers are real: 50% reduction in categorization errors, 30-point improvement in attribute completion, and onboarding time cut from 15 days to under two hours. Rakuten onboarded 6 marketplaces in under a week before Black Friday.

What Anne-Claire learned while building it is what most product teams skip: your data has to be optimized for agents, not just connected to them. Documentation written for humans doesn't retrieve cleanly. Taxonomy built for search doesn't map cleanly to LLM outputs. The quality of work upstream is what determines whether the agent is useful or a liability.

The companies announcing AI features without rearchitecting their data layer are building on sand.

The through-line across all three lessons: this is a product discipline, not a technology experiment. Solve the right problem. Remove the friction. Fix the data. The model is the last variable, not the first.

Dollar General (results before market open on March 12): Trade-down got them here, but you can't comp against your own tailwind forever.

TL;DR: $68.2B in sales, up 9.1%. Net income up 13.8%. Strip out the September 2024 membership fee hike and underlying membership growth is still 7.5% — members absorbed it and didn't leave. That's the whole model confirmed in one data point. The one tension worth watching: digital acquisition is slowly diluting the 92.1% renewal rate. Nobody on the call said it that way.

Costco printed a near-perfect quarter. The press release says it all. The call said something the press release didn't.

The Membership Math

Membership income grew 13.6%. Roughly a third of that came from the September 2024 US/Canada price increase. Strip that out and underlying growth is still 7.5%. Members absorbed the hike and didn't leave. At a 92.1% renewal rate, that confirmation is worth more than the incremental fee revenue.

The Renewal Rate Tension Nobody Named

US/Canada renewal fell 10 basis points sequentially. Costco's explanation: online sign-ups renew at lower rates, and that cohort is growing.

That's not a scandal. It's a structural tension. Digital acquisition is diluting the most important number in the Costco model, one basis point at a time. What I did not hear on this call was where online renewal rates actually sit relative to the in-warehouse cohort — or what Costco is doing to close the gap.

The 92.1% number still looks elite. But the direction matters.

The Digital Build

Personalized carousels alone drove $470M in ecommerce sales. App traffic up 45%. Site traffic up 32%.

This is AI-driven personalization showing up in revenue, not in a slide deck. Costco is building a behavioral data asset inside a warehouse model. The P&L doesn't fully capture the compounding value of that yet. When it does, it will look obvious in hindsight.

The Number Doing the Most Work

Traffic growth: 3.1%.

Egg and produce deflation is compressing ticket. Traffic is carrying the comp. As long as members keep showing up, the model holds. The day traffic softens is the day the story changes. Watch that number above everything else.

The Tariff Question

Vachris called it "extremely fluid" with at least 150 days of ambiguity. Costco's answer is the right one: Kirkland Signature absorbs the cost pressure, sourcing diversification reduces the exposure, and price investment protects the member relationship where necessary. They've run this playbook before. The limited-SKU model means they can move faster than a full-assortment retailer when needed.

The Real Estate Bet Nobody Discussed

Multi-story warehouses with integrated residential in dense urban markets. 30-plus annual openings targeted long-term.

That's not filling white space. That's a thesis that the warehouse format has a decade of runway in markets Costco hasn't touched yet. If that bet is right, the TAM conversation looks completely different in five years.

The Verdict

Every operator competing on price without a membership relationship is watching their structural disadvantage compound. Costco is layering a behavioral data asset on top of the moat. The limited SKU model, the treasure hunt, and the renewal engine — none of that is new. The data flywheel on top of it is.

In short: near-perfect quarter, one structural tension, and a digital and real estate bet that will look either prescient or obvious five years from now.

The 3.1% traffic number is the tell. Watch it.