Written by: Michał Sobieraj, Operations Manager, Digital Colliers
The story making the rounds is that Amazon staff reportedly found internal cases of catastrophically expensive AI usage. One example: around $1.8M spent on Claude just to match author details with book listings. Whether the exact number holds or not, the shape of the problem is real, and it's coming for every retailer that's currently piloting AI without a cost model underneath it.
The uncomfortable part is that most eCommerce teams cannot tell you, per workflow, what a single AI decision costs them today. They can tell you the monthly bill from OpenAI or Anthropic. They cannot tell you what it cost to enrich one product, answer one ticket, or re-rank one search result. That's the gap this article is about.
AI spend behaves like ad spend, but gets booked like SaaS
Here's the mental model shift that hasn't happened yet in most finance functions. SaaS is a fixed seat cost. You pay per user per month, and usage inside that seat is basically free at the margin. AI inference is the opposite. Every call has a variable cost that scales with tokens, model choice, and how chatty your prompts are.
That behaviour is much closer to paid media than it is to Salesforce. You would never run Meta ads without knowing CPM, CAC, and ROAS per campaign. But we're running LLM workflows without the equivalent. And the timing is bad. Meta CPMs for DTC advertisers kept rising through 2024 and 2025, and customer acquisition cost across DTC brands is up roughly 40% since 2023. Margins are already thin. Adding an unmetered variable cost on top of that is how you get a $1.8M line item nobody can explain.
Meanwhile the accounting treatment lags. AI shows up as a single vendor invoice on the P&L, bundled with dev tools, sitting under IT. Nobody in merchandising sees it. Nobody in category management sees it. So nobody can tell if the enrichment workflow on a £12 SKU is eating the margin.
What a workflow-level cost data model looks like
The teams that will avoid the blown-budget headline in 2026 are the ones building a FinOps layer for AI now, before it hurts. The shape of it is not complicated:
- Tag every inference call with a workflow ID. Not a user ID, not a model name. A workflow. "product-enrichment-v3", "returns-triage", "pdp-copy-refresh".
- Log token counts in and out, model used, and unit cost at call time. Store it in your warehouse, not just the provider's dashboard.
- Roll up to cost per unit of business outcome. Cost per SKU enriched. Cost per ticket resolved. Cost per search query served.
- Set a budget per workflow with an alert threshold. If product enrichment starts costing 4x what it did last week because someone swapped the model, you want to know that day, not at month-end.
None of this is exotic. It's the same discipline paid media teams have had for a decade. The lift is mostly about deciding to instrument it before the bill arrives, not after.
Tie the cost back to margin per SKU
This is where it gets useful for retail specifically. Roughly 30% of SKUs at a typical multi-channel eCommerce brand already lose money per order once you account for returns and ad spend. Online return rates sit around 19-20% of gross sales, higher in apparel. If AI enrichment, personalisation, or support automation adds even a small variable cost per order, it lands hardest on the SKUs that were already underwater.
A workflow-level cost feed lets you join AI cost to SKU-level P&L. You find out that enriching a £8 accessory is costing you 40p in inference, and the workflow is running on every catalogue refresh whether the SKU sells or not. That's a decision merchandising should be making, not engineering.
Finance and merchandising need the same view
Around 95% of enterprise AI projects reportedly fail to reach production or ROI. A meaningful share of that failure isn't the model, it's that nobody could prove the unit economics worked. UK eCommerce grew about 3% in 2024. Single-digit growth is the baseline now. There isn't room in the P&L for AI workflows that quietly cost more than they earn.
The operators who'll come out of 2026 ahead are the ones where finance, merchandising, and engineering are looking at the same cost-per-workflow dashboard. Not three different tools telling three different stories. One number, joined to SKU margin, updated daily. That's the layer to build before the first blown budget, not after.