A supply chain can remove millions from cost to serve through AI and automation without reducing the company’s reported cost base by the same amount.
Chewy offers a clear example. More than half of its volume now passes through automated facilities, while AI is being applied across customer care, pharmacy and other operational workflows. The company expects these initiatives to generate savings in the low tens of millions of dollars during fiscal 2026, rising to approximately $50 million on an annualised basis in fiscal 2027.
But management has cautioned that those savings should not be treated as an equivalent increase in earnings. Some may absorb wage inflation and other cost pressures. Some may be reinvested. This exposes an important measurement gap. Improving a workflow and retaining the resulting value are not the same thing.
In Brief
- Chewy expects approximately $50 million in annualised AI savings in fiscal 2027, but not all of that value will reach margin.
- Lower variable cost to serve was the largest contributor to 70 basis points of year-on-year non-GAAP SG&A leverage.
- Separating productivity created from value retained provides a stronger measure of AI economics than deployment or gross savings alone.
AI Savings Are an Intermediate Result
AI business cases are frequently built around the cost of completing a task. A planning application reduces manual intervention. A warehouse system removes labour hours. An AI assistant shortens handling time. An automated workflow allows more transactions to be processed by the same team.
These are valid operational gains, but they are not the final economic result.
The enterprise outcome depends on what happens next. Released capacity may reduce expenditure, absorb inflation, support higher volumes, improve service or be redirected into other priorities. Each creates value, but not in the same form.
Chewy makes that distinction unusually visible.
Kai, its customer-facing AI assistant, is resolving approximately 30% of chats through self-service for common customer needs within its limited rollout. AI is also being used to support customer service employees and automate data extraction and validation in pharmacy operations.
Alongside these applications, more than half of Chewy’s volume now flows through automated facilities. The company has a multiyear programme focused on reducing variable cost to serve.
Together, those initiatives are expected to generate meaningful productivity. But the approximately $50 million annualised savings estimate does not show how much will ultimately improve earnings.
That requires three different outcomes to be separated:
- Productivity created by changing the work
- Capacity used to absorb pressure or support growth
- Financial value retained in the cost base
The first shows whether the technology is improving the operation. The other two show what that improvement has done for the enterprise.
Cost to Serve Connects Technology With Performance
Chewy’s second-quarter results provide evidence that its underlying cost base is improving.
Non-GAAP SG&A declined to 18.4% of net sales from 19.1% a year earlier, producing 70 basis points of leverage. Lower variable cost to serve was the largest contributor.
Chewy also identified improved fulfilment centre utilisation, disciplined headcount management and productivity gains from automation, process improvements and AI-enabled tools.
The company has not separated the contribution of each lever, so the improvement cannot be attributed to AI alone. That is precisely what makes the case instructive.
Operational performance rarely changes because of one isolated technology. Results emerge from the interaction between process design, network utilisation, labour deployment, automation, volume and management decisions.
A credible value case therefore needs to follow the complete sequence:
- AI or automation changes a defined workflow.
- The workflow requires less time, labour or intervention.
- Unit cost, throughput, accuracy or service improves.
- The released capacity is removed, redeployed or used to absorb pressure.
- The enterprise retains a measurable operational or financial benefit.
Many programmes measure deployment and workflow productivity. Far fewer follow the value through the final two stages.
Productivity Can Be Real Without Lowering Total Cost
An operation can become more productive while total expenditure remains unchanged.
If an AI-enabled process removes a dollar of recurring labour cost while wage inflation adds a dollar elsewhere, the productivity gain still occurred. It prevented the overall cost base from rising further rather than reducing reported expenditure.
The same applies when volumes are growing.
Automation may allow a fulfilment network to process more orders without adding equivalent labour or infrastructure. An AI tool may enable a planning or customer service team to manage greater complexity without increasing headcount. The value appears as avoided cost or scalable capacity rather than an immediate reduction in spending.
The benefit may also be deliberately reinvested. Released capacity could be used to improve service, strengthen resilience, develop new capabilities or address work that was previously under-resourced.
These are legitimate enterprise outcomes, but they need to be identified and measured.
Otherwise, operational teams can report substantial savings while finance sees little movement in the cost base. Successful transformation then appears to have delivered less than it actually has, weakening confidence in subsequent investment.
Reported Margin Can Obscure the Same Distinction
Chewy’s second-quarter margin performance shows how reported financial results can complicate the picture in the opposite direction. Adjusted EBITDA margin reached 6.8%, above guidance of 6.3% to 6.4%. Yet management attributed essentially all of the outperformance to timing-related and discrete benefits.
Approximately $10 million primarily reflected earlier tariff refunds and shifted rebate benefits. More than $5 million came from items including gift card breakage, inventory adjustments and vendor-funded merchandising.
Those benefits improved the reported quarter without creating an equivalent structural improvement in Chewy’s recurring operating model.
AI savings can create the reverse effect. Structural productivity may be generated without appearing fully in reported margin because inflation, growth or reinvestment absorbs the benefit.
Margin can therefore overstate underlying progress when temporary items lift a quarter and understate it when genuine productivity is used elsewhere. Understanding performance requires a clearer view of what sits behind the movement.
The Missing Measure
Chewy has not disclosed enough information to calculate how much of its forecast productivity the business will retain.
It has not separated savings by workflow, detailed the full cost of implementation or shown how much of the expected benefit will offset inflation, support growth, be reinvested or improve earnings.
Kai is also at an early stage. Less than 10% to 15% of customer traffic had been exposed to the assistant after less than a month in the market. Its current performance cannot yet be extrapolated across the entire operation.
But Chewy’s disclosures point towards a better measurement model.
For every material workflow, companies should be able to show:
- The original cost and performance baseline
- The recurring productivity created
- The technology and transformation costs incurred
- The capacity removed or redeployed
- The cost increases absorbed
- The improvement in service, throughput or resilience
- The net value retained by the enterprise
This changes the question from whether AI has produced savings to what those savings have enabled.
Chewy’s approximately $50 million estimate suggests that AI can create meaningful economic capacity. Its greater significance is that management has acknowledged that productivity creation and profit improvement are not the same thing.
Without measuring both, a company can transform how work is performed while remaining unable to show where the value of that transformation went.
