What AI costs in retail and ecommerce: budgets that hold up

Most retail and ecommerce AI builds land between $12,500 and $70,000 to deliver, or roughly ₹8 lakh to ₹46 lakh, plus a running bill of a few hundred to a few thousand dollars a month. A support agent sits at the bottom of that range; a personalisation engine spanning catalogue, email and WhatsApp sits at the top.

That range is wide because the model licence is rarely the expensive part. This article breaks the AI cost retail and ecommerce teams actually pay into four lines, prices each of them with real published figures, shows the costs a first budget usually omits, and says plainly when the number should be zero.

The four lines in a retail AI budget

Every retail and ecommerce AI programme we quote resolves into four lines, and finance teams get surprised when they only planned for the first one.

The build line is engineering: retrieval over your catalogue and policy documents, tool contracts against your order management system, guardrails, evaluation suites and the admin screens your operations team needs. The running line is inference, vector storage, telephony or messaging fees and monitoring. The integration line is the work of reaching your commerce platform, warehouse system, courier APIs and customer data platform, which is usually underestimated by a factor of two. The change line is training staff, rewriting policy so an agent can follow it, and the weeks of shadow mode before anything acts alone.

A useful rule from the projects we have shipped: if the build line is X, expect integration to be a third to a half of X again when the commerce stack is older than five years, and expect running costs in year one to be five to fifteen per cent of build.

Retail and ecommerce automation also carries a line that software buyers in other sectors do not face as sharply: seasonality. A manufacturing copilot has roughly the same load in March and October. A storefront agent does not, and a budget built on twelve equal months will be wrong in both directions, over-provisioned for nine of them and short for the three that generate the margin.

What does AI cost in retail and ecommerce by use case?

Published starting prices give you a floor per use case, and the table below maps the five things retailers ask for most against what they cost and what drives their monthly bill. All figures are Eazyware's own published starting prices, listed in full on the pricing page.

The upper bounds matter as much as the floors. Personalisation engines run to $70,000 or ₹46,40,000 when they cover several surfaces and a real experimentation framework. A customer service agent runs to $42,000 or ₹28,00,000 when it acts on refunds and exchanges rather than only answering.

The lines a naive retail budget leaves out

These are the items that turn an approved number into an overrun, in the order they usually bite.

• Catalogue hygiene. Retrieval quality is capped by attribute quality. Budget a fortnight of data work before anyone measures the model.
• Courier and marketplace APIs. Every carrier has its own tracking contract and every marketplace its own feed schema. Each one is integration work, not configuration.
• Messaging fees. WhatsApp conversations are billed by the platform per conversation category, separately from model tokens. Volume forecasts belong in the budget.
• Policy rewriting. An agent cannot apply a returns policy written as three paragraphs of prose with four unwritten exceptions. Somebody has to codify it.
• Peak-season headroom. Sale traffic multiplies inference and support volume in the same week. Capacity you never tested is capacity you do not have.
• Evaluation maintenance. Golden question sets go stale as the catalogue changes. Refreshing them is a recurring cost, not a project task.
• Returns fraud review. Automating exchanges without a fraud check moves cost rather than removing it.

What you pay every month once it is live

Running cost splits into usage and care. Usage is billed by your own model provider on your own account, which is how we set every engagement up: you hold the keys, we set budgets, routing and dashboards. Model vendors publish per-token rates openly, and OpenAI's pricing documentation lists cost per million input and output tokens, so a forecast is arithmetic rather than guesswork once you know volume. The LLM inference cost calculator turns conversation volume into a monthly figure.

Care is the other half. Eazyware Care Plans start at $1,000 or ₹68,000 a month for Essential, which covers business hours in IST with an eight-hour response and ten hours of work. Standard is $2,500 or ₹1,60,000 with 24x5 cover and four-hour response. Enterprise is $5,250 or ₹3,40,000 with 24x7 cover, one-hour response and a named engineer, which is what retailers on a sale calendar usually want. The AI system add-on at $750 or ₹40,000 covers evaluations, cost monitoring, prompt regression and re-indexing, and for a catalogue that changes weekly it is not optional.

Peak season is a budget line, not a risk register entry

Indian retail concentrates a disproportionate share of the year into a handful of sale weeks, and an AI system priced on average volume fails in exactly those weeks. Two numbers change together: inference spend, because conversations and ranking calls multiply, and human escalation, because the long tail of odd questions grows faster than the common ones. Model your peak week explicitly, set a hard spend ceiling with routing to a cheaper model above a threshold, and load-test the retrieval path rather than only the storefront. The engineering side is covered in peak-season readiness.

There is a second-order effect worth pricing. During a sale, the questions customers ask change shape as well as volume: fewer product questions, far more order status, delivery promise and cancellation questions, all of which need a live read from the order management system rather than a document lookup. That shifts load from cheap retrieval calls to expensive tool calls against systems that are themselves under strain, which is why peak-week capacity planning has to cover your order APIs and not only your model spend.

When AI is the wrong line in your retail budget

There are three situations where we tell retailers to spend the money elsewhere, and we would rather say it before an invoice than after.

The first is a catalogue nobody owns. If product attributes are inconsistent across categories, if the same item has three titles, or if stock levels in the storefront disagree with the warehouse, no amount of retrieval engineering fixes the underlying data. Spend the first tranche on data ownership and pipelines instead.

The second is a support queue that is small or dominated by one fixable defect. If eighty per cent of tickets are about a broken tracking page, fix the tracking page. A support agent that answers the same question elegantly forty thousand times has industrialised a problem rather than solved it.

The third is low-margin, low-volume catalogue work where the payback never arrives. A personalisation engine at $21,000 or ₹13,60,000 needs enough sessions for uplift to be measurable against noise. Below a few hundred thousand sessions a month, the honest answer is better merchandising rules and a cheaper build.

What a real retail engagement looks like on the invoice

A growing direct-to-consumer brand came to us wanting personalisation and a WhatsApp support agent at the same time. We sequenced them rather than running both, starting with the recommendation surface where uplift could be measured, then adding the support agent on the same customer data once the plumbing existed. The work is described in the personalisation and WhatsApp case study, and the sequencing mattered more to the budget than any single price. Shared retrieval and shared customer context meant the second build reused the first one's integration work instead of repeating it.

Where the use case is not yet settled, a ten-day Sprint Zero is the cheapest way to find out. The AI Discovery Sprint is $3,250 or ₹2,00,000, fixed, and credited against the build that follows. A three-week ProofRun at $6,250 or ₹4,00,000 proves the hardest intent against your real catalogue before you commit to a full programme.

A costing checklist before you approve anything

• Write down peak-week volume, not average volume, for every metric the system touches
• List each external system the AI will read or write, and whether it has a documented API
• Name the person who owns catalogue data quality and give them time in the plan
• Decide which actions need human approval, because gates cost engineering but save refunds
• Set a monthly model spend ceiling and the routing rule that enforces it
• Choose a Care Plan tier against your sale calendar, not your average month
• Agree the success metric in rupees or dollars before the first sprint, not after launch

Related reading

Demand forecasting for retail inventory covers the use case with the clearest payback arithmetic, total cost of ownership for AI systems extends the four-line model across three years, and the retail and ecommerce industry page sets out how we approach the sector. If you want a number against your own volumes, the contact page is the fastest route.

Budget the integration and the peak week honestly, and the model bill turns out to be the least interesting line on the page.