Noibu blog

What Do AI Agents for Ecommerce Actually Cost? Tokens, Budgets, and Caps

Megan Glover

October 19, 2026

Blogs

Token-based AI cost meter for ecommerce analytics workflows, with usage caps and alerts

AI agents for ecommerce are usually priced on tokens, the units of text a model reads and writes. A quick question over summarized data costs very little; an agent that reads thousands of sessions every morning costs far more. Costs stay predictable when every workflow has a known scope, a cap, and an alert before overage.

TL;DR

Tokens are the meter. Every word an agent reads or writes counts, and reading store data is usually the biggest share.
Cost scales with scope and frequency, not with how hard the question sounds.
Bill shock comes from workflows that fan out: scheduled runs across every page, loops over hundreds of sessions, open access with no caps.
Before you sign, ask who pays for model usage: the vendor, or your own AI account.

Most conversations about AI for ecommerce skip the bill until it arrives. That is a mistake, because the cost of an AI agent depends far more on how a workflow is set up than on which model it uses. A team that understands what drives the meter can run agents every day without surprises.

Noibu is AI for ecommerce operations, the always-on optimization system that keeps online stores running at their best. A team of supervised AI agents finds what's broken, slow, or leaking revenue, does the work, and hands results back for human approval.

Why are AI agents for ecommerce priced in tokens?

AI agents for ecommerce are priced in tokens because the language models underneath them are. A token is a fragment of a word. Model providers charge for the tokens a model reads (input) and the tokens it writes (output), and output tokens usually cost more per token than input.

Ecommerce agents read far more than they write. To explain a drop in checkout conversion, an agent might read error groups, funnel steps, performance data, and a handful of sessions, then write a few paragraphs. That is why the volume of store data an agent reads drives most of the cost.

Vendors handle this in two ways. Some pass model usage through as metered pricing. Others bundle it into a plan with seats, credits, or usage tiers. Either way, the customer is paying for tokens; the difference is how predictable the bill is.

What makes an AI agent's bill go up?

Five things raise the cost of an ecommerce AI workflow, roughly in order of impact:

Data volume per run. Reading raw sessions or a whole catalog costs more than reading a ranked list of issues.
Frequency. A report that runs hourly costs about 24 times as much as the same report run daily.
Fan-out. One request that becomes a separate run per page, per product, or per session multiplies quickly.
Model size. Larger models cost more per token. Routine summaries rarely need the largest one.
Re-reading. Workflows that reload the same data on every run pay for it every time.

~6,000

real queries ecommerce teams have run through the Noibu plugin, led by error investigation, checkout analysis, on-demand dashboards, and daily briefs.

Noibu plugin usage analysis, 2026

Daily briefs show up near the top of real usage, and they are exactly the kind of workflow worth scoping carefully. A morning brief that summarizes ranked issues is cheap to run every day. The same brief rewritten to open every affected session is not.

What does an unexpected AI bill look like?

An unexpected AI bill usually starts with a workflow that worked well on a small scope. Someone builds a useful check for one page template, then points it at the whole store, or schedules it every hour, or shares it with the whole team. Nothing breaks, so nobody looks at usage until the invoice arrives.

The answer is not to use AI less. It is to treat each workflow like any other recurring cost: know what it reads, how often it runs, and what it is allowed to spend.

Who pays for the tokens: the vendor or you?

Who pays for the tokens depends on where the agent runs. When an agent runs inside a vendor's product, the vendor pays the model provider and prices that cost into its plan. When a vendor's data is connected to your own AI assistant through MCP (Model Context Protocol), the model runs in your AI account, so model usage counts against your AI plan.

Many teams use both. Noibu's agents work inside the Noibu console, and teams can also work with them from Claude through Noibu MCP. Agree up front which account carries which workflows, because “whose tokens are we burning?” is a common question once usage grows.

Which pricing models stay predictable?

Pricing model	How it works	Predictability	Watch for
Per seat	A flat price per user; the vendor absorbs usage variance	High	Fair-use limits that throttle heavy users
Credits or bundles with alerts	A monthly allowance with alerts before overage	Medium to high	The overage rate, and what happens at zero
Per-run caps	Each run has a maximum spend	High per run	Analysis that stops partway through
Pure pass-through	Pay for every token consumed	Low	No ceiling unless you set one

How do you keep AI agent costs under control?

Six controls for predictable AI costs

Scope every workflow. Name the pages, date range, and data it reads.
Set frequency on purpose. Daily is enough for most briefs; save hourly for peak events.
Cap per workflow and per user. A cap turns a runaway workflow into a stopped one.
Alert before the limit. A warning at 75% of budget beats a statement after the fact.
Prefer ranked summaries over raw re-reads. Reading a ranked issue list costs less than opening every session.
Review the top workflows monthly. Compare each one's cost with the decisions and changes it produced.

What should you ask a vendor before you sign?

What is metered: tokens, runs, seats, or a mix?
What is included in the base plan, and what triggers overage?
Can I see usage in real time, or only on the invoice?
What is the most I could spend in one day if a workflow ran unchecked?
Can I cap specific users, workflows, or agents?
When I use your data from my own AI assistant, which usage lands on my AI plan?

A vendor that can answer those quickly has thought about cost. A vendor that cannot is handing the problem to you.

Frequently asked questions

How much do AI tokens cost for ecommerce analytics workflows?

It depends on the model and on how much store data each run reads. A short question over summarized data costs very little, while a long run over raw sessions or a whole catalog costs much more. Scheduled workflows multiply that cost by every run, so scope and frequency matter more than the question itself.

Why did my AI bill spike unexpectedly?

Usually because a workflow ran more often, or over more data, than anyone planned. The common causes are scheduled agents pointed at every page, loops that open hundreds of sessions per question, and team-wide access with no per-user limits.

Is per-seat pricing cheaper than usage-based pricing?

Per-seat pricing is more predictable, not always cheaper. It suits teams with steady, moderate use. Usage-based or credit pricing can cost less for teams with occasional heavy analysis and light day-to-day use, as long as caps and alerts are in place.

Who pays for the tokens when I connect store data to Claude through MCP?

The model runs in your AI assistant, so model usage counts against your AI plan, while the data vendor charges for its own product. When an agent runs inside a vendor's product instead, the vendor pays the model provider and prices that into its plan.

How do I estimate the cost of a daily AI report?

Run the report once, read the usage it consumed, and multiply by how often it runs and how many stores or page groups it covers. Add headroom for busy periods, then set the cap at that number so a change in scope cannot quietly multiply the bill.