What an AI Agent Actually Costs to Build for an E-commerce Business

Why nobody can quote you an AI agent cost off a one-line brief

When someone asks "what does an AI agent cost to build for my store?", the honest first answer is a question back: what do you want it to actually do, and to how many systems does it need to talk? The AI agent cost is not driven by the AI. It is driven by everything the agent has to touch to do useful work.

The model — the part everyone thinks they are paying for — is close to a commodity. You call an API, you pay per token, and the price keeps drifting down. The expensive, slow, judgment-heavy part is the integration: wiring the agent into Shopify, into WhatsApp or email, into your payment stack, into your helpdesk, and into whatever spreadsheet or internal tool your team quietly runs the business on. That is where most of the build hours go, and it is why any single number you have seen quoted online is almost meaningless without the scope behind it.

This post gives you a framework to estimate and compare, not a price. If you want to see the shape of the work we do here, that lives on the AI agents page. What follows is how to think about the bill.

The model is the cheap part

Treat the language model as a utility. It reasons, it drafts, it classifies, it decides. You rent that capability by the token, and for most e-commerce workloads the per-request cost is small enough that it is not the thing to optimize first.

• API usage scales with volume, not complexity. A support-triage agent handling a few hundred conversations a day sits in a token-cost range that is a rounding error next to a person's salary.
• Prompt and context design is engineering, not model choice. Getting the agent to behave reliably is about what you feed it and how you constrain it — not about paying for the largest model available.
• You can usually mix models. A cheaper model handles the routine classification; a stronger one is reserved for the hard cases. That is a design decision, and it is where cost control actually happens.

If a vendor's pitch is mostly about which model they use, they are selling you the cheap part.

The real cost drivers

Here is what the estimate actually turns on. Every one of these moves the number more than the choice of model does.

• Scope. "Answer WISMO questions" is one agent. "Answer WISMO, issue refunds within policy, escalate fraud signals, and update the helpdesk ticket" is four capabilities, each with its own failure modes. Scope is the single largest lever on cost.
• Number of integrations. Every system the agent reads from or writes to is a connection to build, authenticate, error-handle, and maintain. One integration is a project. Five integrations is a much larger project, because the failure combinations multiply, not add.
• Data access and permissions. Reading a Shopify order is easy. Getting scoped, auditable, least-privilege access to a payment processor or a customer database — without handing an autonomous system keys it should not have — is real work, and it is the part that protects you later.
• Human-in-the-loop review. An agent that drafts a reply for a human to approve is cheaper and safer to build than one that acts on its own. The moment you want it to move money or make irreversible changes unattended, you are paying for guardrails, approval flows, and audit trails — and you should be.
• Edge cases. The demo works on the happy path in an afternoon. The cost is in the long tail: the partial refund, the duplicate order, the customer who replies in three languages, the webhook that arrives twice. Most of the build budget is spent making the boring 15% not break things.

A useful mental model: estimate the demo, then assume the production version is several times that, and the difference is almost entirely edge cases and integration hardening.

One-time build vs. ongoing cost

This is the part that surprises people, so it is worth being blunt. An AI agent is not build-once software you install and forget. It is an operated system, closer to a small service you run than to a plugin you buy.

The one-time build is the design, the integrations, the guardrails, and the testing. Then there is the ongoing cost, which does not go away:

• Hosting and infrastructure. The agent runs somewhere. Servers, queues, logs, and databases have a monthly cost even when nothing is broken.
• Monitoring. You need to see what the agent is doing, catch when it starts behaving oddly, and get alerted before a quiet failure becomes a week of wrong answers. It is the same discipline reliable Shopify webhook systems need — an unattended process that fails silently is worse than no automation, because you trust it until you find out.
• Model and API changes. Providers deprecate models, change behavior, and shift pricing. Shopify changes API versions. WhatsApp changes its messaging rules. [VERIFY: WhatsApp Business Platform messaging-policy and pricing changes over the past year]. Each change can require work to keep the agent running.
• Fixes and drift. Your catalog changes, your policies change, your team changes how they work. An agent tuned for last quarter's process slowly falls out of step unless someone maintains it.

If a quote covers only the build and says nothing about running the thing, it is quoting half the cost. Ask what the monthly number looks like once it is live.

Build vs. buy vs. hire in-house

Three honest paths, and the right one depends on how specific your problem is.

• Buy an off-the-shelf tool. Fastest and cheapest to start, if a product happens to fit your process closely. The catch is that these tools are built for the average merchant, and the value of a custom agent is usually in the parts that are specific to you. [VERIFY: pricing tiers of the main off-the-shelf e-commerce support-AI tools]. Good for common, standardized workflows; frustrating when your business does something unusual.
• Build custom. Higher upfront cost, but the agent fits your actual systems and rules, and you own how it behaves. This makes sense when the workflow is core to your operation and the off-the-shelf fit is poor.
• Hire in-house. Bringing on an automation engineer is not a one-time cost — it is a salaried, managed, ongoing commitment, plus the ramp-up time before they know your stack, plus the management overhead. In the US, a machine-learning engineer averages roughly $165,000–$190,000 a year in base salary alone (Glassdoor, Indeed), before benefits and management time. For one agent, that is usually more than the agent is worth. For a roadmap of ten automations over two years, it can be the cheaper path.

The comparison is not build-price vs. build-price. It is total cost of ownership over a couple of years, including the months you would spend hiring and the salary you would keep paying.

How a fixed-price pilot de-risks the first build

Because the biggest unknown is scope and edge cases, the riskiest thing you can do is sign an open-ended contract for a system nobody has fully specified. That is how budgets run over.

We work the other way around, with a fixed-price pilot followed by a monthly ops plan. The pilot does a few specific things:

• It puts a ceiling on the first cheque. You know the pilot number before work starts, so a scoping mistake is our problem to absorb, not a surprise on your invoice.
• It forces a narrow, real scope. One workflow, wired into your actual systems, run against your actual data. That is worth more than a broad plan, because it surfaces the edge cases early — while the price is fixed.
• It gives you a real basis to compare. After a pilot you are no longer estimating from a brief. You have seen the agent handle your traffic, so the decision to expand it is made on evidence, and the ongoing number is grounded in observed volume rather than a guess.

The pilot answers "will this work for us and roughly what will it cost to run" for a known, capped price. The details of how the pilot and the ongoing plan are structured are on the pricing page.

Red flags in a quote that looks too cheap

A very low number usually means the quote is missing something you will pay for later. Watch for these:

• No mention of ongoing cost. A build-only price for an operated system is not a real total. The hosting, monitoring, and maintenance are coming whether the quote names them or not.
• Integrations waved away. If "connect it to your helpdesk and payment stack" is one line with no discovery, the integration work has not been thought through, and integration is the expensive part.
• No talk of edge cases, review, or guardrails. A quote that only prices the happy path is pricing the demo, not the production system.
• Autonomy promised cheaply. An agent that acts on its own, safely, costs more than one that drafts for human approval. A cheap quote for full autonomy is quietly skipping the guardrails — which is where the risk to your business actually sits.

The goal of all this is not to make an AI agent sound expensive. It is to let you read a quote and know what it is really covering. If you can describe your scope, count your integrations, and separate the build from the running cost, you can compare offers honestly — and you will know when a number that looks like a bargain is just an incomplete one. When you are ready to put real scope against a fixed number, that is what the AI agents work is built to do.