Most advice on building an AI agent is written for developers with an API key and endless patience. This is written for people who run a store and need a straight answer on cost. Three ways to put AI to work, what each one runs and builds for, and how to tell which one you actually need.
Some version of this question lands with me almost every week. Should I build an agent for this, or is a Claude Project enough? And what will it actually cost me?
The honest answer starts with a distinction most people skip. Every AI setup carries two separate costs, and mixing them up is where budgets go wrong.
The first is what it costs to run: platform fees and API usage. That number is almost always small.
The second is what it costs to build well: the thinking, the integration, the testing, the guardrails. That number scales with one thing, and it is not the tool. It is how much damage is done if the system gets it wrong.
The tools are cheap. Getting them to solve the right problem, safely, at scale, is what you pay for.
The Landscape
Three ways to put AI to work, and what each one costs
Platform and API prices below are indicative and quoted in USD. Build estimates are quoted in AUD. Treat every figure as a starting point, not a quote. Pricing in this space moves fast.
01 ·
Claude Project or custom GPT
A shared workspace with your business loaded into it. You upload documents, write instructions, and your team works through a configured interface. No code. A human still drives every conversation. You have simply made sure the AI knows who you are before it starts.
Runs for: roughly the price of a team seat, in the ballpark of $20 to $30 per user per month. No API fees, no infrastructure.
Builds for: a few hours of your own time if you set it up yourself. The output is only ever as good as the information you load in, which is why some brands pay to have it done properly. Most teams can get a working version on their own.
Maintenance: a document refresh each quarter. Anyone on the team can do it.
Right for you if: your team needs consistent output from the same source of truth, and the work still involves a human prompting and reviewing. This is roughly 80% of what most retailers actually need.
In practice: an email producer for a fashion label. You hand it a topic. It reviews the brand’s best-performing subject lines through the Klaviyo connector, drafts content in the voice loaded into its instructions, then produces a brief and a mock-up with product images pulled straight from Shopify. Here is what changed in the last year: you no longer upload everything. The connectors pull the live information themselves. A person still picks the topic, approves the copy, and hits send. Every piece is judgement work, so the human in the loop is not a limitation. It is the point.
02 ·
Workflow automation with n8n, Zapier or Make
An automated sequence that calls AI at one or more steps. A trigger fires: a new order, a product added to the range, a review landing. AI does a defined job: drafts the personalised email, logs the sentiment. An action follows: it sends through Klaviyo, updates the report. Once built, it runs on its own.
Runs for: $50 to $300 a month. That covers the platform plus API usage, which is usually modest at this volume but varies with the model and how much information each task processes.
Builds for: here the honest answer is a range, and it depends entirely on the stakes.
Build it yourself 4 to 20 hours for a simple automation. Expect the first version to break a few times before it settles. | Have it built properly $5,000 to $20,000, depending on how many systems it touches and what happens if it gets something wrong. |
That money does not buy software. It buys the discovery to pick the right task, prompt design so the output sounds like you, integration with your stack, edge-case testing, error handling, and making it work inside your security and approval processes. At enterprise scale, that last part is most of the work.
The dividing line is not the tool. It is the stakes. If it breaks and only you notice, build it yourself. If it feeds work your team relies on, touches multiple systems, handles real data, or publishes anything without a human in between, build it properly.
Maintenance: 1 to 2 hours a month reviewing runs, repairing connections when a platform changes its API, adjusting prompts as needs shift. Someone has to own it.
Right for you if: you have a repeatable, high-volume task with a predictable trigger, where only the inputs change.
In practice: an AI visibility pipeline. It runs the questions customers actually ask in a category through search and through the AI assistants, checks whether the brand appears in the answer and what gets cited instead, then turns the gaps into briefs and on-brand drafts ready for review. It runs to a schedule, not to a person remembering. And it is not an agent. It checks, compares, and drafts. It does not decide anything and it does not publish.
03 ·
A custom AI agent, built on the API
An autonomous system that decides and acts across your tools with minimal prompting. Built directly on Claude’s or OpenAI’s API. This is the “works while you sleep” agent people picture when they say the word.
Runs for: it varies with the build. API and hosting costs track the volume it handles and the number of systems it touches, so anywhere from a couple of hundred dollars a month to several thousand. A medium-complexity agent handling 10,000 tasks a month might spend $150 to $500 in API fees alone, before hosting.
Builds for: $5,000 to $30,000 and up in developer time, across 4 to 12 weeks. Then budget 10 to 20% of the build cost each year for maintenance, because models update, APIs change, and errors need watching.
Right for you if: the task is too complex and variable for a workflow tool, and the return is obvious. Most retailers under $10M in revenue are not here yet, and that is completely fine.
In practice: a retailer handling 8,000 “where’s my order” and returns tickets a month. The agent reads the ticket, pulls the order, checks carrier tracking, applies the returns policy, and decides: respond, refund, replace, or escalate. Refunds above a threshold go to a human. Build cost sits around $20,000 to $30,000. Set against two or three full-time roles on routine tickets, the return shows up within months.
Where To Start
Start small. Earn your way up.
My honest recommendation: start at Option 1. Move to Option 2 when a specific, high-volume, repeatable task earns it. Only reach Option 3 once you have exhausted the first two and the return is clear.
And notice what runs underneath all three: the quality of your information. Your product data, your tone of voice, your policies. Get that right and the tooling choice gets easier. Skip it and no build budget saves you.
Do not start with the agent. Start with the task, and pick the smallest thing that solves it.
Not sure which option fits your use case? That is the conversation we have with retailers most weeks. Tell us the task, and we will tell you which tier makes sense and give you a realistic read on cost, before you spend anything.
