Skip to content
Request a consultation

Studio note

What AI automation costs a small business

What drives the cost of a small-business AI pilot: tools, exceptions, whether it only drafts, and who must still approve an action.

People ask what AI automation costs as if it were a single product. It is not. A page that answers order status from a shop is a different piece of work from an assistant that drafts replies, and both are different from a system that is allowed to change records. The weight of the work moves with the number of tools, the number of exceptions, and how much a person still has to approve. A flat price on a homepage usually hides that. Nexen does not publish a rate card, a starting price, or a figure you could treat as a quote. This note will not either.

You were not given a number here on purpose. A number without a named queue is a guess, and a guess printed as a price is a claim. Nothing in this note is an invoice, a market survey, or a description of what any business paid. This page is not reporting a range drawn from delivered work, because that would be a claim about jobs the page is not documenting. Use it to see what makes one proposal heavier than another, so you can compare documents that use different words for the same work.

Three kinds of work

A discovery maps one queue, the tools it touches, and the places a person must still say yes. The output is a map and a recommendation, not software. It is lighter when someone has already written the process down and the exceptions are listed. It is heavier when several teams disagree about what done means, when the work has to be watched rather than described, or when the source of truth is a pile of old threads. If a vendor calls that step free and the map only exists inside a sales deck, you did not receive a map you can keep. Ask who owns the notes when the meeting ends. A map you cannot take away is a conversation, not a deliverable.

A single-process pilot reads one system and either drafts or sends a narrow class of replies. The weight is not the model. It is whether the tool has an interface a program can use, or only a file someone exports at the end of the week. It is whether the rules fit on a page, or whether every tenth message is a special case. It is whether a person still handles everything outside that page. A reply that returns tracking when the email on the order matches the email in the chat is a smaller job than a reply that also edits an address, moves money, and answers in more than one language. Those are not decorations on the same job. Each one changes what the system is allowed to do, and what a person must still check before anything leaves the building.

A system that touches several tools, writes back to them, and has real approval paths is a different undertaking again. It gets heavier when the records disagree with each other, when the process crosses finance and support, or when the second tool has no interface and a person has to sit in the middle. Care after launch is a separate decision, not a footnote. Time reserved for fixes and rule changes is not the same purchase as time that quietly adds new workflows. Ask which one you are being offered. If the answer is that both will be sorted out later, you do not yet have a scope. Later is where an unfinished boundary becomes an argument.

What actually moves the cost

  • How many systems must be read, and whether they have an interface or only an export.
  • Whether the assistant may send, or only draft for a person.
  • How many exceptions are blocked, and who approves them.
  • Languages, channels, and whether the channel is already a business account you control.
  • How clean the source is. A help center that contradicts itself takes more effort than a newer model.
  • Whether you need a log, a permission model, and a way to delete a person’s data.
  • Who tests the failures, and whether that time is written into the proposal or assumed to be free.

Usage fees for a model are rarely the whole bill on a small queue. A short lookup adds little usage. A pilot that summarizes long documents uses more, and you should ask for a cap so a busy week cannot surprise you. The heavy part is the unglamorous work: matching fields, writing the sentence the assistant must use when it refuses, connecting the handoff, and reading real items until the refusal holds. If a proposal is mostly about tokens, ask who is paid to read the failures. If the answer is that you will do it after the studio leaves, the proposal is unfinished even when the headline looks light. Testing is not a courtesy. It is how you find the message the rules did not imagine.

Two sketches, not invoices

The sketches in this section are illustrations of scope. They are not customers, not deployments, and not amounts anyone has been charged. They are here so two requests that sound alike can be told apart before anyone names a figure.

Picture a shop that wants a chat reply to read the store and return tracking when the email matches. One channel. No refunds. A person gets everything else. The job is narrow: one lookup, one match, one handoff. The rules can be written in advance. The assistant does not change the order. When the email does not match, or the person asks for something else, the message goes to a person with the original words attached.

Picture a second shop that starts with the same sentence and then adds address edits, partial refunds under a limit the shop sets, and a summary posted into a warehouse tool that has no interface. That is not the first job plus a feature. It is a different permission, a different risk, and a different kind of connection. Someone has to decide what the limit means, who can raise it, and what the log must show after money moves. A vendor who treats the two requests as the same piece of work has not looked at what the assistant is allowed to change. The second request can be a responsible project. It cannot be described with the same sentence as the first.

A clinic sketch makes the same distinction, and it is also only a sketch, not a clinic Nexen is describing and not medical advice. Hours and appointment reminders drawn from a calendar are a narrow job. The words can be approved in advance, and the assistant is not deciding whether someone is well. Adding symptom triage is not a small extra. It is a scope to refuse, or to accept only when a clinician designs the stop and when the clinic has advice this note is not offering. The design that will not answer still has to be built. Someone writes the stop list, the fixed sentence, and the alert to the desk. That refusal is part of the product. Leaving it out does not make the project lighter. It makes the project unsafe, which is a larger problem than a missing feature.

How to compare proposals

Line the proposals up on the same queue. Which statuses are in? Which messages are blocked? What does the handoff contain: the original words, the record it read, and the action it did not take? What do you own when the project ends: the prompts, the workflow, the accounts, the log? What is usage, what is labor, and what happens after the first release, when the first unexpected message arrives? A proposal that looks lighter because it excludes the handoff is not lighter. It is unfinished. A proposal that looks heavier because it includes time to read real messages may be the one that named the work.

Ask for the assumption that would make the proposal wrong. The help articles agree with each other. An administrator can connect the store. The assistant does not send refunds. The channel is already a business account, not a personal one. If the assumption is false, the change should be visible before you agree, not after a demo. Nexen would rather decline a project than publish a price that only works if the work is simpler than it is. That is a rule for this website. It is not a story about a past job, and it is not a promise that a form on this site has already produced a number.

What this page will not claim

This page will not say that businesses save a stated share of their time. It will not say what anyone pays. It will not offer a starting price that is really the price of a conversation and not the pilot. It will not pretend a subscription to a chat tool is the same thing as a workflow with a stop and a log. A real number comes after a person has seen the queue. The sample plan on this site does not collect a project fee and is not a quote. A consultation is how a person looks at the work with you. The consultation page says the same thing in operational language: the form does not book a meeting, and it does not promise how quickly anyone will reply.

If you are setting a budget this year, pick one queue and write the drivers beside it. How many tools. Whether the assistant drafts or acts. How many exceptions a person must see. Whether the source of truth is clean or contradictory. Whether you want someone to change the rules after launch, and whether that time is included or treated as a new request. That list will tell you more than a single round figure with no scope under it. When you ask a studio for a number, ask them to point at the same list and say which lines they included. If they cannot, you are not looking at a price. You are looking at a slogan.

Written by Nexen AI Solution. Published 14 October 2026.