When is an AI agent worth it for a small business?

The word is used for three different things, and the confusion is expensive, because the three cost different amounts and fail in different ways. Anthropic draws the line by asking who decides the steps. In a workflow, the steps are orchestrated through predefined code paths that a person wrote in advance. In an agent, the model directs its own process and its own tool use, keeping control of how the task gets done.

That distinction is the whole decision. If you can describe the steps, you have a workflow, whatever anyone calls it in a sales conversation.

Most of what a small office wants sits in the middle two rows. That is worth knowing before anyone quotes for the bottom one.

When the number of steps genuinely cannot be known in advance. Anthropic recommends agents for open-ended problems where it is difficult or impossible to predict how many steps a task will take, and recommends against reaching for them before simpler approaches have been tried. Research across an unknown number of sources qualifies. Producing the same monthly report from the same four systems does not.

For splitting work across several agents rather than one, Anthropic names three conditions. Any one of them can justify the architecture; none of them is common in a ten-person office.

Anthropic is direct about the default: a well-designed single agent with the right tools accomplishes far more than many builders expect. The burden of proof sits with the more complicated option.

When it is bought as a category instead of scoped to a job. An agent that nobody can describe the success condition for will run, produce something, and leave a person to work out whether it was right, which is the cost the purchase was meant to remove.

The specific ways these systems fail are documented rather than hypothetical, and each one has a recognizable office version.

Three shapes of work are called out as poor candidates for splitting: sequential phases of the same job, division by type of work instead of by context boundary, and tightly coupled pieces that need constant synchronizing. A great deal of professional work is exactly that shape, which is why the answer for an office is so often one careful setup instead of several clever ones.

Four things, in this order, because each one is cheaper to run and easier to correct than the one after it. Most offices get what they wanted somewhere in the first three.

1. 01A better prompt against the real documents. Anthropic’s guidance is to start with simple prompts and improve them under evaluation before adding any machinery at all. This step is free and it is the one most often skipped.
2. 02A saved Skill. When the same task recurs with the same shape, a Skill captures the office’s format and conventions once so nobody restates them every time. Skills come with every paid Claude plan.
3. 03Claude Cowork with a person reviewing. Cowork already takes a whole multi-step task, works across files and connected tools, and shows each step for a person to check. For a great many jobs this is the agent the office was picturing.
4. 04A scheduled Cowork task. When the job is genuinely recurring and its steps are stable, running it on a schedule delivers the unattended part without any custom build.

What Cowork does and where it stops is set out in Claude Cowork, explained, and the catalog of jobs offices hand it first is in What small teams actually do with Claude Cowork. If step three does the job, the build was never needed, and that is a good outcome rather than a disappointing one.

Building past this point is developer work. Anthropic publishes an Agent SDK in Python and TypeScript for it, which is a useful thing to know before a proposal describes a custom agent as a configuration exercise.

More than the seat, and the multiple is published. Anthropic reports that agents use roughly four times the tokens of a chat interaction and that multi-agent systems use roughly fifteen times, with the overhead coming from duplicated context, coordination between agents, and summarizing results. Multi-agent implementations typically use three to ten times the tokens of a single-agent approach.

For a small office on seat pricing this shows up as usage limits arriving sooner rather than as a surprise invoice, which is the friendlier failure but still a real constraint. The seat arithmetic is in What Claude actually costs a small business.

The larger cost is the one that never appears on a bill. Somebody has to define what finished means, review output until trust is earned, and maintain the thing when a connected system changes. An agent nobody owns stops being used within a month, and the subscription carries on.

Four conditions, and they are the same four whether the thing running is a scheduled Cowork task or a custom build. An office that cannot answer all four is not ready for the unattended version, and the supervised version is available today.

• A checkable definition of done. If success cannot be described in a sentence, no run can be judged, and compounding errors have nowhere to be caught.
• Scoped access. The agent reaches the folder and the systems the job needs, and nothing else. Access decided by convenience is the decision that gets regretted.
• A named reviewer for each category of output, and a rule that nothing reaches a client, a court, or a counterparty without that review.
• A record of what ran. Enterprise plans carry audit logs and a Compliance API; below that tier, the review step is the record, which is a reason to keep it.

These belong in writing before the first unattended run, not after the first surprising one. Where each of them sits in a signed document is set out in An AI use policy for a small firm, and the connector permission modes that enforce the second one are covered in How to set up Claude Cowork.

By assessment, and never from a price list. Justo does not sell agents as a cold package, because the question of whether a business needs one cannot be answered before somebody has looked at where its hours actually go. The assessment produces the answer, and for a good number of offices the answer is that a Skill and a reviewed Cowork task cover the ground.

Where a build is justified, it is scoped and priced against that specific job, with the definition of done written before anything is built. A proposal that names an agent before it names the workflow has the order backwards.

The assessment is part of Justo’s flat-fee engagement: plan selection, secure setup, training on your real work, and a signed AI use policy, from $3,500 for a team of 5 to 25. Scoped builds are quoted separately, out of what the assessment finds. Scope is on the pricing section of our main page.