Implementing Claude in a small business: the first 30 days

Most AI rollouts I see fail the same way. Somebody buys seats for the whole office, sends a link, and three weeks later two people use it and nobody can say what it did. The tool was fine. The rollout had no job attached to it.

This is the 30 day plan I use instead. It is deliberately small. One job, one workflow, measured, and THEN you decide whether to expand.

Week 1: pick one job and collect examples

Pick a job that happens a lot, takes real time, and has a clear “good” answer. Good first candidates:

• Sorting a shared inbox into the right buckets.
• Turning a messy intake form or phone note into a clean record.
• First drafts of the same kind of reply you send every week.
• Summarizing a long document into the five things a person needs to know.

Bad first candidates are anything with no clear right answer, or anything where a mistake goes straight to a client.

Then collect 20 to 30 real examples of that job from the last month. Include the weird ones. These become your test set, and they are the most useful thing you’ll make all month. Five spreadsheet jobs AI does well has more ideas if you’re stuck picking.

Week 2: set the rules before anyone types

Before the team starts pasting things in, write one page that answers:

• What data is allowed in, and what is NOT (client confidential material, health information, passwords, full card numbers).
• Which account to use. Business accounts, not personal ones, so the company controls the data and the settings.
• Who reviews AI output before it reaches a client. For now the answer should be “a person, every time.”

That page doesn’t need a lawyer to start, BUT check it against your client contracts, some of them already say what you can and cannot do. I wrote a fuller version of this in Writing an AI policy for a ten person company.

Same week, let people use the Claude app on their own work, inside those rules. You’ll learn fast which tasks they reach for it on. That list is your roadmap for month two.

Week 3: build the one workflow

Now automate the job you picked. The shape I use almost every time:

• The workflow reads the input (the email, the form, the document).
• Claude does one clearly defined thing with it and returns structured output, not a paragraph.
• The output lands somewhere a person already looks: a draft in their mail, a row in their sheet, a card in their task list.
• A person approves it. Nothing goes to a client on its own.

Run your test set through it before it touches live data. Write down what it got wrong. Fix the instructions, not the examples, and run it again.

Pick the smallest model that passes your test set. I explain how I decide in Route AI work by how much judgment it needs.

Week 4: measure, then decide

Measure two things:

• Time. How long did the job take before, and how long does it take now, including the review?
• Corrections. How often did the person reviewing have to change the output, and how much?

If the time saving is real and corrections are dropping, keep it and pick the next job. If people are rewriting most of the output, the job was a bad fit or the instructions need work. Both are fine outcomes. Both are cheaper to learn in 30 days than in six months.

What I’d skip in month one

• Building a custom chatbot for your website. It is the most visible project and usually the least useful first one.
• Connecting the AI to every system you own. Each connection is a new thing that can read your data, so add them one at a time, when a job needs it.
• Agonizing over the model. Start in the middle, test, adjust.

After 30 days

You’ll have one workflow that saves real time, a test set you can rerun whenever a new model ships, a one-page set of rules, and a list of what your team actually wants help with. That is a much better place to spend money from than a seat license and a hope.

If you’d like help picking the first job or building it, give me a buzz. Please let me know if you have any questions, thanks!