AI for routine office work just got about 75% cheaper

Two announcements this week point the same way. On October 7, Anthropic released Claude Haiku 5.5, the newest version of its smallest and fastest AI model. Anthropic says it costs around 75% less to run than Haiku 4.5 on average, and its prices are 90% lower for most requests. A day later, Google Cloud introduced its Gemini agent, which it calls “a universal agent for work” that “chooses the best model for the job” and has “built-in cost controls.”

Different companies, same message: routine work is getting very cheap to hand to AI. Here’s what that means if you run a small or midsize business.

What actually changed

Small models are the workhorses of business automation. They handle the boring, high-volume jobs: reading an email and deciding where it goes, pulling the details out of a form or a PDF, summarizing a call, drafting a first reply.

Three things moved this week:

• Price. Haiku 5.5 lists at $0.10 per million input tokens and $0.50 per million output tokens for requests under 100,000 tokens, down from $1 and $5 for Haiku 4.5. A token is a chunk of text, about the size of a short word.
• Skill at using software. On OSWorld, a standard test of operating a computer the way a person does, Anthropic reports that Haiku 5.5 scored 72.4%, up from 15.7% for Haiku 4.5. That’s the ability behind form filling and data entry between systems that don’t talk to each other.
• Control over cost. Haiku 5.5 is Anthropic’s first Haiku model with an adjustable effort setting, so it can think less on simple jobs and more on harder ones. Google is building the same idea into its agent: match the model to the task to “raise accuracy on the difficult jobs and lower cost on the simple ones,” as its announcement puts it.

Why it matters for a small business

By my rough math, having a small model read 2,000 customer emails a month and write a short summary of each now comes to less than a dollar a month in AI usage at Anthropic’s published prices. Your actual cost depends on how long the emails are and how it’s set up, but the order of magnitude is the point.

The AI is no longer the expensive part. The real costs are:

• Setup: connecting it to your inbox, your forms, your spreadsheets or your accounting system.
• Rules: deciding what “done right” looks like, which takes a few hours of your best person’s time.
• Oversight: someone checking the output, at least until you trust it.

That changes the question to ask. It’s no longer “can we afford AI?” It’s “which of our repetitive jobs is worth setting up, and who will check it?”

Where I’d start

The sweet spot is work that is high-volume, repetitive and low-stakes, where a person reads something and retypes it somewhere else. For example:

• Sorting a shared inbox and flagging the messages that need a fast answer
• Pulling job details from web forms and emails into your scheduling sheet
• Turning PDFs, like purchase orders or supplier invoices, into rows a person can review
• Summarizing calls or site notes into a consistent format

Here’s how that might look at a 12-person HVAC company. Service requests arrive by web form, email and voicemail transcript. A small model reads each one, pulls out the address, the equipment and the problem, adds a row to the dispatch sheet, and drafts a confirmation for the office manager to approve. The office manager stops retyping and starts reviewing. At this week’s prices, the AI usage for a few hundred requests a month is a rounding error. The setup is the real project.

Cheap doesn’t mean unsupervised

Small models still make mistakes, and Anthropic still points customers to its larger models for complex work like advanced coding. Google’s announcement put real emphasis on giving each agent its own identity, limited permissions and an audit trail of everything it does. That’s the right instinct at any size.

Keep a person in the approval step for anything a customer will see. Run a new automation in the background for a couple of weeks before you rely on it. And measure: how long each item took before, how long it takes now, and how often the AI gets something wrong.

The takeaway

If you’ve been waiting for AI to get cheap enough to try, for routine paperwork it is. Pick one repetitive job, set it up carefully, keep a person in charge, and measure the result.

Not sure which job to start with? Talk it through with the AI advisor on this site, or request a free 20-minute call.