Anthropic just shipped the most ambitious small business AI product yet

Claude for Small Business. Anthropic shipped it last week. And if you run a business without an IT department, this is the closest anyone has come to handing you a real lever instead of a chat box. It deserves credit before it deserves critique.

Here's what's in the box. Fifteen ready-made recipes: weekly business brief, invoice chasing, payroll planning, contract review, month-end close, cash-flow monitoring, lead triage, marketing campaign drafting — the recurring jobs every small business owner does and most of them hate. Fifteen skills layered on top — each one a recipe Claude follows the same way every time. Ways to plug Claude into QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365 — the office tools a typical small business already uses. You install the add-on in Cowork, connect the tools you already use, and Claude starts doing the work.

What it does well is the office work. Pulling cash position from QuickBooks. Drafting polite reminder emails to customers whose invoices are 30 days late. Reviewing a vendor contract for red flags. Building a Monday-morning brief that combines calendar + email + accounting. None of these are demos. They're real work that real owners do every week, and Claude does most of them in minutes instead of hours.

But the office work is also the limit of what comes in the box. The list of tools Claude can reach into is built for desk-job small businesses, not for restaurants. Anthropic's own launch announcement is honest about who it's for — owners whose work happens in the same handful of office apps Anthropic has already wired up.

Monday 6:30am. A brewery owner in Bend, Oregon opens her laptop. Claude has already drafted polite invoice-reminder emails to the three restaurants that haven't paid this month, pulled her cash position from QuickBooks, flagged that next Thursday's tap-takeover event needs a vendor confirmation, and queued up a contract from her hop supplier with three highlighted clauses worth re-reading. She approves the invoice emails, sends them, and is at the brewhouse by 7:15. Claude saved her ninety minutes. It's a real product.

That brewery scene is what most of the launch coverage is describing. It works. Not as a future promise — as a thing that runs today.

What changes when you operate five-plus stores

At one location, your sales might live in Toast and your books in QuickBooks. Claude reaches both. The answer it gives is real.

At twelve locations, the picture shifts. Your sales aren't in one place — they're in Toast or Par or Xenial or NBO, often a mix across brands if you operate more than one concept. Your labor isn't in Google Workspace; it's in Crunchtime or Restaurant365 or a fragmented set of spreadsheets that your district managers maintain by hand. Your inventory isn't anywhere clean. And your delivery is split across DoorDash, Uber Eats, and whatever native channel you've stood up. Anthropic's QuickBooks hookup tells you what's in QuickBooks. But it cannot tell you your food cost percentage, because food cost lives across five systems and isn't added up anywhere a single hookup can reach.

Most of the "AI for restaurants" pitches in 2026 fall apart right here. The "AI-powered restaurant operations" demo at NRA last spring. The "AI for franchisees" webinar three vendors ran last quarter. The "AI agent for QSR" booth that turned out to be an empty chat sidebar bolted onto a 2019 dashboard. None of them survived this question: which of my stores had the worst labor variance last Sunday, and why? Because the AI couldn't see across systems to answer.

The vendor playbook has been the same since "data-driven" landed in 2015. Pick the loudest word of the year. Slap it on the same software. Charge thirty percent more. Hope the buyer doesn't notice. What's different in 2026 is the AI is real — Claude and ChatGPT are not the 2018 chatbots — but the "connected restaurant" pitch the AI gets dressed up in is still the same fragmented mess of systems that don't talk to each other.

The "agent-readable" gap nobody on launch day is talking about

Agent-readable data. The term shows up exactly nowhere in the launch coverage. That's the tell.

Think about mise en place. A line cook standing at his station has his diced onions in one container, his minced garlic in another, his stock in a quart on the rail, his salt in the ramekin to the left, his oil within reach to the right. Everything labeled. Everything where it's supposed to be. Everything ready to grab without looking. He can cook fast because the prep is done.

Agent-readable data is mise en place for an AI. Your numbers, labeled, organized, in containers Claude can grab from without asking. When the prep is done, Claude cooks fast. When the prep isn't done — when your sales are in a JSON file with no labels, your store codes don't match across systems, your dayparts mean different things in different places — Claude is the cook who walked into a kitchen with no prep. He'll make you something. He'll do it with confidence. It won't be what you ordered.

For a restaurant operator, your data is sitting in the walk-in. None of it is on the rail. Store codes that don't match across systems — Toast calls it loc_id, your accounting system calls it restaurant_code, your delivery aggregator calls it unit, and not one of them maps to the others without translation. Dayparts defined differently per brand. Labor calculated three different ways across vendors. Refunds posted as positive numbers in one system and negative in another. The numbers exist. But nothing is labeled, nothing is portioned, and nothing is where the cook can reach for it.

AI on dirty data gives confident, wrong answers. Connecting Claude is the easy part. And making the answers trustworthy is the work.

Write down the five questions

While you're waiting for the data work to catch up, do this. Write down the five operational questions you'd actually want to ask Claude about your business. Not "summarize my sales" — the real ones. Which stores are over labor target this week? Which menu items lost margin month-over-month? Which delivery channel is bleeding? Which GMs are running their P&L cleanest? Which item is being 86'd most often across the portfolio? Write them in the sentence form you'd type into Claude if you could. Save the list as a Note or a doc.

That list is the spec for the data work that has to come next. It tells you which systems have to be prepped and labeled before any AI tool — Claude, ChatGPT, or whatever the next one is — can give you a trustworthy answer. The list is also a vendor screen: when someone pitches you "AI for restaurants" next quarter, ask them to answer your five questions using their product on your data, live. The ones who can't are not selling you AI.

On the partner question

Doing the mise en place on your operational data — connecting it to the systems where it sits, cleaning it, labeling it, putting it on the rail so the AI can reach for it without asking — is what platforms like Expo exist to do. There are others. The point is the work is real, and it's not the work Anthropic is going to do for you, because it's not their job. Anthropic is building the brain. The data has to come from somewhere. Operators who treat that as a partner question rather than a competitor question move fastest.

Frequently Asked Questions

Should I sign up for Claude for Small Business if I run a multi-unit restaurant?

Yes. The office side pays for itself in the first month — investor updates, contract review, email triage, marketing copy, vendor reminders. The operational-data side is a separate question. Sign up for the first, plan for the second.

Will Claude's QuickBooks connector tell me my food cost percentage?

Not accurately, not at multi-unit scale. Food cost is calculated across your POS (what you sold), your invoicing and inventory systems (what you bought and what you used), and your accounting system (what cleared). QuickBooks has the invoices and the cleared payments. It does not have what your stores actually used. Until those three sources are reconciled in one place, any food cost number Claude gives you is built on partial data — it'll be confident, and it'll be wrong.

What does "agent-readable data" actually mean?

Agent-readable means your data is prepped the way a kitchen is prepped before service — labeled, portioned, sitting on the rail where an AI can grab from it. The opposite is a printed PDF of last week's sales report. A human can read the PDF. The AI cannot reach in and use it. Agent-readable is the version the AI can actually work from.

Can I just dump my POS export into Claude?

You can. Don't trust the answers yet. A raw POS export carries stale records, test transactions, store codes Claude has no way to match across systems, and dayparts defined however your POS happens to define them. You'll get a confident answer about a store that closed two weeks ago, or a labor variance for a shift that doesn't exist. The fix isn't more uploading — it's the prep work that has to happen between your POS and the AI.

Is Expo trying to replace Claude?

No. Claude is the brain. Expo, and platforms in the same category, are the data plumbing that makes the brain useful for a multi-unit operator. Different jobs. Don't pick one over the other — they're built to work together. The reason the partner-versus-competitor question keeps coming up is the AI launches keep getting pitched as if connecting to your tools is the whole problem. For a single-shop business on a standard set of office tools, connecting is the whole problem. For a multi-unit restaurant operator, the connecting is step one and the harder work — making the data agent-readable across systems — is what has to happen after.

Will Anthropic eventually do this for restaurants specifically?

Maybe. Industry-specific small business products are a natural next move for them. Until then, the data work has to happen somewhere, and it's separate from picking an AI tool. Doing the mise en place is what makes the AI-tool choice easy to swap out as better tools come along — whichever AI tool wins next year, the data work you do this year still pays off.