Your Newest Hire Started Today (And It Doesn't Need a Desk)
Notes from Anthropic’s Small Business Workshop in San Jose, and what Claude Cowork can actually do for a business like yours
I spent yesterday afternoon at PayPal’s San Jose headquarters at a workshop Anthropic and Tenex put on for small business owners. I went in as someone who already builds AI workflows for a living, expecting a polished demo and a goodie bag. What I got instead was the clearest, least hyped explanation I’ve heard of how an owner with no technical background can hand real work to AI and trust the result. Here’s what I took away, because most of it has nothing to do with being a developer and everything to do with how you already run your shop.
A quick note on why this workshop existed at all. Anthropic’s head of small business, Lena Achman, made the case that software almost always gets built for startups and big enterprises first, and small businesses get it last, watered down. The premise here was the opposite: what if small and medium businesses got the good tools first? She pointed out that the overwhelming majority of businesses in San Jose are small or medium, not venture-backed startups, and that the AI gap for owners is real but closeable. That framing stuck with me, because it matches what I see in my own consulting work. The owners who win with this stuff treat it like a management problem, not a software problem.
The one idea worth taking home
AI is a brilliant new hire on their first day.
That was the anchor for the whole session, and it’s right. The model is well-read and tireless, fast at almost anything you can describe in plain language. But it’s day one. It doesn’t know your customers, your pricing, your tone, how you sign your emails, or what “good” looks like at your shop. Everything that’s productive about working with AI and everything that’s maddening about it comes from that one gap: high capability, zero context. What you get out is almost entirely determined by what you put in.
The good news, and this is the part that should relax you a little, is that you already have the core skill. You’ve onboarded a new employee before. You know how to figure out which jobs to hand them, how to train them on your business, how to check their work until you trust it, and when to finally let them run. That is the entire skill set. Everyone who came in thinking this required a developer background left having built something real, just by getting better at delegating.
What Cowork actually is
There are three ways to work with Claude. Chat is the thinking partner: brainstorming, drafting, research, talking things through. Code is for people writing software, and it’s not what this post is about.
Cowork is the doing partner. It connects to your actual tools and runs multi-step jobs on your behalf, including several at once in the background while you do something else. In the live demo, the facilitator kicked off two jobs side by side: one chasing overdue invoices across QuickBooks and PayPal, another triaging open leads out of HubSpot and ranking them by urgency. Both produced drafted, ready-to-send emails written in her voice in about five minutes. She was assigning work, not typing commands.
Anthropic also released a free “Claude for Small Business” add-on for Cowork, with a set of prebuilt workflows aimed squarely at back-office grind: a Monday morning brief with your top priorities, a cash-flow snapshot and 90-day forecast, open-invoice chasing, a month-end close package for your accountant, lead triage and scoring, a customer pulse check pulled from reviews and tickets. You install it and start adapting the pieces to your business. Think of them as a competent first draft, not a finished product.
The four steps that make an automation actually work
The instinct everyone has is to open the chat and type “automate my invoicing for me.” It rarely works, because you’ve handed your new hire a one-line job description and no context. The repeatable process they taught instead has four steps.
Before you touch any AI tool, write the task out by hand. Open a doc or a piece of paper and list every single step you take to do the thing. Open this email, look at these fields, check this total, then send this Slack. The facilitator described automating her own order-entry process after she’d keyed in hundreds of orders by hand and given herself carpal tunnel. The first move wasn’t a prompt. It was writing down exactly what she did. The nuances you’d never think to mention are usually the ones that make the difference between right and wrong.
The second question is what the AI does and what you keep. For her order process, she had the AI do everything up to the submit button, then checked it and pressed submit herself. Later, once she trusted it, she handed off that step too. You don’t have to delegate the whole thing at once.
Once that’s settled, write the brief and connect the tools. Spell out the output, the audience, the tone, what good looks like, and what it should never do. Then connect the relevant tools so it has real context instead of guessing. Giving it a role helps here: “you are a sharp accounts-receivable specialist” pulls better work out of the model than no framing at all.
The step most people skip is verifying against your own history before going live. Before you let an automation loose on real work, run it against your past. Point it at your last 20 orders, your last 20 leads, last quarter’s numbers. You already know the right answers, so you can catch where it’s wrong and tune it in a low-stakes way. By the time you flip it on for real, you’ve already watched it work.
What to hand off, and what to keep your hands on
Not everything should be delegated, and the workshop was refreshingly blunt about this.
Hand off the standardized, low-judgment work: routine email drafts and data cleanup. If the first try isn’t perfect, the cost of fixing it is small.
Keep the irreversible decisions for yourself: final pricing, contracts, anything legal, hiring and firing, numbers you’d certify to a bank or regulator, anything that goes to a customer unread. The tool can get you 95% of the way there, but if a mistake can’t be undone, you take the last look. For the multi-step work in between, anything customer-facing or with real numbers in it, stay in the loop until you trust it.
The test they offered for anything you’re unsure about is a good one: what’s the cost if it’s wrong, and who catches it before it matters? If the answer is “high” and “nobody,” that one stays under your control.
And the line I keep coming back to: don’t ship anything you wouldn’t stand behind. Your customer cares that you sent it, not who drafted it first.
The stories that made it click
The feature list is fine, but the room came alive during the real examples, and these are the ones I’d want a fellow owner to hear.
A whole-animal butcher shop in Brooklyn was constantly out on the road and had no real process for paying invoices or forecasting cash. Some weeks they paid too much, some weeks they were late. They built a workflow that pulls their invoices, computes their cash position, and recommends what to pay each Monday. The interesting part is what the AI got wrong at first: it logically recommended paying the biggest invoices first. But in their business, the small farms are the relationships that can’t survive a late payment, while the big, established suppliers can wait a few days. So they corrected the logic, told it to prioritize the small farms and draft a friendly heads-up email to the larger ones when payment would run late. Now it runs every Monday and the owner just decides.
A medical device supplier in Chicago automated the afternoon ritual where a sales manager downloaded order PDFs, copied the details into each supplier’s particular template, and emailed them out, fifteen to twenty times a day. They had the AI handle their established suppliers and flag any brand-new supplier whose template wasn’t standardized yet, so a human could step in. Two things that CEO said stuck with me. First, don’t try to automate everything at once; pick one workflow and make it genuinely good. Second, build it alongside the employee who does the job today, so they’re bought in and they catch the things you’d miss.
And my favorite, for how small and human it was: a woman who runs a music lesson studio with five teachers. Her teachers kept forgetting to mark lessons as completed or canceled. She set up an automation that checks her scheduling software at 7am and 3:30pm and nudges the right teacher in Slack about any unlabeled lessons. The teacher fixes it or just tells Claude what happened. As she put it, twice a day, a fifteen-minute task she never has to do again.
None of these owners were engineers. Each one picked a specific task they were tired of doing. They handed the execution to AI and stayed in the loop for anything requiring a judgment call.
The honest part: is this safe?
The first question in the room, every stop on this tour apparently, is some version of “I hook up my Gmail, my PayPal, my bank, and let it loose, and then one morning my account is empty.” It’s the right question to ask, and I’d be skeptical of anyone who waved it off.
A few things genuinely matter here. Your business data isn’t used to train the models when you connect a tool like QuickBooks, full stop. On the prompts you type, training is off by default on the Team plan, and on individual paid plans you can turn it off. You own what you create. Anthropic encrypts your data and walls it off from other organizations; the financial and healthcare institutions that trust the platform aren’t doing so casually.
The part most relevant to the empty-bank-account fear is permissions, and this is where you stay in control. Every connected tool has separate permission levels. You can leave read-only actions on automatic, so it can look at your data freely, while requiring your explicit approval before it ever writes or sends or deletes anything. It cannot fire off an email or move money without you saying yes, unless you’ve specifically told it to. You’re the manager. You decide what your new hire is allowed to touch.
What the room actually built
The session closed with five owners coming up to show what they’d made, and that fifteen minutes sold the point better than any slide. The line they kept repeating, almost word for word, was “I’m not technical.”
A stagehand who runs a software company on the side put it best. He’s not a coder, he said; he just gives clear directions and manages the AI the way he manages his fifty-person crews on a production floor. He’d built an online store for an artist friend that runs itself, taking payment through Stripe and firing off the confirmation emails and social posts on its own. He figures his solo work over the last three months is worth seven figures to his two companies. Take the number with a grain of salt; the pattern underneath it is the real lesson.
Another owner, who runs a creative consultancy, made a point worth stealing. The free small-business toolkit mostly assumes a product business, and hers is a service business, so she had Claude audit the whole set and rework what fit her model. The prebuilt skills are a floor, not a ceiling. You adapt them to your shop.
A third had spent about two hours building a small app so his brother-in-law’s wife could stop hand-typing fuel receipts into a spreadsheet at tax time. Snap a photo, and it pulls the date, state, gallons, and total into a table she can export. “I’m not a developer,” he said, and he’d built three of these.
Hardware monitors and design systems, a four-year-old’s birthday-gift tracker so the relatives stop buying duplicates: all different businesses running on the same underlying discipline. Each owner started by mapping the task carefully, then handed the execution to AI and protected the decisions only they should make.
What I built in the room
During the open-build portion I worked on something from my own consulting practice.
I run KCC Consulting, and a chunk of my work is auditing small businesses’ Google Ads, the kind of account where money leaks every month and nobody has time to find it. I’d been doing it by hand. In the workshop I built it into a Cowork workflow.
The design idea is simple and it’s the same discipline the whole workshop was preaching. Instead of asking the AI to “audit these ads” and trusting whatever it tells me, I split the job in two. One step researches what good Google Ads practice actually looks like and writes the rules down with real sources. A second step takes those written rules and checks the client’s actual data against them, line by line, citing the specific numbers. The AI can’t invent a “best practice” to justify a finding, because every finding has to trace back to a sourced rule. It also runs a real script to count the data rather than eyeballing sixteen hundred rows, which is exactly where AI tends to make things up.
I ran it against a real client, a charter fishing business in Ohio, on their last 30 days. The result: 94.9% of their ad spend, about $481 out of $507, produced zero conversions. More than two hundred search terms were burning money with nothing to show for it. The number came straight from the audit, not a spreadsheet I’d tortured for an hour. It’s the kind of thing a five-minute automated audit surfaces that a busy owner would never have time to dig out by hand.
I built that in a workshop session. That’s the part I want you to take seriously.
Where this leaves you
The technology changes every few weeks. The mental model underneath it is more stable, and that’s what’s worth holding onto. Treat AI like a capable new hire, write the job down, hand off what you’d give any good assistant, and check the work against your own history before you trust it. Start with one workflow, the most annoying recurring task you can think of, and make that one genuinely good before you build the next.
If you’re reading this as an owner thinking “this sounds useful and I still don’t have the hours to build it myself,” that’s fair, and it’s exactly the gap I work in. Whether you build it yourself or bring someone in to build it with you, the thing I’d push back on is doing nothing because it feels technical. It’s delegation with a tireless, context-hungry assistant. You’ve been doing the delegation part for years.
I’m Kevin Cress. I build practical AI workflows for small businesses through KCC Consulting. More at kevincress.com.