In June, Anthropic released Claude Fable 5, the most capable AI model it has ever made available for general use. Along with it, the company published an official guide for the developers and businesses that build on it. Most business owners will never read that guide. That’s a shame, because tucked inside is the clearest picture yet of what AI is turning into: less like a search box, more like an employee. And if your office handles client data, it raises the question every regulated business is asking right now: how do you use AI safely?
We read it so you don’t have to. Here are the five lessons worth stealing, in plain English, each with a prompt you can copy this week, plus when a private AI instance is the right answer for a law firm, medical practice, or financial office.
The Big Shift: AI That Finishes the Job
The headline change isn’t speed. It’s independence. Earlier AI tools answered questions and drafted paragraphs. This generation can take on a full project, split it into pieces, hand parts off to helper copies of itself, check its own work, and keep going for hours or even days without a person steering every step.
Anthropic’s advice to the companies using it says a lot: don’t judge the model on easy work. The teams getting the most out of it hand it their hardest unsolved problems, because simple tasks hide what it can actually do.
Think about your own backlog for a second. The value isn’t AI writing a slightly nicer email. It’s AI taking on the project that has been sitting on your list since March.
Lesson 1: Brief It Like a Sharp New Hire
For years, the standard advice was to write long, detailed AI instructions that spelled out every rule and every edge case. The new guide flips that. The model follows direction well enough that a short, clear request usually beats a page of rules. Anthropic even warns that instruction libraries written for older models can now make results worse, and recommends trimming them.
If someone at your firm built an elaborate set of AI prompts last year, don’t throw it out, but do revisit it. Say what you want the way you’d brief a capable new employee, then let it work.
Prompt to steal
Lesson 2: Explain Why, Not Just What
One line in the guide could have come from any management book: the model does better work when it understands the reason behind a request, who the result is for, and what it needs to accomplish.
Sound familiar? It’s the same brief you’d give a person. “Clean up this spreadsheet” gets you a cleaner spreadsheet. “The bank needs this for our loan renewal on Friday” gets you the right spreadsheet.
Prompt to steal
Lesson 3: Set Boundaries Before You Delegate
The guide is honest about the flip side of independence: an AI that can act on its own will occasionally act when you didn’t ask, like drafting a message nobody requested. Anthropic’s fix is simple and worth copying. Tell it, in writing, what it should do, what it should never do, and when it must stop and ask a human first.
For a law firm, medical practice, or financial firm, this is the section that matters most. Before your team uses AI that can take real actions, write the rules down: what information it may touch, what it may never touch, and which decisions always need a person. You already have this discipline for new employees who handle client information. AI deserves the same handbook.
Prompt to steal
Lesson 4: Don’t Accept “Done” Without Proof
Our favorite part of the guide: Anthropic tells builders to require the model to point to real evidence before reporting a task as finished, and to say plainly when something isn’t verified yet. In the company’s own testing, that one rule nearly wiped out cases of the AI claiming progress it hadn’t made.
Treat that as house policy. AI can be a confident intern: fast, tireless, and occasionally sure of things it shouldn’t be. Ask for the receipt. If it says the report is reconciled, look at the numbers. For a regulated business this was never optional anyway. Your compliance obligations don’t go away because software did the work.
Prompt to steal
Lesson 5: Give It a Memory
The guide recommends giving the model a simple notes file where it records lessons as it works: what worked, what failed, what got corrected. The next time it starts a job, it reads its own notes and starts smarter.
That is a training file, and it’s where firms will quietly pull ahead of each other. A business that treats AI like a disposable gadget starts from zero every morning. A business that builds a small “how we do things here” file gets a little more out of it every week.
Prompt to steal
Private AI: The Part the Manual Doesn’t Cover
Everything above assumes one thing Anthropic’s guide takes for granted: the AI is set up safely for your kind of business. An assistant that works independently with your files is only as trustworthy as the guardrails around your client data.
Free consumer chatbots are the wrong place for that data. What your staff types into them can be used to improve someone else’s model, and the fine print was never written with your confidentiality duties in mind.
For firms with real confidentiality obligations, the cleaner answer is a private AI instance. That’s your own AI workspace, set up so client files, case notes, and patient or financial records stay inside an environment your firm controls. Conversations aren’t used to train anyone’s public model, activity is logged so you can answer an auditor, and the whole thing sits behind the same access rules as the rest of your systems. It’s the difference between staff quietly pasting client data into whatever free tool they found, and giving your team one approved place to use AI with confidence. We build these for compliance-heavy firms in Morris County and northern New Jersey, and it’s usually the step that turns “our people are secretly using AI” into “our firm uses AI, on our terms.”
Round it out with a short written AI policy your staff has actually read. For law firms, medical practices, and financial firms, none of this is extra credit. It’s the entry fee.
The manual for one of the world’s most capable AIs reads a lot like a job description. So manage it like your sharpest new hire: clear briefs, real context, firm boundaries, and proof before you sign off. The firms that learn to manage AI well will get years of compounding value from it. The ones that treat it like a toy will keep wondering what the fuss is about.
Common Questions About AI and Client Data
Is ChatGPT safe for client data?
Not on a free personal account. Consumer chatbots can use what you type to improve their models, and nothing in the fine print is built around your confidentiality duties. The safe habit is simple: keep client information out of public AI tools entirely. And when client data leaks, the case that follows is about whether you protected it.
What is a private AI instance?
Your own AI workspace, running in an environment your firm controls. Client files, case notes, and patient or financial records stay inside your walls, conversations never train a public model, and activity is logged so you can answer an auditor. Think of it as the difference between renting a desk in a shared workspace and having your own locked office.
Does a small business need an AI policy?
If your staff touch client information, yes. One page is enough: which tools are approved, what may and may not be typed into them, and which decisions always need human review. Write it before the habit forms, not after a problem does.