The Claude-Ready Practice: Build Your AI Operations Assistant

Most practices are using AI the way they used a search box. Open a tab, ask a question, copy the answer, close the tab. Useful, but it does not run anything. Nothing happens unless you are sitting there typing.

An AI operations assistant is different. People are starting to call this an agent, and the word is fine to use. What it means in plain terms: an assistant you set up once, connect to the tools your practice already uses, and then leave running. It reads your inbox, drafts your replies, builds your meeting agendas, briefs you on your numbers every week, and assembles your operating documents from material you already have.

This guide is the full build, inside Claude Cowork. By the end you will have an assistant that briefs you every morning before you sit down, drafts the email you were dreading, prepares the agenda for your afternoon, sends you a weekly sales and marketing summary that compares against your goals, recaps your meetings into action items, and on the operations side, turns the protocols you have already written into a structured onboarding plan.

The work splits into two halves. First you do the input work once: connect the tools, set the rules, teach it your voice. Then it scales. Every task after that takes minutes to set up and runs as long as you want it to.

1. What you are actually building

Two jobs, one assistant.

The personal assistant side handles the work that eats your mornings and evenings. Email triage. Drafted replies in your voice. A prepared agenda before every meeting. A recap after it with the action items pulled out. A weekly summary of what happened and what is still owed.

The operations side handles the back office. It briefs you on weekly sales, weekly marketing and ad spend, cost of goods, and anything else you track, measured against your goals. It reads the SOPs you already wrote and assembles a 90-day onboarding plan. It keeps the documents that usually live in one person’s head written down and current.

Both run on the same setup. You build the foundation once. After that, adding a new job is a single prompt.

2. What Cowork is, and what it cannot do

Before you build anything, it helps to know exactly what this tool is, so you trust the output and you are not surprised by its limits.

Claude has two agentic tools. You want Cowork, not Claude Code. Cowork is the desktop app built for knowledge work. It runs the same engine that powers Claude Code, the developer version, but you never touch a terminal and you never write a line of code. Claude Code is for engineers. Cowork is for owners.

What it is. An AI worker that lives inside the Claude desktop app on your computer. Instead of answering one question at a time, it takes on a multi-step job, does the work across the tools you have connected, and comes back with a finished result: a briefing, a drafted email, an organized folder, a built document.

What it can do.

Read and work across the tools you connect to it: your email, your calendar, your files, your team messaging.
Read files saved on your own computer, because it runs on your machine, not only in the cloud.
Draft, analyze, summarize, compare numbers, organize, and build documents.
Run on a schedule, so the same job happens every morning or every Monday without you starting it.

What it cannot do, and you should want these limits.

It only sees what you give it. No connection and no folder access means it cannot reach that information. You are always in control of what it can touch.
It runs while your computer is awake and the desktop app is open. It is not a robot working in the cloud while your laptop is shut. If a scheduled run is missed because the machine was asleep, it catches up the next time you open the app. (There is a setting to keep your computer awake if you want a run to fire on time. Running truly while the computer is off is only possible in the developer tool, which is not this path.)
It is not connected to your patients, and it never should be. It works with summaries, systems, and templates, never identifiable patient records. More on that in section 6.
It drafts and proposes. You keep approval over anything that sends or commits. It does not fire off emails on your behalf unless you deliberately allow it, and for anything that matters, you should not.

Do you need API keys? No. This is the single most common thing that scares owners off, and it does not apply to you. The connections in this guide are a secure one-time sign-in, the same kind you use to log into any app. There is no key to generate, no code to paste, no developer to call. The only place an API key ever appears in your world is a custom chatbot on your website, which is a separate project and not part of this build.

What you need to begin: a paid Claude plan (Cowork is included on Pro and above, not on free), the Claude desktop app on Mac or Windows fully updated, and about one afternoon. One honest note: each email or calendar connection links one account at a time, so connect the account that holds your practice work.

3. What it actually looks like inside Cowork

A quick tour so nothing feels unfamiliar when you open it.

When you open the desktop app and select Cowork, the layout has a few simple pieces:

A left sidebar. Your tasks, a Scheduled section for anything set to run automatically, your connectors, and your projects.
A “+ New task” button, near the top. Every job you give the assistant starts as a task.
A chat input at the bottom. You type what you want and start it running.

Two things make Cowork different from a normal chat, and they are the whole point.

First, the slash menu. Type a forward slash in the chat input and a menu appears. This is how you pull in the Skills you have built (your voice, your rules) so every task starts with your context already loaded.

Second, scheduling. In any task, type /schedule. Cowork walks you through it, usually with a few multiple-choice questions about how often it should run. It then shows you the task name, the schedule, and a plain description of what it does. You confirm, and it lands on your Scheduled page, where you manage everything running for you. You can also open “Scheduled” in the sidebar and build one from a blank form: write the prompt, pick a frequency (hourly, daily, weekly, weekdays, or manual), and choose its folder.

On scheduling, the one rule worth repeating: these tasks run while your computer is awake and the app is open. A morning brief set for 7am runs when you open your laptop at 8am, not while it was shut overnight. If a run is missed, it catches up when you next open the app. For morning briefs and Monday summaries, this works exactly the way you would want.

That is the entire interface. A place to start tasks, a slash for your context, and a schedule for anything on repeat.

4. Connect your tools and get your reports flowing

This is the input work. It is the part most people skip, and it is the reason their AI never does anything useful. Do it properly and everything after is fast.

Install and open Cowork. Download the Claude desktop app, sign in with your paid account, and open Cowork from the left sidebar.

Connect what you already work in. Cowork is not limited to one ecosystem. Connect the tools you actually use:

Email and calendar, whether that is Gmail and Google Calendar or Outlook and Microsoft 365.
Team messaging like Slack, if your team runs there.
Business platforms and CRMs that offer a connection, through Cowork’s connectors.

Each connection is a one-time secure sign-in. No passwords handed over, no keys, and you control what it can see. While you are in your calendar connection, turn on write access if you want the assistant to place events, not just read them.

Get your reports to the agent. Here is the part that makes the operations side work, and it is simpler than it sounds. Most of the platforms a practice runs on, your EMR or practice management system, your marketing CRM, your KPI dashboard, your inventory tool, your ad accounts, do not plug into Cowork directly. They do not need to. Every one of them can either email you a report or let you export one, and the assistant reads both. You just have to point the reports somewhere it can reach.

Two ways, pick whichever is easier for each platform:

Have the report emailed to you on a schedule. Inside the platform, set up a recurring report (most have this) to send to your inbox daily, weekly, or monthly. Create a label or folder in your email so they are easy to find. The assistant reads them from your inbox. Nothing to download.
Have the export saved to a folder on your computer. If a platform only exports on demand, download it to one set folder, or set your downloads to save there automatically. Cowork reads local files, so it reads the latest export.

This is also how you bring in the numbers that have no connector at all, including your Meta and Google ad performance. Schedule the report email from each ad account, and they flow in like everything else.

Make a home for its work. Create one folder called something plain, like Practice Operations, in your cloud storage or on your computer. This is where the assistant reads from and writes to. Inside it, keep:

Your SOPs, role descriptions, and standard email templates.
Your brand and voice notes.
A place for the reports and exports above.
A short Targets note: your monthly revenue goal, your lead goal, your cost-per-lead ceiling, your margin targets. The briefings compare against this, so the assistant can tell you not just what happened but whether you are on pace.

A clear folder is the difference between an assistant that knows your world and one that guesses.

5. Teach it your voice and your practice

Out of the box, the assistant writes like everyone else’s. You fix that by building a Skill.

A Skill is a small instructions file that teaches Claude how you work. You build it once and it loads into every task after that, through the slash menu. Think of it as the briefing you would give a sharp new hire on their first morning, written down so you never repeat it.

Open a new task and run this:

I want to set up a Skill that captures how I run my practice and how I communicate. Interview me one question at a time. Ask me everything you need to know about my voice and tone, who my patients and team are, how I want email and updates written, what I always include, and the things I would never say or do. When we are done, save it as a Skill called “Practice Brain.”

Answer in plain language. A few things worth telling it directly, because they are what make the output sound like you:

How you write. Calm and direct, no exclamation marks, the way you actually sign off.
What you never do over email. Quote prices, make clinical promises, commit to dates without checking.
Who is reading. Owners and team members are spoken to differently than patients.
Your non-negotiables. The brand lines you hold, the words you avoid.

From now on, when you want it to use that briefing, you type a slash and select the Skill.

6. The guardrail that makes this safe

Before you connect this assistant to anything that touches patients, set one rule and never break it.

Identifiable patient information does not go into a consumer AI tool. Not names tied to treatments, not photos, not anything that points to a specific person’s care. The principle is simple to hold onto: AI builds the system, your EMR holds the patients.

This is not a limitation that weakens the build. It is what makes it usable in a regulated practice at all. Your assistant works with SOPs, templates, role descriptions, schedules, and summary-level numbers. The patient-identifiable layer stays in your EMR, where it belongs.

Write this rule directly into your Practice Brain Skill so it is enforced in every task:

Never include identifiable patient information in any output, draft, or document. If a task would require patient-identifiable data, stop and tell me instead of proceeding. Work only with templates, protocols, roles, schedules, and de-identified summaries.

Now you can build the rest with a clear conscience.

7. Build the personal assistant

This is the half you will feel within a day. Five jobs.

The morning brief

The single highest-value thing you can build. One task that checks your inbox and your calendar and hands you a structured briefing before you open anything yourself.

Read my inbox from the last 18 hours and my calendar for today. Give me a single brief with three parts. First, the three things that genuinely need a decision from me, with one line each on what and why. Second, everything else worth knowing, summarized in a sentence or two. Third, my schedule for today with any prep I should do before each item. Use my Practice Brain Skill for tone. Do not draft replies yet, just brief me.

Run it once by hand and read what it gives you. Adjust until it matches how you think about your morning, then type /schedule and set it for weekday mornings. After the first scheduled run, Cowork refines the task based on what it learned, so the second run is usually sharper.

Drafted replies in your voice

You want it drafting, so you approve and send.

Take the three decision items from this morning’s brief and draft a reply to each, in my voice using my Practice Brain Skill. Keep them short. Leave anything that needs a number, a date, or a clinical judgment as a clearly marked blank for me to fill. Do not send anything.

The judgment stays yours. The typing stops being yours.

Meeting agendas

Before any meeting, hand it the calendar event and let it prepare you.

Look at my 2pm today. Pull any related email threads and documents from my Practice Operations folder. Build a one-page agenda: the purpose, the three things we need to decide or move forward, the relevant background in a few lines, and the open questions I should raise. Use my Practice Brain tone.

Meeting recaps and follow-ups

This is where your AI note-taker pays off. When it sends a transcript or summary, the assistant turns it into action, not just a record.

When my AI note-taker sends a transcript or summary, read it. Pull out the decisions made, the action items with who owns each and a due date, and anything left open. Draft a short follow-up email to the attendees in my voice using my Practice Brain Skill, and add the action items to a running notes file in my Practice Operations folder. Do not send anything.

You leave the meeting and the follow-up is already written and the owed items are already logged.

The weekly recap

Friday afternoon, or whenever your week closes.

Review my sent email, my calendar, and the notes file in my Practice Operations folder from this week. Write me a recap: what got decided, what is still open and who owes it, and what is carrying into next week. Keep it to one page.

8. Build the operations briefings

This is the part you asked about most. Not a briefing that just tells you what the numbers were, but one that compares them to last week, to last year, and to the goals you set, then flags what moved and why.

The data gets in through the reports you set flowing in section 4: emailed to your inbox or exported to your folder. The assistant reads them, runs the comparison against your Targets note, and writes you a summary. Summary level only, never patient-identifiable detail.

Build these one at a time.

Weekly sales brief

Each Monday, read the sales reports in my inbox and the latest sales export in my Practice Operations folder. Write me a one-page summary of last week. For each key figure (total revenue, revenue by service category, average ticket, new memberships, rebooking rate), show the number, the change versus the prior week, the change versus the same week last year, and how it compares to the goal in my Targets note. Flag anything that is down or outside its normal range, and give me one line on the most likely reason. End with the single thing I should act on this week. Summary level only, no patient-identifiable data. Save it to my folder and draft it as an email to me so it is in my inbox.

Weekly marketing and ad brief

Each Monday, read the marketing and ad reports from my CRM and from my Meta and Google ad accounts, whether they arrive in my inbox or as exports in my folder. Summarize last week: new leads, booked consults, consult-to-treatment conversion, cost per lead, ad spend, and return on ad spend by platform. For each, show the change versus the prior week and how it compares to my Targets note. Flag any channel where cost per lead rose or leads fell, with a likely reason. Tell me where my money worked best and where it leaked. Draft it as an email to me.

Monthly cost of goods brief

Each month, read the inventory and cost-of-goods report from my inventory tool, in my inbox or folder. Summarize product and injectable cost as a percentage of related revenue, items trending over their target cost, and anything running low or overstocked. Compare to last month and to the margin targets in my Targets note. Flag the items eating into margin and what is driving it. Summary level only.

Weekly team brief

Each Monday, read the team and scheduling report in my Practice Operations folder. Give me a short operational read on last week: provider productivity, payroll as a percentage of revenue, and any schedule gaps coming up. Compare against my Targets note. Keep it high level. No individual patient data.

Set each to run with /schedule. By the time you have your coffee, last week’s sales, marketing, and team picture is waiting in your inbox, read and compared against your goals, instead of in five platforms you mean to log into.

The deeper point: the agent reads whatever you can put in front of it, an email or a file, then analyzes it against the targets you set. You decide the sources. It does the reading, the comparison, and the writing.

9. Turn your SOPs into a 90-day onboarding plan

Most practices write SOPs and then onboard new hires from memory anyway. The documents exist. They just never get assembled into a plan a new person can follow. This is one prompt.

Read every SOP and role description in my Practice Operations folder. Build a complete 90-day onboarding plan for a new front desk hire, organized by week. For each week, list what they should be trained on, which SOPs they should read, and what they should be able to do by the end of it. Pull from my actual documents. Where a step is missing or unclear in my SOPs, flag it as a gap instead of inventing it.

That last line matters. It tells you where your own documentation has holes, which is information you want.

You did the input work once, when you wrote those SOPs. Now it assembles, structures, and scales. Change the role and run it again for a new injector, a new coordinator, a new manager.

10. What to build first, and what comes next

Do not build all of this in one sitting. Staged rollouts hold. Simultaneous ones fall apart, because nothing gets the attention it needs to be trusted.

Week one: the morning brief. Just that. Build it, run it by hand a few times, tune the prompt, schedule it. This one habit changes your mornings and proves the whole thing works.

Week two: drafted replies, agendas, and meeting recaps. Now that the brief is reliable, add the jobs that sit right next to it.

Week three: one operations briefing. Pick sales or marketing, whichever number you most wish you looked at more often. Get its report flowing into a folder or your inbox, set up your Targets note, point a Monday brief at it, schedule it.

Week four and beyond: the rest. Add the other briefings. Build the onboarding plan when you next hire. Add the weekly recap.

It helps to stop thinking of this as software and start thinking of it as roles you are filling. The morning brief and the drafting is your executive assistant. The Monday numbers are your analyst. The onboarding builder is your operations lead. You are not learning a tool. You are hiring a small team into your practice, one role at a time.

11. Tips that make the difference

Run every task by hand before you schedule it. See what it produces, fix the prompt, then automate. Never schedule something you have not watched run once.
Tell it to flag gaps, not invent. Any time it works from your documents, add “flag anything missing instead of making it up.” This keeps the output honest and shows you where your systems are thin.
Keep everything in one folder. Scattered files are the most common reason an assistant gives vague answers.
Name things plainly. “Sales report,” “Targets note,” “front desk SOP.” Clear names work better than clever ones, for the assistant and for the next person you hand this to.
Set your Targets note once and keep it current. A briefing that compares to your goals is worth ten that only report the number.
Let it ask you questions. When it offers multiple-choice options during setup, answering them is faster and more accurate than a longer prompt.
Batch related work. Running tasks uses more of your plan’s allowance than a quick chat, because the assistant is doing multi-step work. Group related jobs into one session.
Keep patient data out, every time. Summaries and systems go in. Identifiable patient detail never does.

12. Your next 48 hours

You do not need the whole system this week. You need the first piece working.

Install the desktop app and connect your email and calendar. Turn on calendar write access.
Make your Practice Operations folder. Drop in your SOPs, templates, a place for reports, and a Targets note.
Build the Practice Brain Skill. Answer the interview honestly.
Write the PHI rule into it.
Build the morning brief, run it once by hand, tune it, then schedule it.

That is the foundation and your first working assistant, in one afternoon. Everything else in this guide is one prompt away once those five steps are done.

You did not open a practice to spend your nights sorting an inbox and logging into five platforms to pull the same numbers every week. This is how you hand that work to something that does it consistently, every day, while the judgment that actually requires you stays exactly where it is.