How to roll Claude out across every QuickBooks Online client (CPA firm or fractional CFO).
I have sat through the partner-and-170-files version of this conversation, and the fractional CFO tired of tab-switching across four entities. Slow down. The model is not the bottleneck. The admin design is.
This is not the migration page.
Connecting three sandbox companies is a demo. Rolling Claude out across every QuickBooks Online client is an operations problem, whether you are a CPA firm with a hundred-plus files or a fractional CFO with eight to forty. Who can add a file, who can remove one, who can update the prompt when the author is on vacation, and whether you are allowed to do that without opening a ticket.
I have sat through the partner-and-170-files version of this conversation, and the fractional CFO tired of tab-switching across four entities. Slow down. The model is not the bottleneck. The admin design is.
Roll out in waves the team can actually review. For a CPA firm, a useful pattern is 30, then 60, then the rest of the book, on the order of 170 files if that is the book. For a fractional or outsourced CFO, the book is often 8 to 40 files, sometimes four entities under one founder. That is still a rollout. Self-serve add and remove. No monthly minimum. More than one admin. Treat every skill publish as a release: each person downloads again and chooses Replace. Run this in the stack you already use (Teams and TaxDome for most CPA firms; Drive and email for most fCFO shops; Slack is optional). Do not confuse this with Intuit’s native one-company Claude connection. That connector is a toy.
This is the rollout. It is not the migration week, which is OAuth, first files, and a parallel run.
Xero files ride the same connection model. The examples below are QBO-heavy because that is most of the book we see. The operating rules do not change.
Migration answers “do we have to re-key history?” (no) and “what does week one look like?” (connect, brand, import rules, parallel run, then the rest). Useful. Not enough once the question is a hundred-plus companies, or eight entities that still need isolation, and a team that cannot wait on a vendor.
Rollout is who is allowed to connect a client without emailing us, what happens when a client leaves mid-month, how the prompt stays current on file 87 when you changed it on file 3, and how you do that without Slack and without a single Claude project that can see the whole book.
If you came here from a hosted AI bookkeeper, you already know why that distinction matters.
Why firms leave hosted AI bookkeepers.
The pattern is stable. A firm tries a product that “does the books.” Onboarding a new QBO company is a ticket. Offboarding is a ticket. The vendor’s capacity becomes the firm’s capacity. Agreements change. The founder goes quiet. The firm has doubled volume and wants a partner who will say “not this quarter” instead of “yes, yes, yes” for six months. Fractional CFOs hit the same wall when they cannot connect Claude to multiple QBO clients without waiting on someone else’s queue.
What those practices say they liked about a connector, in their words: it uses the Claude they already pay for. It is not a whole software thing that wants them out of their own procedure.
Self-serve add and remove is the control they were missing:
- An Owner or Admin connects a QBO or Xero company through OAuth from the portal. One click per client, not a shared onboarding queue.
- The same person disconnects a company when the engagement ends. The books stay in QBO. You are not asking a vendor to delete a shadow ledger they should not have had.
- Staff access is per connection. A Viewer does not see bank data. A contractor can be removed. See roles and access and the data-flow post in this series for the Claude-side revoke.
If add/remove still requires a vendor, you have not left the hosted model. You have re-skinned it.
30, then 60, then the rest of the book (CPA firm).
Do not connect 170 files on a Friday because the plan math works. Close week will not forgive you.
A pattern that has held up for CPA firms:
- 1. First cap (on the order of 30). That is the Scale plan’s listed connection count (pricing is on the homepage). Use it as a cohort, not as a target to smash. Pick a mix of tiers, not only the clean cash-basis files. The prompt will lie to you if it only ever sees clean files.
- 2. Second cohort (on the order of 60). After the team can close the first cohort without a war room. This is where you find out whether the master prompt is actually reusable, and whether the person who built the Claude projects is a single point of failure.
- 3. The rest of the book (we have seen firms plan toward ~170 for a named month-end). Batch it. Thirty a week is a pace a manager can review. A hundred on Monday is how skill versions fork in the dark.
Monthly add and remove, no minimum. Some months you add ten. Some months you drop three. A contract that punishes that is a hosted bookkeeper with nicer copy. If you are still learning the prompts, you will feel the per-connection cost before you feel the time back. Do not buy the whole book in month one to chase a volume rate while the prompt is still v2.
Enterprise exists when the listed caps are the wrong shape. We are a two-person company. We will have that conversation. We will not pretend 170 files is a Friday afternoon.
A fractional CFO book of 8 to 40 files is still a rollout.
Do not treat eight files as a demo because a CPA firm’s first wave is thirty. A concentrated fCFO book is senior work: multi-entity QBO, a 13-week cash view, a board pack, often a construction WIP file sitting next to a SaaS file. Connecting Claude to every QBO client you already have is the whole operating system, not a trial.
A pattern that has held up:
- 1. First two or three companies, including at least one messy inherited file. Audit before you model on that file. If the prompt only ever sees a clean books-you-kept company, it will lie on the next one.
- 2. The related entities. Four entities under one founder is a common shape. That is multi-entity QBO, not four copies of the same prompt pasted into four chats. One project per entity. Combined reporting only for related companies. Read multi-entity reporting before you promise a consolidation. Combined is not consolidated.
- 3. The rest of the 8 to 40. Growth at 7 connections and Scale at 30 exist for this book (homepage pricing). Do not buy thirty to look efficient while the audit prompt is still v2.
Tab-switching across four QBO logins is the pain this is supposed to end. Native one-company Claude-to-QBO does not end it. It recreates it inside Claude. If you cannot name which entity the project is for without opening the chat, you have not rolled out. You have made a faster alt-tab.
Multi-admin, or the prompt freezes when someone is on vacation.
Claude projects have owners. Let one person (usually the most technical staff member) create every client project, and that person becomes the only human who can update the master prompt in those projects. They go on vacation. Close week continues. The prompt does not. I have watched this happen. It is a miserable Tuesday. A solo fractional CFO is not exempt: the contractor who built the projects is the same failure mode.
Fix it before wave two:
- More than one Numbers Game Admin. Owners and Admins can invite, remove, disconnect companies, and maintain firm rules. Do not leave a single Owner who also happens to be the only person who understands Claude. Matrix: roles and access.
- More than one owner on the Claude organization (Teams or Enterprise). Individual Claude plans cannot revoke a departing contractor centrally. Get on an org plan before wave two, not after the first offboard goes wrong.
- A written rule: no one creates a client project as a personal sandbox that the firm then inherits. Projects are firm objects.
The original-admin problem also shows up when two firms share one client file (a fractional CFO and a bookkeeping firm). The person who first connected the company has to invite the second. Budget that email.
Skill refresh is a release, not a suggestion.
Publishing a new firm section changes the next download. It does not push into Claude. Everyone who already installed the skill is on the old one until they download again and choose Replace. Refresh the connector and the skill together. A few minutes. Skipping it is how v4 and v6 close the same night.
What has reduced the chaos:
- Version number in the skill and in the MD file the project holds. If you cannot see the version without opening a chat, people will not look.
- First operation of the close: if this is not current, stop. The close how-to in this series treats that as step one.
- Publish notes, like a changelog. “We changed capitalization on Tier 3” is a release note. “New skill, please update” is not.
- A stale-file view the manager can read. Until the product shows it, the who’s-done board needs a version column.
We send product-update mail when the Wizard skill changes. That mail is not a substitute for the firm’s own release habit. If a live session dies in “Claude skill confusion,” the skill was stale. Replace it.
The SOP layer this is feeding is firm instructions.
Connect Claude to multiple QBO clients without mixing the books.
Native Claude-to-QBO is one Claude account to one QuickBooks company. Firms still alt-tab. Staff cannot replicate the partner’s process. Parallel sessions across the book do not exist. QuickBooks’ own AI has the same file-scale limit. We already wrote that comparison: Numbers Game vs QuickBooks AI and the MCP page.
Firm or fCFO rollout means one Claude (or ChatGPT) organization, one Numbers Game account, and one project per client. Not one project per tier. Not one project for “the four entities, I’ll remember which is which.” Unrelated clients in one project is how entries land in the wrong company. Parent/child grouping is for related entities only (multi-entity reporting).
Per-client MD file with facts (basis, thresholds, last month’s pack). Master prompt per tier, injected, not pasted by hand into 170 chats.
ChatGPT works as the model on the same connections. Scheduling a Saturday bank-feed pass or an overnight close is a Claude scheduled-tasks story today. If the practice is ChatGPT-only, design the rollout around people pressing the trigger phrase.
Gemini does not support MCP. Claude or ChatGPT, on purpose.
Teams and TaxDome are the CPA stack. Slack is optional.
A lot of product screenshots in this category assume Slack is the close channel. CPA firms on this product often are not in Slack. They are in Microsoft Teams and TaxDome. Documents live in SharePoint. Fractional CFOs are often in Drive and email, sometimes Notion. That is a fine stack. I would rather meet you there than make you stand up a Slack workspace for a vendor.
What rollout does not require:
- A Slack workspace.
- A Slack webhook as the source of truth for “who’s done.”
- The team to live in the Numbers Game portal after setup. Claude is the daily surface. The portal is admin: connections, people, rules.
What rollout does require:
- A place the package lands (TaxDome, SharePoint, Drive).
- A who’s-done list the partner or the CFO will actually open.
- The same trigger phrase on every project.
If you do use Slack, it is a notification path, not the close. The multi-client tools guide is the five-layer stack view if you need to place this connector in the rest of the firm’s software.
What the first 90 days actually feel like.
Expect the time to move in this order, not the reverse. Anyone who promised otherwise was selling.
- 1. Week 1-2. Connecting, inviting, installing the skill, fighting stale versions, writing thresholds that were previously “in someone’s head.” This is slower than the old close.
- 2. First full month-end on the first cohort. Prompt quality dominates. Fifteen minutes and forty minutes can both happen, on similar files, based on the prompt. Refine on one client per tier, then ship.
- 3. Second cohort. Multi-admin gets tested. So does offboarding. So does “can we add ten this month without a call.”
- 4. The rest of the book. Only after the board is boring. CPA firm: the rest of the ~170. Fractional CFO: the rest of the 8 to 40.
Do not hire for this. Do not stand up a parallel AI bookkeeping team. The point is the same headcount, the close and the narrative, not a cheaper categorizer. Staff who used to click now review. That is a change-management conversation inside the firm, not a vendor training video.
Proof that retrieval can stop being the job, on a published file: PrismEdge, Nicole Jordan-Dahdal, three-entity opex audit in under two hours instead of days. That is a fractional CFO story. CPA-firm rollout is more files, more admins, more versioning. Same connector.
A rollout checklist you can copy.
- Claude Teams or Enterprise (or ChatGPT org), not a pile of personal plans.
- Two Numbers Game Admins, plus an Owner who is not the only person who can bill.
- Firm section published, trigger phrase agreed, 8k cap respected.
- Client-file facts for cohort one: basis, materiality, capitalization.
- One project per client, MD file dropped in. Related entities grouped only when they are related.
- Skill install documented as a release. Replace, not “I think I have it.”
- Who’s-done board in Teams/TaxDome/SharePoint/Drive.
- Contractor offboard path: portal, Claude org, Intuit connected apps.
- Wave plan: CPA ~30, then ~60, then the rest of the book. Fractional CFO: two messy files, then related entities, then the rest of the 8 to 40.
- Named gaps accepted: no Desktop, no Sage 50, no Rillet, no ProConnect, no Gemini.
If a row on that list is “we’ll figure it out later,” it is a wave-one row. Later is how you get 170 projects and twelve prompt versions.
Rollout questions, answered.
Book a walkthrough with the real roster.
Bring the client list, the admin you want to own it, and the first thirty files (or the first handful, if you are an fCFO). We will connect a live company and walk add/remove, not a slide.