Claude Implementation
Implementing Claude at a CPA Firm: What Actually Gets Installed
The firms getting real leverage from AI are not running a “chatbot pilot.” They are running Claude — Anthropic's own applications, on firm devices, inside an enterprise workspace configured for how a tax and audit practice actually handles client data. This is what that implementation looks like, step by step.
First principle: it's Claude itself, not a wrapper
A firm-grade implementation installs the software Anthropic ships: the Claude desktop and mobile apps on staff devices, and where the practice writes automation or works in code, Claude Code for the people who need it. No intermediary interface sits between your staff and the model. That matters for two reasons. First, capability: Anthropic's own apps get every new Claude feature the day it ships — connectors, projects, file handling — while wrappers trail behind or never catch up. Second, accountability: your engagement letters, your insurer, and your state board all get a cleaner answer when the tool in the workflow is the vendor's own product under an enterprise agreement, not a reseller's layer with its own data path.
The implementation, in six moves
1. Map the data before the tool
Inventory what client information staff would realistically put in front of Claude: return data, source documents, ledgers, workpapers. Sort it by what Section 7216 consent covers, what GLBA treats as customer information, and what never leaves the engagement. The implementation is designed around this map, not the other way round.
2. Stand up the enterprise workspace
A firm runs Claude on a commercial plan — with single sign-on, centralized user management, and commercial data terms — never on personal accounts. Anthropic publishes how enterprise data is handled; your WISP cites it directly.
3. Configure Claude for the practice
Projects scoped per engagement so one client's materials cannot bleed into another's context. Connector allowlists per engagement team. Retention aligned to your document policy. Defaults set so the safe path is the easy path.
4. Pilot on one engagement team
Two weeks, one team, real work: memo drafting, source-document cross-checks, close checklists. The pilot produces the evidence — hours saved, outputs retained as workpapers — that makes the firm-wide rollout a decision instead of a leap.
5. Train on what may and may not leave the building
Staff training is a compliance control, not an amenity. People learn what belongs in a scoped Claude project, what requires client consent first, and how to log AI-assisted work product the same way any workpaper is logged.
6. Govern and score it
A written rubric the firm can hand to a reviewer: which uses are standing-approved, which need sign-off, and how usage is reviewed. Governance is what turns “we use AI” into “we can show you exactly how.”
Why the enterprise workspace is non-negotiable
The difference between a personal Claude account and a firm workspace is the difference between a staffer's browser habit and an institutional control. On Claude for Enterprise, the firm gets single sign-on against its own identity provider, admin visibility, and commercial terms on how data is used — the things a Written Information Security Plan can actually cite. Anthropic documents its security and compliance posture publicly at its Trust Center, which is precisely the artifact you want attached to the WISP section describing third-party service providers. A firm that cannot point to those pages is not implementing AI; it is tolerating shadow AI.
What staff actually do with it
Once the workspace is governed, the use cases are unglamorous and immediately valuable: drafting client memos and correspondence without exposing return data beyond the engagement, summarizing and cross-checking source documents inside a scoped workspace, researching positions with the work product logged, workpaper review support with every prompt and output retained as evidence, and month-end close checklists that run the same way every time. None of this replaces professional judgment — it removes the typing, collating, and first-drafting that sits between judgment and a finished deliverable.
Where Skoor fits
Skoor does the six moves above as a fixed-fee, two-week implementation: the data map, the workspace, the per-engagement configuration, the pilot, the training, and the governance rubric — and you keep the report either way. The offer, the rubric, and a sample report are public: Claude for CPA firms, the governance rubric, and a sample assessment report.
Learn More
Get Claude implemented, governed, and scored
Two weeks, fixed fee, on Anthropic's own Claude apps — and you keep the report.
See the CPA offer