Key takeaways
- Start with two or three specific, repetitive tasks, not a company-wide AI mandate.
- Use business tiers of ChatGPT, Microsoft 365 Copilot or Claude, with SSO and MFA, never personal accounts.
- Fix file permissions and oversharing before connecting AI to SharePoint, OneDrive or Google Drive.
- Keep a human in the loop for anything that reaches clients, patients or financial records.
- Pilot with 5 to 10 people for a few weeks and measure time saved before buying more seats.
The businesses getting real value from AI did not start by buying licenses for everyone. They picked a few specific tasks, chose tools with sound data terms, cleaned up permissions, and tested with a small group before scaling. The ones that ran into trouble usually skipped those steps, or never decided anything and let staff use whatever free tools they found.
This guide walks through an eight-step rollout that works for offices of 5 to 150 people, followed by the safeguards that matter most and concrete use cases by industry. Plan on about 30 days from kickoff to a measured pilot.
Before you start
You will need admin access to Microsoft 365 or Google Workspace, a decision-maker who can approve tools and policy, and a short list of volunteers for the pilot. If you have not yet compared the main tools, read Copilot vs ChatGPT for business first.
The 8-step rollout
STEP 01Pick two or three high-value use cases
List the repetitive, text-heavy tasks that eat staff time. Good candidates are frequent, low-risk, and easy for a person to check. Score each on time spent per week, data sensitivity and how easy errors are to catch, then pick the top two or three.
Avoid starting with anything that makes decisions about people, such as screening job applicants, or that touches restricted data like PHI.
STEP 02Choose business-tier tools with no-training terms
Match the tool to the use case. Microsoft 365 Copilot fits teams that live in Outlook, Teams and SharePoint. ChatGPT Business (formerly ChatGPT Team) and Claude Team fit general drafting, analysis and research. As of 2026, all three vendors state that business-tier content is not used to train their models by default.
Confirm the specifics before buying: data retention, where data is processed, available admin controls, and whether the vendor will sign the agreements you need. For healthcare, that means a HIPAA BAA for the exact product. Not every business tier qualifies.
STEP 03Set up SSO, MFA and admin controls
Connect the AI tool to your identity provider so accounts are created and removed with employees:
- Copilot: licensed and managed in the Microsoft 365 admin center (Billing > Licenses), protected by your existing Entra ID conditional access and MFA.
- ChatGPT Business: configure SAML SSO in the workspace admin settings, verify your domain, and decide which apps and connectors are enabled. OpenAI's documentation says apps are on by default for Business workspaces, so review them.
- Claude Team: enable SSO, assign seats centrally and set which connectors are allowed.
Then block or discourage personal AI accounts for work. Use your web filter or endpoint tools where you can, and say it plainly in policy.
STEP 04Write and publish an AI acceptable use policy
Your policy should name the approved tools, list prohibited data (PHI, client confidential information, credentials, payment data, HR records), require human review of AI output before external use, and explain how to report a mistake. Keep it to two pages. Download our AI acceptable use policy template to start, and see our AI governance framework for the broader structure.
STEP 05Connect business data safely
AI becomes much more useful when it can reach your files, and much riskier. Do this in order:
- Copilot: run a permissions and oversharing review first. Check SharePoint sites with "Everyone" or "Everyone except external users" access, remove stale "Anyone" sharing links, and lock down HR, finance, legal and executive folders. Microsoft's SharePoint Advanced Management and Purview tools can help find overshared content, depending on your licensing.
- ChatGPT connectors: enable only the connectors your use cases need, such as SharePoint or Google Drive. OpenAI states source permissions still apply, so the same cleanup matters.
- Claude Projects: upload only the reference documents a workflow needs, such as SOPs, templates or a style guide, and share the project with the people who use it.
AI does not create permission problems, it reveals them. A file shared with the whole company in 2019 becomes one question away from anyone once AI can search your tenant.
STEP 06Run a pilot with 5 to 10 users
Pick volunteers from the departments that own your chosen use cases, including at least one skeptic. Give them 30 to 60 minutes of training on the tool, the policy and good prompting. Run the pilot for two to four weeks on real work.
STEP 07Measure results
Set a baseline before the pilot starts: minutes per proposal, time to clear a morning inbox, hours spent on meeting notes. Have pilot users log time saved and rate quality weekly. Also track the time they spend checking and correcting AI output, which is real cost.
At the end, compare time saved against license cost. Keep the use cases that clearly pay off and drop the ones that do not.
STEP 08Scale what works
Roll out licenses department by department, not all at once. Turn successful prompts into shared templates, custom GPTs or Claude Projects so new users start from what works. Review usage, licenses and enabled connectors quarterly, and revisit your use case list every six months as the tools change.
Safeguards that matter most
Data loss prevention (DLP). Microsoft Purview DLP and similar tools can detect and block sensitive data types, such as credit card or Social Security numbers, in email, files and, depending on licensing and configuration, some AI interactions. Endpoint and browser controls can limit pasting into unapproved sites.
Sensitivity labels. Labeling documents as Confidential or Restricted in Microsoft Purview lets Copilot inherit those protections. Microsoft states Copilot respects sensitivity labels and the permissions they enforce.
Oversharing review before Copilot. Covered in Step 5, but worth repeating. It is the most common gap we see.
Human in the loop. A qualified person reviews AI output before it reaches a client, patient, court, regulator or financial record. AI tools can produce confident errors, including invented citations and wrong numbers.
Write review rules by risk. A social post may need a quick read. A legal brief or a patient letter needs a line-by-line check by the responsible professional.
AI use cases by industry
Dental and medical front desk
- Drafting appointment reminder templates, recall letters and new-patient instructions with no patient details
- Summarizing insurance policy documents and payer rules for staff reference
- Writing office SOPs and training guides
- Patient-specific work only through a vendor that signs a BAA and is configured for PHI
Law firms
- First drafts of engagement letters, routine correspondence and internal memos
- Summarizing long depositions or contracts for an attorney's review
- Building clause libraries and checklists
- Strict rules: an attorney verifies every citation and fact, and privileged material goes only into approved, contractually protected tools
Hospitality, golf and country clubs
- Event descriptions, menus, member newsletters and social content
- Draft replies to guest reviews and common inquiries
- Staff training materials and multilingual versions of SOPs
- Summarizing member feedback and survey results
Finance and accounting
- Excel formulas, reconciliation checks and variance explanations
- Drafting client emails and engagement summaries for review
- Summarizing tax law updates and policy documents
- Client financial data only in approved tools, with recordkeeping rules confirmed before using AI meeting notes
If your business handles personal information of California residents, the CCPA and CPRA govern how you and your vendors process it. California's newer regulations on automated decision-making technology add requirements when technology substantially replaces human judgment in significant decisions such as hiring, lending or healthcare, with ADMT requirements phasing in from 2027. Keep a person meaningfully in charge of those decisions and confirm your obligations with counsel.
If data cannot leave the building
Some firms decide certain data should never go to a cloud AI vendor at all. A self-hosted model on your own hardware is an option for those workloads, with real tradeoffs in capability and upkeep. See how to run an LLM locally.
When to call an IT professional
Steps 1, 4 and 7 are business decisions you can lead yourself. Steps 3, 5 and the safeguards, SSO, permissions cleanup, DLP and sensitivity labels, are where an IT partner saves time and prevents expensive mistakes. Los Angeles IT helps LA businesses plan and roll out ChatGPT, Claude, Copilot and self-hosted AI at $95/hr with no long-term contract. Our Microsoft 365 support team handles Copilot readiness reviews.
AI Acceptable Use Policy Template
A fill-in policy for using ChatGPT, Claude, Copilot, and Gemini at work