What to Prepare Before Setting Up AI for Your Business

When an AI setup disappoints, the cause is often the starting point rather than the tool: a vague goal, a shared folder of everything, a login passed around the office. Most of the preparation below is decisions and documents you already have, not technical work.

1. Choose specific tasks, not “use AI more”

A goal like “use AI more” gives nobody anything to configure, test or train. A task does. It has an input, a person who does it today, and a result that person would recognize as good or bad. Compare:

• Vague: “Help with customer emails.” Specific: “Draft a first reply to a service inquiry using our current service list and response templates, for a coordinator to edit and send.”
• Vague: “Use AI for HR.” Specific: “Answer employee questions about the PTO and expense policies, pointing to the relevant section, and send anything about an individual’s pay or a complaint to the office manager.”
• Vague: “Make proposals faster.” Specific: “Turn intake notes into a first draft of our standard scope section, flagging missing details instead of filling them in.”

Good first candidates are frequent, repetitive and text-heavy, and are already done by someone who can tell a good result from a bad one. Poor first candidates are rare tasks, tasks the team doesn’t agree on, and anything where the assistant would effectively be deciding something about a person. Pick two or three. In our workspace setups we limit the first engagement to three bounded uses for one team, so each one can be tested properly.

2. Name one owner for the reference material

Every document the assistant relies on needs someone who decides what goes in, what comes out and when it changes. Without that person, the material drifts out of date, and the assistant keeps repeating last year’s price list or a retired policy.

The owner doesn’t need to be technical. It is usually whoever already maintains the service list, handbook or templates. Agree on three duties: review the sources on a schedule and whenever a policy changes, replace outdated files rather than adding new versions alongside them, and be the person the team tells when an answer was wrong. For more, see how to build a company knowledge base for AI.

3. Gather approved source material, and decide what to leave out

“Approved” means someone with authority has confirmed the document is current and that everyone using the workspace may see it. Typical sources are service descriptions, the price sheet staff already quote from, internal policies, standard templates, answers the team gives to frequent questions, a short tone guide and a list of who handles what. A small, current set is more useful than a large, complete one.

What to leave out

• Drafts, superseded versions and “final_v3” duplicates
• Personnel files, compensation, health information and records about individual employees or customers
• Client contracts and confidential terms, unless a chosen task requires them and everyone in the workspace is allowed to see them
• Passwords, API keys, bank details and account numbers

Pay most attention to who can see an uploaded file. OpenAI’s help center says that in a shared ChatGPT project, “All members of a project can view and download files added to the project” (Projects in ChatGPT). Anthropic says members with view access to a shared Claude project “can see project contents, knowledge, and instructions” (Manage project visibility and sharing). Neither says an uploaded copy keeps the permissions of the folder it came from. The practical rule: upload only what everyone in the project is allowed to see.

Connected sources work differently. OpenAI says its Company Knowledge feature “respects permissions in the connected source” (Company knowledge in ChatGPT), and Anthropic says “Claude inherits each person’s permissions from the connected service.” Anthropic also notes that custom connectors using one shared credential “reach whatever that credential can access” (Use connectors to extend Claude’s capabilities). How a connection is set up matters as much as which folder it points to.

Model training is a separate question. Both vendors’ published policies say that, by default, they don’t use data from their business plans to train their models (OpenAI enterprise privacy; Anthropic: Is my data used for model training?). That covers training only. It doesn’t change who inside your company can see what you upload.

4. Set up accounts and permissions the company owns

The workspace should belong to the business, not to an employee’s personal login or a consultant’s account.

• A business plan in the company’s name. ChatGPT Business and Claude Team are the vendors’ plans for teams. Each requires at least two paid seats or members, according to the ChatGPT Business FAQ and What is the Team plan?. Plans change, so check the current terms before buying.
• Named users, one seat per person. No shared logins and no passwords passed around in a group chat. With named seats, you can see who has access and remove someone the day they leave.
• An administrator and a backup on your side, so access doesn’t depend on one person’s inbox.
• A decision about connections. If the assistant will connect to Google Drive, Slack or similar tools, decide who approves each connection and what it may do. Anthropic says that on Claude Team and Enterprise, an Owner must enable a connector before members can use it, and each person still signs in individually. In ChatGPT Business, apps are enabled by default and admins can change their availability for the whole workspace (OpenAI admin controls for plugins and apps).
• Temporary, limited access for any outside help. We configure workspaces in accounts the client owns, with named users and limited permissions, and never through consultant-owned accounts or shared passwords.

5. Collect real examples to test with

Before anyone relies on the assistant, test it against the work it will actually see. Collect around 20 real examples across the chosen tasks, with names and personal details removed or replaced. Include a mix:

• Routine cases: the everyday questions and drafts it should handle well.
• Missing information: a request without a date, a location or a service type, where the right behavior is to ask or flag rather than guess.
• Conflicting or outdated sources: cases where two documents disagree, to check that it says so.
• Requests it should refuse: questions outside the approved material, requests for confidential information, or a discount that no policy supports.
• Requests it should escalate: complaints, questions about an individual’s pay, legal threats, safety issues and requests for exceptions to policy, each with the name of the person it should go to.

For each example, write down what a good answer or the right handoff looks like. In our workspace setups, about 20 such cases form the acceptance test the client reviews before the team relies on the setup.

6. Define success realistically

A useful definition is a team member can complete the agreed task with the assistant, without being coached through every step, and knows when to check its work or hand off. You can observe this by watching someone do the task.

Avoid defining success as an accuracy percentage. A score from a handful of examples says little about the next unusual request, and it encourages people to stop checking. Better signals are that drafts need light edits rather than rewrites, answers point to the right source document, out-of-scope requests get flagged instead of answered, and people are still using the setup a month later.

What you don’t need

• A data-science team. Setting up ChatGPT or Claude for a team means configuring an existing assistant with instructions and approved material. It isn’t training a model or writing software.
• Perfect documentation of everything. You need current material for the chosen tasks, not a written version of every process. Gaps found during testing are useful: they show where the team relies on knowledge nobody has written down.
• Every department at once. Start with the one team whose repetitive work is clearest, then decide what to extend.

If you’d like help with the setup, see how an AI workspace setup is scoped and how we work. Current prices are on the pricing page.