CoDNA Request access

Setting up Claude or ChatGPT for your company

A practical guide for directors and IT leads in Georgia. It covers which plan to buy, how to set up sign-in and data rules, connecting your documents, a one-page AI policy, and the first tasks to automate.

Checked and updated on 27 October 2026. · ქართულად

In most companies we talk to, people already use Claude or ChatGPT at work. They use personal accounts, often paid for themselves, and paste in company documents to get answers. It helps, but the company doesn't control any of it.

This guide is how to turn that into something the company owns: a workspace the company pays for, people signing in with their company accounts, written rules on what data may go in, the assistant connected to your documents, and a few weekly tasks it does for you. These are the steps we take in the first weeks of a project, written down so that you can take them yourself.

Plans, prices and features change often. Everything here about them was checked on the vendors' own pages on 27 October 2026, and those pages are listed under Sources at the end.

Start from the work, not the tool

Before you buy anything, spend a week finding out where AI would help. Talk to five to eight people from different teams, for half an hour each, and ask them three things:

  • What do you do every week that follows the same steps each time?
  • What do people ask you over and over?
  • What do you write that others skim once and forget?

The same answers come up in most companies: meeting notes nobody writes up, the weekly report, questions about procedures and prices, getting a new colleague up to speed, code review. List five to ten of these. For two or three of them, write down how long they take today. That is your starting point, and it's how you'll know in two months whether any of this worked.

Choose Claude or ChatGPT

Both are good enough for most company work, and their company plans cost the same per person. Pick on what matters for your company:

  • What your people already use. People who already like one tool will use it more. That matters more than the small differences between the two.
  • Engineers. Each plan includes its coding agent in every seat: Claude Code with Claude, Codex with ChatGPT.
  • Where your data is stored. Anthropic says Claude's data is stored in the United States, and lists no other region for its company plans. ChatGPT Enterprise lets new customers choose where data is stored, Europe among the choices, at no extra cost. For ChatGPT Business this is rolling out and not yet available to everyone.
  • How many people. Claude Team is for 2 to 150 people, and ChatGPT Business for 2 to 200. Larger companies need an Enterprise plan from either.

Both list Georgia among the countries they serve.

The plans for a company of 10 to 150 people:

Claude Team ChatGPT Business
Standard seat, a month $25, or $20 a month paid yearly $25, or $20 a month paid yearly
Premium seat, a month, for heavy users $125, or $100 a month paid yearly $125, or $100 a month paid yearly
People 2 to 150 2 to 200
Sign-in with company accounts (SSO) Yes Yes
Accounts removed automatically when someone leaves (SCIM) No: Enterprise only No: Enterprise only
Coding agent in every seat Claude Code Codex
Paying By card By card

Prices are in US dollars before tax, and both vendors say they can differ by country. You can mix seat types: give Premium seats to people who use the tool all day, and Standard seats to everyone else. Premium gives about five times the usage of Standard on both.

Enterprise plans add automatic account removal, audit logs, control over how long data is kept, invoices instead of card payments, and on ChatGPT the choice of where data is stored. Claude Enterprise costs $20 per seat a month, paid yearly, plus everything people use at Anthropic's API prices. You can buy it online from 20 seats, or through sales from 50. ChatGPT Enterprise is sold through sales only, at a price agreed with you.

Buy it for the company, not for each person

Set up the workspace as the company's, not one person's:

  • Create it with a company address and pay with a company card. The Team and Business plans take card payments only: no invoices or bank transfers.
  • Have two owners, so the workspace doesn't depend on one person. Owners handle billing and settings; admins can manage people.
  • Verify your company's domain. You add a DNS record that proves the domain is yours. That's what lets people sign in with their company accounts, and on ChatGPT Business it lets people who sign up with a company address find the workspace and ask to join.
  • Bring people over from personal accounts. Ask them to move their work into the company workspace and cancel the plans they pay for themselves. Moving existing personal accounts over automatically is an Enterprise feature on both.
  • Enter the company's tax ID at checkout where the form offers it, and ask your accountant how the purchase is treated in Georgia.

Sign in with company accounts

Sign-in with company accounts (single sign-on, or SSO) is included with Claude Team and ChatGPT Business. It works with Google Workspace, Microsoft Entra ID and other identity providers. Once your domain is verified:

  1. Set up SSO with your identity provider, following the vendor's steps.
  2. Require it, so nobody signs in to the company workspace with a password of their own.
  3. Let accounts be created at first sign-in, so a new colleague doesn't wait for an invitation. Both plans can do this for people with an address on your domain.
  4. Add the AI workspace to your checklist for people who leave. Removing accounts automatically when someone leaves (SCIM) is an Enterprise feature on both, so on Team and Business someone has to do it by hand.

Decide what may go in

Most of the risk isn't in the tool. It's in people pasting what they shouldn't into accounts the company doesn't control. Write down which kinds of information may go into the company's AI workspace. Four kinds are enough to start:

Kind Examples In the company's AI workspace?
Public your website, published prices, product docs Yes
Internal procedures, meeting notes, plans, most code Yes
Confidential contracts, salaries, unreleased figures, security details Only for named people, in named projects
Personal data customers' and employees' details, health records, ID numbers Not until you've checked the law, below

Passwords, keys and payment card numbers never go in at all.

What the vendors do with what goes in:

  • Training. Neither trains its models on company-plan data by default. The exception is a chat someone sends as feedback, with a thumbs up or down, which may be used. On Claude, an owner can turn feedback off for the whole company.
  • Keeping. Chats are kept until someone deletes them. Deleted chats are gone from the vendors' systems within 30 days. Setting how long chats are kept is an Enterprise feature on Claude, and OpenAI's own pages disagree on whether ChatGPT Business has it.
  • Who can read chats. On ChatGPT Business, workspace admins can view, export and delete people's conversations. Tell people so.
  • Where it's stored. As above: in the US for Claude; for ChatGPT, in a region you choose on Enterprise, and on Business where that has reached you.
  • A data processing agreement. Anthropic's is part of its commercial terms, with nothing extra to sign. OpenAI signs one with ChatGPT Business and Enterprise customers who ask for it through its form.

Personal data. If customers' or employees' personal data goes into an AI tool, the vendor processes it on your company's behalf. Georgia's Law on Personal Data Protection sets out what that asks of you, such as an agreement with the vendor and telling the people concerned. This guide isn't legal advice: ask your lawyer before personal data goes in, and until then keep it out.

Connect your documents and tools

An assistant that can't see your documents answers from what it knows about the world, not about your company. Connecting it to where your documents live changes the answers more than any other step.

Both connect to Google Drive, Microsoft 365 (SharePoint, OneDrive, Outlook and Teams), Slack, Notion and GitHub. Claude also connects to Jira and Confluence, Linear, Asana and HubSpot, and ChatGPT to GitLab and Box. On Claude, an owner turns each connector on for the company. On ChatGPT Business many are on from the start, so an admin should look through the list. On both, each person then connects with their own account.

Three rules keep this safe:

  • Let each person connect as themselves. The assistant then sees only what that person can see in the source, and nothing more.
  • Start with two or three sources: the drive where documents live, the ticket tracker, and the code host for engineers. Add more when people ask.
  • Decide which actions need approval. A connector can write as well as read: post a message, create a ticket. On Claude, an owner sets each action to allowed, needs approval or blocked. On ChatGPT you choose whether it always asks first, or may go ahead with reading, with low-risk actions, or with everything. A document can also carry instructions written for the AI rather than for people (this is called prompt injection), and an assistant that acts on its own may follow them. Start with writing actions set to need approval.

Your own systems connect through MCP, an open standard both support for custom connectors. On Claude Team only owners can add one, and on ChatGPT Business owners and admins. Claude reaches a custom connector from Anthropic's servers, so it has to be reachable from the internet; ChatGPT offers a tunnel for systems on a private network.

Give it your company's context

The assistant answers better when it knows how your company works:

  • Shared projects. Both let you share a project, a set of instructions and files for one team or piece of work, with the whole company or with named people.
  • Skills (Claude). Instructions for a task that Claude uses whenever the task comes up, such as how to write a proposal your way. Owners can give a skill to the whole company.
  • Plugins (ChatGPT). OpenAI is retiring custom GPTs in favour of plugins, which bring instructions, files and connected apps together. Most workspaces lose GPTs on 11 December 2026, so build new things as projects or plugins.
  • Memory. Both can remember things about each person between chats. An owner decides whether it's on for the company, and on Claude each person then turns it on for themselves.

Whichever you use, write one page of what a new colleague would need to be told. Say what the company does, the names of products and teams, the words you use for things, and how you write to customers. If you work in Georgian, add your Georgian terms and an example or two of the tone you want, so the assistant gets your industry's words right.

If you use more than one AI, or your data has to stay in Georgia, the vendors' projects and memory aren't enough on their own. Each keeps what it learns to itself, on its own servers abroad. Keep the company's knowledge in one place that every AI reads, on servers you choose. That's what CoDNA does; it's one way to do it, not the only one.

Set up the coding tools

Claude Code comes with every Claude Team seat, and Codex with every ChatGPT Business seat. Engineers who use them all day may reach the seat's limit. Then give them a Premium seat, or let the company buy extra usage: Claude charges it at its API prices, and ChatGPT in credits. OpenAI's own estimate is that Codex costs about $100 to $200 per developer a month on average.

A few habits decide whether engineers keep using a coding agent after the first week:

  • Write the agent's instructions into each repository, in CLAUDE.md for Claude Code or AGENTS.md for Codex: how to build and test, the conventions, what not to touch. Keep them short and current.
  • Agree a review rule. For example: the author can explain every line, and tests come with the change. Code an agent wrote is reviewed like anyone else's.
  • Set permissions on purpose. Let the agent run tests without asking; make it ask before it pushes, deletes, or touches anything live.
  • Have AI review every pull request against your written conventions before a person looks.

Write a one-page AI use policy

People use AI more, not less, when they know what's allowed. A policy that fits on one page gets read; a long one doesn't. Cover these points:

  1. Which tools are approved, and that work happens in the company's accounts, not personal ones.
  2. What may go in, using the four kinds above.
  3. Who checks the output. Anything that goes to a customer, a regulator or into production is read first by a person, who answers for it.
  4. Code. The review rule, and that secrets never go into prompts.
  5. What to do when something goes wrong, such as a wrong answer that went out, or something pasted that shouldn't have been. Say who to tell, and that telling quickly is never punished.
  6. Who owns the policy, and when it's next reviewed. Every six months is about right, because the tools change that fast.

Automate two or three weekly tasks

Go back to the list from your first week and pick two or three tasks. Good first choices happen often and take real time. They should also do little harm when a draft is wrong, because a person checks it before it goes anywhere.

Three patterns cover most of what companies automate first:

  • On a schedule. Every Friday morning, a draft of the weekly report, built from the week's tickets, meetings and changes.
  • When something happens. When a meeting ends, a draft of its decisions and action items. When a pull request opens, a review against your conventions.
  • A draft a person approves. An answer to a question people keep asking, drafted from your approved pages and checked by someone before it goes out.

Start inside the assistant, with a shared project and a prompt people run themselves. When a task needs to run without anyone starting it, the coding agents (Claude Code and Codex) can run it on a schedule from your own server or build system. For each task, write down how long it took before, and check again after a month.

Train people on their own work

A general course teaches people what a prompt is. What changes how they work is an hour spent with the tool on their own tasks, with someone beside them. Three things work well:

  • Short sessions by role, an hour or two, on real work: sales on their proposals, support on their tickets, engineers on a ticket from the backlog.
  • Two to four champions: curious, respected people who learn more and help the others. They're also the ones who'll build the next automation.
  • One exercise for everyone. Ask the assistant a question about your company and find the detail it made up. Then ask again once it's connected to your documents, and compare. Nothing teaches people to check the output faster.

Measure what changed

After two months, check the numbers you wrote down in the first week. Look at how many people use it each week, too: both admin consoles show it. When few people use it, the usual reason is that it isn't connected to their work yet, rather than that they don't like it.

A checklist

  • Five to ten repeat tasks listed, with today's time for two or three
  • Claude or ChatGPT chosen, on a company plan, with two owners
  • The company's domain verified, and sign-in with company accounts required
  • Feedback for training and data retention checked in the settings
  • What may go in written down, in four kinds
  • Two or three sources connected, with writing actions set to need approval
  • A page of company context, with your Georgian terms
  • CLAUDE.md or AGENTS.md in each repository, and a review rule
  • A one-page AI use policy, with an owner
  • Two or three tasks automated, with their before numbers
  • Champions named, and a session for each team
  • A date, two months out, to check the numbers

Sources

Checked on 27 October 2026.

Claude

ChatGPT

Want help setting it up?

We set up Claude or ChatGPT for companies in Georgia, connect it to your documents and tools, and train your team on its own work. The first call is free.

Book a 30-minute call