Work Theory Privacy Policy
Effective date: November 3, 2026 · Version: 3
Work Theory Inc ("Work Theory", "we") runs app.worktheory.ai and the Work Theory connector (at api.worktheory.ai). This policy says what we collect, why, who sees it, how long we keep it, and what you can ask us to do with it.
Questions or requests: hello@worktheory.ai. Postal address: Work Theory Inc, 242 Newkirk Ave #7C, Brooklyn, NY 11230.
1. What Work Theory is
Work Theory lets people invite others to help guide their AI agent. An agent's clients and members use it through their own chatbot. Guides, each in a team, teach the agent: they set rules, review what it made, and answer its questions. The agent does the work. Everyone adds the Work Theory connector to their own chatbot.
Work Theory keeps two kinds of knowledge:
- each agent's memories: where things stand, the decisions made, and the results;
- each team's playbook: its rules and its lessons, which it brings to every agent it guides.
2. Who is responsible for which data
- An agent's memories belong to its clients. We hold and process them to provide the service to the clients and to the teams they engage.
- A team's playbook and public profile belong to the team.
- We are responsible for account data (names, email addresses, sign-in records), billing records and the activity log described below.
3. What we collect
| What | Examples | Where it comes from |
|---|---|---|
| Account | Your name and email address, which agents and team you are in, and whether you are a client. Invitations: the address invited, who invited it, and the agent's or team's name. | You, or the person who invited you. |
| Sign-in records | The time of each sign-in, your IP address and your browser type, as Supabase's sign-in service records them. When a new account is started, the IP address of that request, kept apart from the account. | Automatically, when you sign in. |
| Agent content | Memories, including handoff notes. Questions and answers. A team's rules for that agent. What the agent brings to a guide for feedback: what it is about, its draft, and links to the work (never the files). Which teams each memory was shared with. | Saved by the agent's clients and members, or by its guides, almost always through their chatbot. |
| Team content | Rules, lessons, notes linking similar situations across the agents the team guides, and the guides' reviews. The public profile: the team's name, scope, description, services and contact address, and each guide's name and a line about them. A guide's outcomes and areas, in their own words. | The team's guides. |
| The activity log | Every call a chatbot makes to the connector: who made it, when, through which connection, which action, what it sent and what came back, and how long it took. Refused calls are included. | Automatically, whenever a chatbot uses the connector. |
| Fees, agreements and billing | Each fee proposed and agreed, the version of the team's terms agreed with it, and each box ticked when agreeing, with the words it showed. Each bill: what the client was charged, what Work Theory kept, what the team received, and anything the team owes. Notices to end guidance, with any note. Stripe's identifiers for your customer or connected account. | You, and Stripe. |
| One-time secrets | A password or key one person shares with another through a one-time link. | The person who shares it. |
| Limits | A count of the emails we sent for your account and the AI calls made for it. Nothing about what they said. | Automatically. |
The activity log contains copies of agent and team content, because it records what was sent and returned. Two things never reach it:
- secrets, which can't pass through a chatbot at all;
- what a client says while finding a guide: the log keeps only that a search happened, for which agent, and whether it worked.
What we don't collect. We never see or store a card number or bank account number: Stripe collects those on its own pages. We use no advertising or analytics trackers. See section 10 for cookies.
4. Your chatbot, and the AI we call ourselves
What we receive from your chatbot. You reach Work Theory through your own chatbot (Claude at launch). Work Theory receives only the calls your chatbot makes to the connector:
- who you are signed in as, and which connection the call came through;
- what the chatbot sends in that call, such as a memory to save, a question for a guide, or which guides to include in the chat.
We don't receive your conversations, your files or your chatbot's own memory, except what your chatbot puts into a call.
What we send back. Work Theory sends your chatbot the memories, rules and playbooks you are allowed to see. From then on, they are in your conversation. They are handled under your agreement with your chatbot's provider (Anthropic, for Claude), and its privacy policy applies to them, not this one. Your chatbot may also keep some of it in its own memory.
AI we call ourselves. A few steps in the portal send what you type or say there to Anthropic's API (a Claude model), to pick out the main points for you to check:
- a guide describing what they have achieved for the people they guide;
- a client describing what they want to get done, when finding a guide, once that is offered.
Only the words you entered on that step are sent. The step keeps nothing from the call: you see the result and edit it, and only what you then save is kept. Anthropic's terms for its API say it doesn't train on what we send.
Speaking instead of typing. Some pages let you speak. That uses your browser's own speech-to-text, and Work Theory receives only the text. Depending on your browser, your voice may be processed by its maker under its own terms.
Search. To help your chatbot find the right memories and lessons, their text is turned into a search index. The model that does this runs on our own hosting, at Supabase. That text isn't sent to an AI provider.
5. How we use it
- To provide the service. We store memories and playbooks, show each person what they are allowed to see, run sign-in, and send sign-in codes, invitations and notices.
- To support teams' lessons. A guide can draw a lesson from guiding an agent and use it with other agents, but only in de-identified form, and only if the agent's clients permit it. Before a lesson is saved, the database checks it. It refuses a lesson that names an agent the team guides or has guided, or that contains an email address, a web address, a phone number or an exact large figure. A client can withdraw permission (section 8).
- To improve Work Theory. We study the activity log to learn how people use the product and where it falls short. When finding a guide is offered, we count which kinds of help clients look for, by week, never who.
- To bill and pay out through Stripe.
- To keep the service secure, investigate misuse, and meet legal obligations.
We do not sell personal information. We don't share it for advertising, and we don't use your content to train AI models.
6. Who sees what
- An agent's clients and members see all of its memories.
- Each team guiding the agent sees only what was shared with it:
- what its own guides saved;
- what the agent's clients and members saved in a chat that included that team. At the start of each chat, the agent asks which guides to include, unless the chat already says. "None" is an answer.
It also sees the questions and work brought to it. Once a memory is saved, nobody can change which teams it is shared with. A team sees nothing of an agent while its engagement is paused or after it ends.
- A team's playbook is seen by its own guides. It is used in the chats of the agents it guides, when a chat includes that team. Lessons drawn from one agent can reach other agents' chats, de-identified. When an agent has several teams, its people see each team's name and what it covers, but one team never sees another's playbook.
- Anyone can see an approved team's public profile.
- Work Theory admins see the names and email addresses of teams, agents and their people, along with counts. The admin pages never show memories, lessons or rules. Every fix an admin makes is recorded.
- Work Theory engineers can reach the underlying database. They do so only to run, secure or repair the service, or when the law requires it. Our engineers work with an AI coding assistant (Claude Code, from Anthropic). Anything it reads from the database, such as the activity log while investigating a problem, is sent to Anthropic as part of that work.
- Service providers who host or run parts of Work Theory for us:
| Provider | What for | What it receives | Where |
|---|---|---|---|
| Supabase | The database, sign-in, the connector and our server functions, and the search model | Everything Work Theory holds | United States (US West) |
| Vercel | Hosting the portal at app.worktheory.ai | The portal's pages and requests, including your IP address, in its own request logs | United States (US West) and its global network |
| Resend | Sending our emails | Each recipient's address and the email: sign-in codes, invitations, notices when guidance is ending (with up to 1,000 characters of the note), failed payments, and requests to guide | United States |
| Stripe | Payments, teams' identity checks and payouts | For clients: the agent's name, the paying client's email address, card or bank details (on Stripe's own pages) and each invoice, which names the team and the agent. For teams: the team's name, the team owner's email address, and what Stripe asks for to verify identity and pay out. | United States |
| Anthropic | The AI steps in section 4 | Only the words entered on those steps | United States |
- The law. We disclose information when the law requires it, to protect people's safety, or as part of a sale or merger of Work Theory. In a sale or merger, this policy continues to apply.
7. How long we keep it
- Agent and team content is kept until it is deleted on request. Corrections don't overwrite anything. A memory that is corrected or retracted keeps its history, so the record of what was decided stays accurate. Ending an engagement deletes nothing.
- Deletion on request. See section 8.
- Activity log: kept for the life of the service. After a deletion request, it is anonymized rather than deleted.
- Billing records, agreements and notices: kept as long as tax and accounting law require.
- IP addresses of new-account requests: kept with no set limit today.
- Limit counts: kept one day.
- One-time secrets are destroyed after their one reading, or after 7 days unread. We keep a record that a secret was shared: its label, who asked, who pasted it, and who read it and when. We don't keep the secret itself.
8. Your choices and rights
- See and correct. You can change your name and email address in the portal. You can correct a memory through your chatbot, and the correction keeps the earlier version's history. Only whoever saved a memory, or a client of the agent, can correct or retract it.
- Choose which guides see a chat. At the start of each chat, you decide which of your agent's teams to include.
- Delete. Anyone can ask us to delete their account. A client of an agent can ask us to delete the agent's data. Send the request to hello@worktheory.ai from the address on the account. A Work Theory admin checks that it came from that address and then carries it out:
- Deleted: the agent's memories, questions, rules, and the records of its secrets.
- Anonymized: your activity log entries, which stay only as counts that can't be linked back to you.
- Kept: lessons a team has already drawn in de-identified form. They belong to the team and identify no one.
We record each request and what we did.
- Shared lessons. A client of the agent decides whether its teams may draw de-identified lessons from it. Withdrawing permission stops new lessons. Today a client does this by asking their chatbot.
- Access and copies. You can ask for a copy of your data at hello@worktheory.ai.
- State privacy rights. Depending on where you live, you may have more rights, such as the right to know, to delete, to correct, and not to be discriminated against for using them. Write to hello@worktheory.ai.
9. Security
- Data is encrypted in transit.
- Access rules are enforced by the database itself, so no tool or page can get around them.
- One-time secrets are sealed with a key that is kept outside the database.
- Teams are kept apart: a team sees an agent's memories only as they were shared with it, and never another team's playbook.
No system is perfectly secure. If a breach affects your information, we will tell you as the law requires.
10. Cookies and browser storage
- One cookie. The portal sets one cookie, to keep you signed in. When it is long, your browser may store it in more than one piece. It has no other use.
- No trackers. The portal has no analytics, advertising or tracking cookies, and loads no scripts from other companies. Its fonts are served by the portal itself.
- Browser storage. The portal keeps three small things in your browser's own storage. They stay on your device and are never sent to anyone:
- the address you were invited at, so the sign-in form can fill it in;
- whether connecting your chatbot worked, for the page that is waiting to hear;
- a team owner's unfinished invitation, until that browser tab is closed.
- Other sites. Stripe's payment pages and your chatbot's own site set their own cookies, under their own policies.
- The marketing site, worktheory.ai, isn't run from the portal yet, and this section doesn't describe it.
11. Where we operate
At launch, Work Theory is for people based in the United States, and your data is stored in the United States.
12. Children
Work Theory is for adults acting for themselves or a business. It is not meant for anyone under 18, and we don't knowingly collect their information.
13. Changes
Each version of this policy is numbered and dated. A new version takes effect when it is published, and we tell account holders by email when it is.