Is Poke AI Safe? A Founder’s Privacy Checklist for Messaging AI Assistants

The short answer: Poke looks serious, but “safe” is the wrong first question

If you are asking “is Poke AI safe?”, you are probably not really asking about Poke only. You are asking whether it is sane to connect an AI assistant to your email, calendar, messages, documents, tasks, and half the operational nervous system of your company. That is a good question. Slightly terrifying, actually. Not because Poke is doing something uniquely weird, but because the entire category is moving from “chatbot that writes text” to “agent that can act inside your life.” Once an assistant can read your inbox, schedule meetings, update Notion, and send messages from the same place where your friends text you, the privacy model matters more than the demo. Poke’s official site positions it as a messaging-native assistant that now works in Apple Messages, with integrations across tools like Gmail, Google Calendar, GitHub, Notion, Linear, Todoist, Outlook, and others. Its FAQ says it supports Gmail and Outlook inboxes, calendars, and contact sync, and its pricing page lists a free plan, Pro at $19/month, and Ultra at $199/month. That is not a toy. It is infrastructure for personal operations. So the useful answer is not “yes, safe” or “no, run.” The useful answer is a checklist. Before connecting any messaging AI assistant, including Poke, Notis, or the next beautifully animated thing on your timeline, you want to know what data it touches, who can see it, how long it is kept, whether it trains models, and what happens when you leave.

What Poke says publicly about privacy

Poke’s privacy policy, last updated June 23, 2026, says Interaction collects categories of personal data depending on how you use the service. That includes contact information, account information, customer content such as files, documents, audio, video, images, data, and communications you provide, and sensitive information that may be revealed through connected email, including financial or health information. That sounds intense, but it is also the honest shape of the category. If you want an AI assistant to summarize email, triage your calendar, remember context, and act across apps, it cannot do that by politely staring at a wall. It needs access. The real trust question is whether that access is scoped, understandable, revocable, and used only for the job you asked it to do. Poke’s FAQ also says account deletion is available from privacy settings and that the full deletion process may take up to 24 hours, while backup archives may not be immediate and legal holds can extend retention. It also describes Poke Human, included with Ultra, as real-world tasks handled by a real person, and says operators only see the specific details needed for the task, not chat history, memories, or integrations. That last point matters. Human-in-the-loop products can be extremely useful. They can also be privacy-sensitive. A founder should not treat “AI” as one magic black box. Sometimes there is a model. Sometimes there is an API. Sometimes there is a human operator. Each layer needs its own trust boundary.

The founder privacy checklist

Here is the checklist I would use before connecting a messaging assistant to serious work. First, look at permissions. Can you understand what the assistant can access before you connect it? Email and calendar access are not small permissions. They expose customer names, investor threads, invoices, medical appointments, family plans, and the embarrassing draft you wrote at midnight and should never send. If the product makes scopes clear and lets you disconnect them, good. If it hand-waves with “we use AI magic,” close the tab. Second, look at purpose limitation. The assistant should use connected data to provide the features you request, not to build an ad profile, sell your data, or train random systems without a clear basis. Notis’ own privacy policy says Google user data is used solely for user-facing features you request, that Notis complies with Google API Services Limited Use requirements, and that Google data is not sold or used for ads. That is the kind of plain statement you want to find from any assistant. Third, look at human access. I am not religious about “no human ever sees anything.” Support, abuse prevention, legal obligations, and user-consented debugging exist in the real world. But the policy should say when humans may access data, and the product should avoid broad human access by default. Notis says humans do not read Google data unless you explicitly consent, it is necessary for security or legal reasons, or required by law. Poke says Poke Human operators only see details needed for a task and cannot access chat history, memories, or integrations. Different products, different workflows, same principle: least necessary access. Fourth, look at retention and deletion. Deleting an account should not feel like cancelling a gym membership in 2008. Poke says deletion may take up to 24 hours, with backup and legal-hold caveats. Notis says account data is kept while active and typically up to 24 months after closure, backups and logs are typically retained up to 180 days, and cached Google content is deleted within a defined window after disconnecting Google, subject to backup cycles. You may prefer shorter or longer defaults, but the point is transparency. Vague retention is where trust goes to die. Fifth, look at model training. This is the question founders ask badly. They say, “does OpenAI train on my data?” The better version is, “which model providers process my content, under what terms, and is my data used to train models for unrelated purposes?” Notis’ policy says content may be processed by model providers under terms restricting use to providing the service, and says providers are instructed not to use data to train models where feasible. If a product does not explain this, assume you need to ask before connecting sensitive accounts.

Poke vs Notis: the fair comparison

Poke and Notis are both part of the messaging-native assistant category. That is the interesting part. Neither is asking you to live in yet another dashboard all day. The assistant is reachable from the places where you already communicate. Where Poke feels especially ambitious is the consumer-life assistant angle. Its homepage emphasizes Apple Messages, rich actions, many app-style integrations, proactive help, and on Ultra, Poke Human for real-world tasks. If your dream is one assistant that can help across personal life, real-world errands, apps, and messaging, I understand the appeal. Notis is narrower in personality and deeper in founder operations. The product is built around the idea of an AI intern one message away: WhatsApp, Telegram, iMessage, Slack, email, automations, skills, memory, and connected work tools. In the privacy policy, Notis is explicit about Google Limited Use, TLS in transit, encryption at rest including OAuth tokens, least-privilege RBAC, audit logging, token scoping and expiry, secret rotation, anomaly detection, and separate environments. The real distinction is not “safe vs unsafe.” That would be lazy. The distinction is which trust model matches the work you are delegating.

Choose Poke if

Choose Poke if you want a broad personal AI companion that lives in your messaging experience, you like the Apple Messages direction, and you value the idea of Poke Human for real-world errands enough to pay for the higher tier. Also choose it if your use case is more life assistant than company operations assistant.

Choose Notis if

Choose Notis if your primary problem is founder execution: capture a thought in WhatsApp, delegate a task, run an automation, query memory, update Notion, draft an email, prepare for a meeting, and get back to work without opening another cockpit. Also choose Notis if you want the privacy policy to spell out operational controls like encryption at rest, OAuth token protection, RBAC, audit logging, and Google Limited Use in very plain language.

Choose neither yet if

Choose neither yet if you cannot explain what data you are connecting, what actions the assistant can take, and what would happen if it misunderstood a prompt. This is not anti-AI. This is adult supervision. Start with low-risk accounts, test the workflow, read the privacy policy, and only then connect the inbox where customers, investors, and your accountant live.

My take

I do not think the future belongs to the assistant with the prettiest chat bubble. It belongs to the assistant that earns enough trust to sit inside your actual workflow. Safety is not a vibe. It is permissions, retention, deletion, human access, model training, and operational controls. Poke appears to be a serious product with public privacy documentation and clear pricing. Notis has its own bias: we built it for founders who want a work AI intern in their messages, with explicit privacy and security language because trust is not a footer link. It is the product. So, is Poke AI safe? Read the policy, check the scopes, and decide what you are comfortable delegating. Then ask the same questions of Notis. Ask them of every AI assistant. The moment you connect your inbox, you are not choosing a chatbot anymore. You are choosing a junior operator with keys.