IrisGo Privacy Review: Is It Safe for Freelance Client Work? — AI tool privacy review for freelancers

IrisGo Privacy Review: Is It Safe for Freelance Client Work?

Short answer: IrisGo keeps your screenshots, transcripts, and recorded workflows on your own computer, but any task you send to a cloud model travels through OpenRouter to OpenAI, Anthropic, or Google, and its firm no-training promise covers only your Google data. Our verdict is USE WITH CAUTION for client work. Policy reviewed October 2026, a week after IrisGo updated its privacy policy and two days after its solopreneur launch. IrisGo is a desktop assistant for Windows and Mac that watches how you work, learns repeat tasks, triages your inbox, and runs workflows on its own. For a solo consultant, that means client emails, invoices, and contract files can all become AI context. Here is what the documents say, measured against how we vet privacy claims.

What IrisGo does with your data

Privacy at a glance

Privacy dimensionIrisGo’s answer
Trains on your data?Never for Google data; silent for screen and file content
Training opt-outNo toggle documented; choosing a local model avoids the cloud
Data retentionLocal device; rolling 14-day copy of Gmail and Calendar
Third-party sharingPrompts and context routed via OpenRouter to cloud models
Storage regionYour device; company is US-based, transfers to US possible
Enterprise-team tierNone published; no DPA offered; free during beta
Screen captureScreenshots only; periodic scans if task detection is on

IrisGo’s privacy policy took effect on September 29, 2026 (per IrisGo’s privacy policy, retrieved 2026-10-07). It describes a hybrid design. Account details such as your email or phone number sit on IrisGo’s side. Everything else, including screenshots, audio transcripts, chat history, and recorded workflows, is meant to live on your machine. The app captures still screenshots rather than video, and it transcribes speech on the device, encrypts the transcript, and throws away the raw audio.

The cloud enters the picture in three ways:

  • Cloud models. When a request goes to a cloud model, the prompt plus whatever context it needs is sent through OpenRouter, a model router, to the provider you picked. The policy names OpenAI, Anthropic, and Google as examples. You can also plug in your own provider key, or pick a local model so nothing leaves the device.
  • Google Workspace. Connecting Gmail, Drive, or Calendar under Settings, then Integrations, lets IrisGo read, summarize, label, draft, and send. It keeps roughly the last two weeks of mail and events in a local database. Drive files are fetched only when a request needs them.
  • Background features. If you switch on automatic task detection, IrisGo may look at your screen and transcripts on a schedule to spot to-dos, and it uses your chosen model to do it.

The strongest promise in the document is narrow. For Google data, IrisGo says neither it nor the model providers it calls may train general-purpose models on it, and that summaries derived from that data get the same protection. For everything else you show it, such as a client PDF on screen, a contract in a local folder, or a transcript of a call, the policy only says vendors are contractually limited to helping IrisGo run and improve its service. OpenRouter’s own policy adds that some model providers may use inputs for training, depending on the model chosen (per OpenRouter’s privacy policy, retrieved 2026-10-07).

IrisGo’s Responsible AI policy, also dated September 29, 2026, says contextual processing happens only when the user starts it and mentions reviewing training datasets for bias, without saying whose data those datasets contain (per IrisGo’s Responsible AI policy, retrieved 2026-10-07).

What this means for solo freelancers

IrisGo’s local-first design is better than most agent tools we have reviewed, including the screen-memory apps covered in our Earlyn privacy review. The risk sits in how the cloud path is configured.

Scenario 1: you teach Watch & Learn a client invoicing workflow. You open the client’s portal, pull their purchase order, fill in your invoice, and send it. IrisGo captures screenshots of each step, and when it turns them into a reusable skill, the relevant screen text goes to a cloud model. Unlike your Gmail data, that portal content has no explicit no-training clause. Based on the policy as written, whether it stays out of a training set depends on the model and provider behind OpenRouter, which you have to check yourself.

Scenario 2: you leave automatic task detection on during client calls. Task detection periodically reads your screen and transcripts. If a client shares a confidential roadmap or a contract draft, those words can become context sent to the cloud to extract a follow-up item. If your NDA limits where client material can be processed, this approach carries a contract-breach risk you cannot audit after the fact.

Scenario 3: you serve EU clients from the free beta. IrisGo’s only stated legal basis is your consent, it says data may move to the United States, and it does not mention a data processing agreement or the transfer safeguards behind cross-border processing. Based on the policy as written, if you feed EU client names and emails into IrisGo, you hold the controller duties with little paperwork to show for the processor side. Our explainer on what a data processing agreement means for solo consultants explains why that gap matters.

One more detail: the Terms of Service, last updated July 30, 2025, cap IrisGo’s total liability at $100 and place disputes in California courts (per IrisGo’s Terms of Service, retrieved 2026-10-07). If client data leaks through an automation, you carry the loss.

How to use it safely

IrisGo gives you more controls than most agents. Set them before any client material touches the app:

  1. Pick a local model for client work. In the app’s model settings, choose an on-device model for any workflow that touches client files or mail. Per the policy, local processing sends no request content to a cloud provider. Keep cloud models for public research such as news digests.
  2. If you need a cloud model, bring your own key. Connect your own provider account instead of the default OpenRouter route, and choose a provider whose API terms rule out training by default. That puts the training question under a contract you can read.
  3. Leave automatic task detection off. It is opt-in. Keep it that way, and run IrisGo on demand rather than letting it scan your screen in the background.
  4. Record Watch & Learn demos with dummy data. Teach the workflow using a test client, fake invoice numbers, and a sample PDF. Swap in real inputs only at run time, ideally with a local model.
  5. Connect only the Google services you need. Each one is authorized separately under Settings, then Integrations. Email triage needs Gmail; it does not need Drive. Disconnect when a project ends, which also wipes the local mail and calendar copies.
  6. Turn on multi-factor authentication for your IrisGo account, and full-disk encryption (FileVault or BitLocker), since your vault of screenshots lives on the laptop.
  7. Delete your vault between clients. Clear recorded workflows and stored memory at the end of each engagement, and email help@irisgo.ai to request account deletion if you stop using it.

Privacy-friendlier alternatives

If your contracts rule out sending client screens or mail to US cloud models, these options cover the same jobs with tighter data handling. All links are direct, with no affiliate tracking.

  • Proton Mail with Lumo for inbox triage and drafting. What it gives you that IrisGo doesn’t: an AI assistant from a company whose stated policy is no chat logs and no training on your conversations, next to an end-to-end encrypted mailbox, run from Switzerland. Pricing band: free tiers; paid Mail and Lumo plans roughly $5 to $15 per month. Best for: freelancers whose inbox is the core of their client work.
  • A local model runner such as LM Studio or Ollama for summaries, drafting, and document questions. What it gives you that IrisGo doesn’t: every prompt and every file stays on your hardware, with no router and no third-party model provider. Pricing band: free; needs a recent laptop with 16 GB of RAM or more for comfortable use. Best for: consultants handling NDA-bound documents who can live without screen automation.
  • Bitwarden or 1Password for any workflow that logs in or pays. What it gives you that IrisGo doesn’t: credentials and card numbers stay in an encrypted vault and are filled at run time, so a recorded workflow does not need to contain them. Pricing band: Bitwarden has a free plan and Premium under $20 per year; 1Password individual plans cost roughly $3 to $5 per month. Best for: anyone teaching an agent to work in client portals or payment pages.

None of these replace Watch & Learn’s show-once automation. The trade-off is simple: less automation, far less client content leaving your machine. A practical middle path is to keep IrisGo for your own admin, such as personal news briefs and calendar tidying, and route anything involving client names, documents, or payment details through the encrypted or local tools above. That split keeps the time savings where the stakes are low.

The verdict

ATP Privacy-Vetted: USE WITH CAUTION

Use with caution: IrisGo is a reasonable assistant for client admin only if you run client workflows on a local model, keep automatic task detection off, and record Watch & Learn demos with dummy data. Its local-first storage, consent-based features, and strict rules on Google data beat most desktop agents. But the default cloud route sends screen and file context through OpenRouter to third-party models, and the policy gives that content no explicit no-training guarantee. EU-facing freelancers also get no data processing agreement, and the terms cap liability at $100.

Frequently asked questions

Does IrisGo train AI models on my screen and files?

For Google data, the policy is clear: neither IrisGo nor its model providers may use it to train general-purpose models. For screenshots, local files, and transcripts, the policy is silent on training and only limits vendors to running and improving the service. The founder said on Product Hunt that personal activity is not used for public models. Choose a local model or your own API key to remove the doubt.

Is IrisGo GDPR-friendly for EU client work?

Based on the policy as written, it is a partial fit. IrisGo lists GDPR rights such as access, erasure, and portability, and relies on your consent as its legal basis. But it is US-headquartered, says data may be transferred to the United States, and does not mention a data processing agreement. A freelancer processing EU client data should document their own safeguards or keep client work on a local model.

Can I use IrisGo for HIPAA-covered information?

IrisGo does not mention HIPAA or offer a Business Associate Agreement anywhere in its privacy policy, terms, or Responsible AI policy. Based on the policy as written, sending protected health information through its cloud route would leave you without the agreement HIPAA expects from a vendor. If you work with covered clients, keep health data out of IrisGo entirely, or limit it to a fully local model on an encrypted device.

Does IrisGo record my screen all the time?

No. The policy says IrisGo does not record continuous video, only still screenshots. Watch & Learn captures screenshots while you demonstrate a task. Separately, if you opt into automatic task detection, IrisGo may analyze your screen and transcripts periodically to find to-dos. That feature is off unless you enable it, and you can revoke access in the app’s settings at any time.

Where does IrisGo store my Gmail data?

On your device. Connection tokens are stored locally, and IrisGo keeps a rolling local copy of about the last 14 days of Gmail messages and Calendar events to power to-do detection. It says it does not store Gmail, Drive, or Calendar content on its own servers. When a cloud model handles a request, the needed content is sent encrypted to that provider, which may only use it to answer.

Is IrisGo safer than other AI desktop agents?

On storage, yes: screenshots, transcripts, and workflows stay local, and a local model option exists, which many cloud agents lack. On cloud processing, it is average. The default OpenRouter route leaves the training question to whichever provider serves your model, and there is no business tier with a signed DPA. For NDA-bound work, a local model runner gives you stronger guarantees.

Sources

Reviewed by Jérémy, founder of AidTaskPro and GreenBudgetHub. Based in central France. Privacy posture sourced from public policies and vendor documentation as of October 7, 2026.

Get Your Free Cybersecurity Checklist

Protect your digital life in 5 minutes. Free checklist + weekly productivity & security tips.

Similar Posts