Wajo launched its first consumer agent, Fo, at the end of September 2026. The pitch is simple: an assistant that does not stop at drafting. It calls businesses, sends emails, pays for things, and when the AI gets stuck, it hands the task to a trained human assistant.
That last part is the interesting design choice. It is also the reason to read the fine print before connecting anything to it.
This review covers what Fo does, how its payment and privacy model works, what the launch benchmark does and does not show, and where the marketing and the legal documents disagree.
What Fo is
Fo is a personal "action agent" from Wajo, Inc., a San Francisco startup founded by Shivani Poddar, who previously led AI engineering work at Google DeepMind and Google Labs. The company has at least one publicly listed angel backer. It is an early-stage company, and Fo had been on the market for only a few days when this review was written.
Fo works through chat. You describe an errand, and it plans, researches and acts. The examples Wajo shows are ordinary life admin: book a dentist, find a barber open on Sunday, get a plumber out, cancel a gym membership, send flowers, collect quotes from caterers.
Three capabilities separate it from an inbox assistant:
- It makes phone calls. It can sit through phone trees, call businesses with no online booking, and cancel subscriptions that can only be cancelled by phone.
- It acts under its own identity. Fo emails from its own address and calls from its own number. It does not send as you.
- It completes purchases. Each checkout uses a single-use payment card.
When automation fails, Fo escalates to a human executive assistant. According to launch coverage, that person receives only the information needed for the specific task, not your full chat history.
Pricing
| Plan | Price | What you get |
|---|---|---|
| Fo Mini | Free | Daily task limit that resets each day |
| OG Fo | Free, invite only | Unlimited tasks and call time, group chats |
| Pro | Contact sales | Heavier workloads, MCP access, custom agents |
The homepage lists a free-tier limit of 1,000 tasks a day, which is not a meaningful constraint for one person. With paid human assistants behind the product, expect this pricing to change once usage grows. Treat it as a launch offer.
How payments and logins work
Payments are the strongest part of the design. Fo checks out with a single-use card for each purchase, so the merchant never sees your real card and neither does the agent. If one transaction goes wrong, the exposure is limited to that transaction. Where merchants support newer agent checkout standards, Fo can use a one-transaction token instead.
You also control how much autonomy it has. Wajo describes three modes: draft only, ask before acting, or act independently on defined work. It says Fo is designed to ask before hard-to-reverse actions.
Logins are where the story gets less clean. More on that below.
What the launch benchmark shows
Wajo published its own evaluation: 104 real-world tasks, each run three times.
| Agent | Task completion (no human fallback) |
|---|---|
| Fo | 71% |
| OpenClaw | 42% |
| Hermes Agent | 35% |
The headline claim was "2x better." That holds against Hermes Agent. Against OpenClaw, the stronger competitor, it is about 1.7x. Still a large gap, but not the one in the headline.
The safety numbers deserve the same careful reading. Wajo ran "trust trap" scenarios designed to tempt the agent into contacting people without authorization, leaking private information or making decisions it should not. Fo passed 94% of them, against 74% for OpenClaw. In raw terms that is 10 violations in 172 hazard runs, roughly one in seventeen.
Two caveats apply. First, this is a self-run benchmark, and no independent lab has replicated it. Second, 94% is a strong benchmark result and also a meaningful failure rate for software holding your email and a payment method. Both are true at once.
Where the homepage and the fine print disagree
The homepage leads with "built on trust." The privacy notice and terms of service are standard startup documents, and on several points they are looser than the marketing.
Passwords. The homepage says Fo signs in to websites without seeing or storing your password. The terms say that when you direct the agent to sign in to an account you hold, you authorize Wajo to store and use those credentials, and it may buy things with a payment method saved in that account. In that scenario, the single-use card protection does not apply.
AI training. The homepage says your data is never used to train third-party models. The privacy notice lists developing, training and fine-tuning Wajo's own models as a use of personal information. Google Workspace data is excluded under Google's API policy. Your prompts and other data are not.
Advertising. The privacy notice permits sharing personal information with advertising partners for targeted ads. Google Workspace data is again excluded.
Call recordings. When Fo contacts a business for you, Wajo may keep the recording, transcript, summary and outcome.
Retention. Data is kept for as long as you use the service, or as long as needed for Wajo's business purposes. There is no fixed deletion window.
Liability. The terms place responsibility for spending controls on you. If you do not configure them, Wajo can act on instructions it reasonably believes came from you. Gen Digital's Agent Trust Hub, which grades AI products on their policies, gave Wajo its lowest tier and flagged a $100 cap on what Wajo pays out if something goes wrong. That rating is itself AI-generated, so treat it as a reason to read the terms yourself rather than as a verdict.
None of this is unusual for a seed-stage startup. It is unusual next to a homepage whose main promise is trust and privacy.
Do you need it if you already use Gemini or Claude?
For most tasks, probably not. Gemini in Gmail already summarizes threads and drafts replies. Desktop agents such as Claude Cowork handle research, documents and browser tasks, with your approval at each consequential step. Between them, they cover most of Fo's demo list: rescheduling meetings, comparing vendors, requesting quotes by email, drafting a parking ticket appeal.
Fo's real differentiator is the phone. Neither tool will wait on hold with an insurer or call a barber who has no booking page. If a meaningful share of your errands are phone-only, Fo solves a problem nothing else in a typical AI stack does.
Location matters too. Wajo's demos and merchant examples are mostly American. In markets where small businesses run on WhatsApp rather than phone lines, the calling advantage shrinks, and WhatsApp support is not mentioned at launch.
How to try it safely
- Use a secondary email account, not your main or work one.
- Turn on spending limits and "ask before acting" before the first task.
- Start with low-stakes, phone-only errands: a booking, a quote, a cancellation.
- Do not give it logins for banking, investment, government or work accounts.
- Reassess after a few months of real-world use and, ideally, an independent security audit.
Do not connect a work account under any circumstances. Your employer has not vetted this vendor, and a personal agent reading company email is a policy problem as well as a security one.
Verdict
Fo is a credible product from a credible team, built around one genuinely useful idea: an agent that knows when it needs a human, paired with a payment model that limits damage per transaction. Its lead over open-source rivals is real, even if the headline overstates it.
The trust claim is ahead of the evidence. The product is days old, the testing is self-reported, and the privacy notice permits model training, ad sharing and open-ended retention that the homepage does not mention.
Use it as a beta for phone errands. Do not make it the assistant that holds your inbox.
This review is based on Wajo's homepage, privacy notice (last updated August 24, 2026) and terms of service as of October 2, 2026, plus launch coverage. Wajo's benchmark figures are self-reported and have not been independently verified.
