Fo hands hard AI errands to humans. That is not a workaround - it may be the product
Wajo’s action agent can call, email, book and pay. When automation reaches its limit, trained assistants finish the job. The revealing part is not that AI still needs people - it is where the product chooses to put them.
Wajo’s Fo is built to do more than suggest the next step. The company says it can call businesses, exchange emails and messages, manage schedules, book services and make authorised purchases. It has its own communications identity and uses one-time payment cards, turning a prompt such as “find a plumber” into an attempt to arrange the visit.
Its most revealing feature is the escape hatch. When a task requires context or judgment the agent cannot supply, Wajo says Fo can bring in trained executive assistants. That is not hidden as a temporary patch; human escalation is part of the product pitch. The system is designed around the fact that real errands contain unanswered calls, changing prices, ambiguous consent and people who do not expose an API.
The launch story is therefore less “AI replaces the assistant” than “software decides which parts still need one.” Wajo’s own evaluation suggests that its harness and guardrails improve autonomous completion and trust. It does not prove that Fo can finish most real-world errands, nor that human oversight automatically makes agentic action safe.
What Fo’s hybrid design tells us about the next generation of AI agents
1. An action agent is judged by the changed world, not the answer
A chatbot can return a list of dentists and stop. Fo is intended to contact them, negotiate available times, obtain approval where needed and leave the appointment booked. That changes the failure standard. An inaccurate answer is inconvenient; an unauthorised email, purchase or cancellation has already affected another person or system. Wajo’s definition of an action agent - owning a job until there is an outcome - is marketing language, but it captures a real architectural shift.
2. The human handoff is a routing decision, not continuous supervision
Wajo says its agents escalate when they need more context or human judgment and that Fo can work with trained executive assistants. That is different from asking one user to watch every step. The agent handles routine searching and coordination, then routes exceptions to someone better equipped to interpret an unstructured situation. The design can make human attention more selective, but only if the system reliably recognizes the cases it should not handle alone.
3. Wajo’s 71% result is promising - and company-authored
In a paper written by five Wajo team members, Fo completed 71% of simulated errands and preserved the user’s stated boundaries on 94% of adversarial “trap” runs. OpenClaw completed 42% of the shared errands and preserved trust on 74% under the authors’ setup. The evaluation is more informative than a polished demo because tasks run three times and outcomes are checked in simulated websites, inboxes and phone lines. It is still a vendor-designed, vendor-run benchmark that has not yet been independently reproduced.
4. The headline score explicitly excludes human help
The 71% autonomous-completion figure does not count errands handed to Fo’s human assistant. That distinction is essential. The benchmark measures what the automated system and its guardrails can do on their own; the product promises an additional path for cases outside that envelope. We therefore do not know from the published study what percentage of real customer requests the combined AI-and-human service completes, how long escalations take or what they cost.
5. The harness may matter as much as the underlying model
Wajo’s results attribute much of Fo’s advantage to the system around the model: memory for unfinished work, tools, task state, approval rules and checks before an action leaves the system. Its examples show guardrails holding an unrequested email, removing private context or stopping checkout when shipping pushes the total over budget. Anthropic describes the same four-layer pattern - model, harness, tools and environment - and warns that weakness in any one layer can make a capable model unsafe.
6. “Ask a human” is not itself a safety guarantee
Escalation works only when the agent notices uncertainty, supplies enough context and leaves the human time to intervene before an irreversible action. A separate 2026 position paper by Margaret Mitchell, Avijit Ghosh and Samir Passi argues that long exposure to automation can weaken the attention and skills effective oversight requires. The relevant metric is not whether a person appears somewhere in the loop; it is whether that person can understand, challenge and stop the action at the right moment.
7. Human labour is displaced from the interface, not necessarily removed
The user may experience one agent while a trained assistant handles the exceptional call or ambiguous negotiation behind it. That can be a valuable service and a sensible division of labour. It also creates questions the product category will need to answer: who performs the escalated work, under what conditions, how much sensitive context they see, how the intervention is disclosed and whether the price of the service reflects the human labour inside it.
8. Data claims need the same separation as autonomy claims
Wajo says Fo uses single-use payment cards, does not see stored website passwords and does not use personal data to train third-party AI models. Its privacy notice is broader: it says personal information may be used to develop, train and fine-tune Wajo’s own AI technologies and may be shared with service providers or third parties needed to perform the task. That is not a contradiction, but “not used by third-party models” should not be read as “not used for training at all.”
9. The useful question is where the handoff belongs
Search and scheduling can often proceed automatically. Spending money, disclosing personal information, accepting a changed price, interpreting a sensitive request or negotiating an exception may deserve explicit approval or a skilled human. The boundary should depend on reversibility, cost, privacy and uncertainty - not on whether the agent technically has a tool that can click the button. Products that expose this boundary clearly will be easier to trust than products that advertise a vague percentage of autonomy.
Fo’s human layer is easy to mock because it interrupts the fantasy of a universal autonomous assistant. It may be the more mature product choice. Most useful systems are not made reliable by pretending exceptions disappeared; they are made reliable by detecting exceptions, routing them well and learning which decisions should never be automatic.
The harder test comes after launch: independently measured completion, clear escalation records, auditable approval boundaries, transparent labour practices and evidence that the combined service works outside a simulated world. Until then, Fo is an interesting hybrid architecture with a thoughtful internal benchmark - not proof that the last mile of real-world agency has been solved.
Where we track this
- The AI agent harness, explained →
Why tools, memory, permissions and guardrails often matter as much as the model.
- AI changes tasks before it removes jobs →
A broader framework for understanding task redistribution rather than binary job replacement.
- Four practical agents you can build today →
Smaller, bounded workflows where the action and review points can be defined in advance.
Questions people ask
- What is Wajo’s Fo AI agent?
- Fo is Wajo’s consumer action agent. Wajo says it can use email, messaging, phone calls, calendars, websites and payment mechanisms to carry an errand through to an outcome, rather than only recommending what the user should do next.
- Does Fo use human assistants?
- Yes. Wajo publicly says Fo can bring in trained executive assistants when a task needs human context or judgment. The company presents this escalation path as a product capability, not as fully autonomous performance.
- Did Fo complete 71% of real-world tasks?
- Not exactly. A Wajo-authored study reports 71% autonomous completion in a simulated world of invented businesses, websites, inboxes, people and phone lines. Human-assistant help was excluded, real businesses could not be contacted, and the result has not yet been independently reproduced.
- Does keeping a human in the loop make an AI agent safe?
- Not by itself. The system must recognize when to escalate, preserve enough context for review and give the person meaningful authority and time to intervene. Approval rules, permissions, audit trails and the reversibility of each action remain important.
- Does Wajo train AI on user data?
- Wajo’s product page says personal data is not used to train third-party AI models. Its privacy notice says personal information may be used to develop, train and fine-tune Wajo’s own AI technologies. Users should distinguish those two claims and review the policy before connecting sensitive accounts.
Written 1 October 2026 from an owner-supplied French article published 30 September. Product capabilities, the human-assistant escalation path, one-time payment cards, account connections, approval boundaries and current plan descriptions were checked against Wajo’s official product, support, terms and privacy pages. Founder background was checked against Wajo’s About page: it describes Shivani Poddar as an AI engineering leader who launched frontier models at Google DeepMind and worked at Google Labs and Meta; it does not support the stronger wording that she was Google DeepMind’s head of engineering. The benchmark section was checked against Wajo’s 26 September paper, Trust and Task Completion in the World of Consumer AI Agents (arXiv:2609.33017), and Wajo’s accompanying methodology article. The paper reports 71% autonomous completion and 94% trust preservation for Fo, versus 42% and 74% for OpenClaw on shared tasks, but it was authored by Wajo staff, used simulated businesses and people, excluded human-assistant help from Fo’s completion score and has not yet been independently reproduced. Anthropic’s 9 April trustworthy-agents article supports the distinction between model, harness, tools and environment and its warning about unintended actions. The oversight caveat comes from the independent August 2026 position paper AI Agents Push Humans Out of the Loop by Margaret Mitchell, Avijit Ghosh and Samir Passi, not from Anthropic. The supplied article’s Stanford figures claiming 12% real-world versus 89.4% simulated household-task success could not be matched to the cited primary Stanford material and are omitted. The claim that human-assistant interactions directly train Fo was also not established by the reviewed primary pages; Wajo says feedback tunes the agent more generally. TaskNorth has no commercial relationship with Wajo, its founders or the other organizations cited.
Trying to work out which AI tools fit your task? Describe the outcome and get a Blueprint: the tools, the prompt, and the steps, with pricing we verified ourselves.