OpenAI’s new errand benchmark scores agents on chores drawn largely from American daily life. A Canadian office has to redraw the line at the government sign-in.
Most offices already let software finish small errands: the banking app that pays a recurring bill, the supplier portal that reorders toner on the first of the month. OpenAI LifeVal, published November 2, grades how well OpenAI’s agents finish much bigger errands across a browser, a computer and connected apps. For a Canadian office, the score is a fair guide for shopping and planning chores and a poor guide for government errands, where the Canada Revenue Agency’s own sign-in rules keep a person at the submit button.
We read OpenAI’s LifeVal page on November 3. It names three sources behind the task list: the American Time Use Survey, the Multinational Time Use Study and a survey OpenAI ran with 3,995 respondents. One sample task, a home search in Luzerne County that ends with a deed placed in a records cart for review, is shown at 100%. That is an impressive chain of steps. It is also an American errand, and the page shows no Canadian one.
Key points
What does OpenAI LifeVal measure?
OpenAI LifeVal measures how often an OpenAI model completes an everyday errand from start to finish using browser use, computer use and connectors. OpenAI picked the errands from two time use studies and its own survey of 3,995 respondents, and says it used the test internally for the past year to track its own models.
Think of it as a road test where the examiner also drew the map. Each task is a chain: find the options, compare them, fill in details, then stop at a point where a person can review. The task areas run from caregiving and schooling to groceries, budgets and dealing with government. Scores appear per task as a percentage, and the page we read names no other company’s model.
That design is sensible for OpenAI’s purpose. A vendor that wants its assistant to book, buy and enrol needs a yardstick for those actions, and knowledge tests alone do not measure them. A buyer has a narrower question. Which of your errands look like the ones on the map?
Where the OpenAI LifeVal score stops for a Canadian office
The score stops where the errand list stops. LifeVal’s sample tasks are American in shape, and government services, the area with the most legal weight for a business, depend on rules a benchmark cannot see. We found no published LifeVal result for a Canada Revenue Agency, Service Canada or provincial errand on the page we read.
The Canada Revenue Agency is plain about sign-ins. Its account security guidance says it in a single line: “Do not share user IDs and passwords with anyone else.” The route it offers for help runs through a person: a representative registered for Represent a Client, whom the account holder authorizes from the Profile page in My Business Account. An agent that clicks through your business account on your behalf is using somebody’s sign-in. We see no slot for it in that process.
Banks drew the same line earlier, and we covered how Canadian bank agreements already restrict AI agents that log in for you. Government accounts carry higher stakes, because a wrong filing can bring interest or penalties.
A note on what we did. We have not run any OpenAI LifeVal task, and we have not watched an OpenAI agent work inside a CRA account. Everything above comes from OpenAI’s page and CRA’s published guidance, both read on November 3.
Sort your office errands by who presses submit
Before an agent touches office errands, list the ones your team repeats each month and mark each by who presses the final button: the agent, a person after the agent prepares it, or a person alone. Give it one afternoon. Anything that needs a government sign-in, a payroll remittance or a signature goes in the last two columns.
Here is the check, with an owner and a deadline.
- The office lead pulls every repeat errand from last month’s calendar, inbox and card statement.
- Beside each, write the account it needs and whose sign-in that account uses today.
- Place each errand in one of the three groups above.
- Pass mark: every errand that touches a government account lands in the second or third group, and the first group has a dollar cap written next to it.
- Finish by Friday, November 6, before anyone switches on an agent for office work.
The second group is where most of the value sits. An agent that gathers the receipts and fills the working file saves the bookkeeper real time, and the person who submits still signs in with their own credentials. Our earlier piece on CRA rules for scanned receipts in an AI mailroom covers the record keeping that sits underneath that work.
Is a person at the submit button just slowing the agent down?
Sometimes it is. For low-stakes chores like reordering paper or booking a room, approving every step wastes the hours an agent saves. That is why the sort keeps those chores in the first group. The slowdown is reserved for errands where a wrong submission costs money, a penalty or a client’s trust.
There is also a cost argument for fewer, larger approvals. Agents spend usage on every step they take, and an AI agent token budget grows fast when a person bounces work back three times. Batch the reviews: one approval for the week’s prepared filings beats ten interruptions.
Scores built in one market can also hide gaps in another. We saw that with a French AI chatbot that reached Quebec 53 days after English, and errands are no different. Canadian forms, French screens and provincial portals are untested ground until someone publishes results on them.
Two things would change this sort. One is OpenAI LifeVal results being published on Canadian government errands. The other is CRA publishing guidance that names software agents as a permitted way to act on an account. Until either appears, keep the government column with a person, and weigh the rest against the AI tools we compare for Canadian firms.
FAQ
What is OpenAI LifeVal?
OpenAI LifeVal is an evaluation OpenAI published on November 2, 2026. It scores how well OpenAI models complete everyday errands, using browser use, computer use and connectors, with tasks chosen from two time use studies and a survey of 3,995 respondents.
Can an AI agent file with the CRA for my business?
CRA’s account security page says user IDs and passwords should not be shared with anyone, and its route for help is a representative registered for Represent a Client. Our view: let an agent prepare the working file, and have a person sign in and submit.
Does a high LifeVal score mean an agent will handle Canadian websites well?
Nobody has shown that yet. The LifeVal page we read on November 3 shows American sample errands and no Canadian ones, so test on your own provincial and federal portals before relying on it.
