Note /
Which tasks to automate first
The first task is rarely the most valuable one you can imagine. It is the one your existing tools already reach, and the one that is safe to get wrong while you learn what the thing is like to work with.
Start where your tools already are
Read the connector items on the add-on menu before you read your own wish list. They name the systems we connect agents to: chat surfaces, CRM, accounting, Google Workspace or Microsoft 365, helpdesk. A first task living inside one of those is a standard connector at $400 one-off, and a Solo Install already includes one integration from the menu inside its $3,000. A first task living somewhere else is not refused, it is simply more work. Your own internal API is reached through a bespoke MCP server, which is a complex build at $700, and anything past the menu is priced on application once it has been scoped. The same job carries a different starting price depending on where it already sits.
Start with work that comes round again
The work listed on the home page as jobs agents take on from day one is all work that repeats. Invoices come due again, and the Monday numbers are wanted every Monday. That is the shape the tooling is priced for: the scheduled automations pack covers crons, reports and watches as a single item. A task that comes round once a quarter is a poor first candidate, because you will spend more attention specifying it than it gives back before you have forgotten it exists.
Pick something a draft is useful for
Under the default hardened posture, actions with consequences wait for human sign-off. An agent drafts the reminder or stages the change, and a person approves it before it leaves. That shapes which first task actually pays off. Choose work where most of the effort sits in the assembling rather than in the deciding. A chase email is mostly assembling, so a drafted chase saves real time even though someone still presses send. A judgement that takes ten seconds once the facts are in front of you saves very little, because the facts were never the slow part.
Whose job it is settles the product
Choosing the first task usually chooses the install for you. Work that sits inside one person’s day points to a Solo Install at $3,000, a single personal-class agent fitted to that workflow. Work that passes between people, or waits in a queue several people watch, points to a Team Install at $6,000 on a Slack or Teams surface with scoped workspaces. This is the same reasoning we apply to an enquiry, so picking the task first and the product second is the right order rather than a shortcut.
What not to start with
Not work whose whole value depends on the model writing well without anyone reading it. Model output quality is on the published exclusion list as your responsibility, so a first task should be one where a person is glad to review the output rather than obliged to. Not work that only becomes possible with the guardrails off, since a dangerous-equivalent posture is refused outright at any price. And not work that has no system to reach into, which is a question rather than a job, and is better served by a chatbot.
You do not have to have decided
The enquiry form asks what you want agents doing as optional free text, and lets you answer unsure on the product, because the description is what the recommendation is made from. The businesses this is built for run between roughly ten and two hundred people, usually with nobody whose job is platform engineering, so an unfinished answer is the normal starting position rather than a sign you are not ready.