RoleAxis
Start free

Insights · The method · September 10, 2026

Not every back-office task deserves an AI agent.

That sentence is an odd thing for an AI agent company to publish. It is also the reason our assessment exists. Every role is a mix: some tasks belong with a person, some need better tools in the hands of the team you already have, and a specific kind of task deserves a governed agent. Knowing which is which, before you spend, is the whole method.

The industry has an honesty problem

Gartner expects 40% of agentic AI projects to be cancelled by the end of 2027, mostly for failing to show measured outcomes. Meanwhile Anthropic's Economic Index rates business and finance occupations at 94% theoretical task coverage by current AI, with observed real usage far lower. Read together, those numbers say something specific: the technology can address much of the work, and a large share of attempts to apply it are expected to fail, because they start from what AI could do rather than from what this task, in this business, actually needs.

Three outcomes, and what earns each one

Our assessment scores a role task by task with a deterministic engine: the same answers produce the same verdict every time, with no model judgment call in the number. Each task lands in one of three outcomes. Hire a human when the task runs on judgment, relationships, or consequence: negotiating with a supplier, answering an auditor, deciding who to trust. AI can draft around such work, but the work itself is a person's. Enable the team when the people are right and the tooling is slow: the task needs faster drafts or a workflow that removes the repetitive part, not a new worker. Deploy an agent only when a task is recurring, rules-based, high-volume, and checkable against records a system already holds: matching invoices to purchase orders, recomputing VAT lines, reconciling stock counts.

The engine actually distinguishes six classes underneath, including one that means fix your data or system access first, then score again. A task can be perfect agent material and still not be ready, because the records it would read do not exist in structured form yet.

Why an agent still is not autonomous

Even a task that earns an agent does not earn autonomy. In RoleAxis, every action an agent drafts waits in an approval queue for a named person; larger amounts escalate to the owner; and some things, tax lines, bank details, master data, are never changed by an agent at all. This is not caution theatre. The audit trail of who approved what is what makes the agent's value countable at month end, and countable value is exactly what the cancelled 40% could not show.

What this means if you run a back office

Do not start from the technology. Start from a role, list what it actually does with a rough share of the week, and sort the tasks. The honest result is usually a mix: some tasks argue for a hire, some for better tooling in the hands of the team you have, and the recurring arithmetic on records your ERP already holds argues for a governed agent. Our free assessment does that sorting in about three minutes per role, and when no live agent fits a task, it says so rather than stretching.

Score a role, freeWhy we build this

Sources

More from Insights: all posts. Related: Vision 2030's SME math, and where AI actually fits.