We gave OpenAI Codex full access to our Mac desktop and asked it to prepare a company tax return through OmaVero. The login had been handled in advance. Before that, Claude and Codex had gone through the accounting material together and reconciled the numbers.
And it worked.
This kind of annual tax-return routine is a perfect test for AI agents. Not as a demo, not as a toy, but as real office work: accounting figures, PDFs, an authority form, attachments, dates and the familiar feeling that one wrong field can ruin an afternoon.
This is where agents start to feel genuinely useful.
Not because they write better slogans, but because they can stay inside a messy workflow: one report here, another PDF there, OmaVero in the middle, numbers to reconcile, attachments to name correctly, and context that must not disappear halfway through.
That is a very different category from answering questions in a chat.
A useful agent is patient, careful with numbers, able to understand files and persistent enough to take the work to the finish line.
That is why we believe the near-term value of agents will often be found in boring back-office work. Not in robot-CEO fantasies, but in helping a team get through an authority form without five extra browser tabs and one small existential crisis.