Celeste for accounting firms: the desk that scales with the deadline

Ask anyone at a tax firm what March feels like and they will describe the same inbox: what documents do you need from me, how do I get into the portal, is my return done yet, when is the extension deadline. The volume triples exactly when the people who could answer have the least time to. Clients are not being difficult. They ask because the answers matter to them and the firm is the only place to get them.

This surge shape is what an AI desk is for. Software takes the fourth “is my return done” of the hour with the same patience as the first, at 11pm, in April.

The routine queue, handled

Load your document checklists, engagement letter terms, deadline calendar, and portal instructions once. Celeste answers from them on email and web chat: which documents a Schedule C client still owes you, what the extension actually changes, how to reset a portal password, what your fee covers. When a client asks something the material does not answer, the agent logs it as a named knowledge gap instead of guessing, and your office writes that answer exactly once.

The “where’s my return” question deserves special mention because it is the one clients ask most and staff enjoy least. Connect your practice management or workflow system as a read-only tool and the agent answers with real status: received, in preparation, in review, filed on this date. The client who asks verifies their identity first, and the lookup is scoped to their own engagement by code that runs outside the AI. Nobody’s return status is visible to anyone but them, and an unverified sender gets no record data at all.

What stays with the professionals

Tax advice does not come from the agent. “Should I take the home office deduction” is not a portal question, and topic boundaries keep it out of bounds: the agent’s live topics are logistics, documents, deadlines, and status, and a message that crosses into judgment is held for a person. You can add hard tripwires on top. Put “IRS notice,” “audit,” or “amend” on the escalation keyword list and any message containing one is barred from an autonomous reply, queued for a human with the trigger flagged.

Anything that changes data follows the same rule every write follows in this product: drafted by the agent, executed only on a person’s approval, until you have deliberately decided a specific action is safe to automate.

Practice through one season first

The agent starts in shadow mode, drafting silently against your real client questions while your team reads what it would have said. A firm that turns this on in the off-season gets a low-stakes rehearsal: by January you know exactly which topics draft well, and you graduate them one at a time. Checklist and portal questions first. Status lookups when the scoping has proven itself. Anything touching advice, never, and the agent still triages those into a clean queue.

The economics are seasonal in the firm’s favor. The desk absorbs the surge without seasonal hires, and your staff’s billable hours stop leaking into questions the checklist already answers. Your clients, meanwhile, stop waiting two days for a one-line answer during the exact weeks they are most anxious.

See how it works for shadow practice and per-topic graduation, or the product for the day-to-day view. Starting the trial in the off-season means the agent has practiced on a full quarter of real questions before the season hits.