Every finance system now ships with an AI assistant. Ask it why travel spend spiked in March and it will tell you, usually well. This is the copilot pattern: a conversational layer over your data that informs the human who still does everything.
Copilots are useful. They're also, structurally, a very sophisticated read-only mode.
The line that matters
The question that separates finance AI into two categories: when the AI finds the problem, who fixes it?
A copilot surfaces the unreconciled difference and hands it back to you. An agent, the unit of agentic accounting, takes the next step itself: pulls the supporting documents, identifies the missing transaction, drafts the correcting entry, routes it for approval or posts it within policy. The copilot compresses the finding; the agent compresses the fixing. And fixing is where the month goes.
Why the distinction gets blurred
"AI-powered" covers both patterns, and vendors have every incentive to blur them, because retrofitting agency onto a batch-era architecture is hard. An agent can only own a workflow end to end if the workflow lives in one system: the invoice, the approval, the payment, the ledger entry, the audit trail. Scatter those across five tools and the agent is reduced to a copilot again, describing work it can't touch.
That's why the copilot/agent line usually traces the architecture line. Chat layers bolt on anywhere. Agency has to be built into the system of record.
What to ask a vendor
Skip "do you use AI" (the answer is always yes). Ask instead: Can it post an entry? Chase a receipt without being told to? Complete a reconciliation, or only display one? What percentage of exceptions does it resolve without a human touch? The answers sort copilots from agents in about ninety seconds.
Copilots make the historian faster. Agentic accounting retires the historian role altogether, and lets the team do the work that was always supposed to come after the books were current.
Meet Astra, Light's always-on analyst, and the agents that do the posting.