Operator guide

What an AI staff accountant actually books — and what it refuses to.

Buyers search for an AI bookkeeper; what exists here is Lena, a staff accountant on a roster of ten, working at a trust level you set. This guide is what that means in practice: the entries she books, the ladder she climbs, and the things she will not do at any level — which are the reason the rest is safe to hand over.

The role

The backward numbers.

Lena books what was actually spent and earned: the actuals, entry by entry, each one attributed on the ledger. Month-end lands as booked actuals, not a model — and her booked actuals override everyone else’s projections, including Gabriel’s financial planning. That boundary is structural: the forward numbers belong to a different name, so the books never quietly become a forecast.

Like everyone on the roster, she discloses that she is AI and works under the same machinery: a trust level you set, a gate that is human-only, a ledger that records every widening with a written reason.

The ladder

What T1 looks like in the books.

At T1 Drafter — where every name starts — Lena’s entries are real work on your real inputs, parked for your release. You see each categorization, each accrual, each month-end draft before it lands. The first weeks are you reading drafts, which is slower than autopay and the point: trust granted by a marketing page is not trust.

Promotion is earned on evidence you have watched, one rung at a time, and reversible work earns it first: recurring categorization before adjustments, adjustments before close. Every promotion is your decision, on the ledger, with a reason.

The refusals

What she will not do, at any level.

The forward numbers: Projections, pricing models, and unit economics are Gabriel’s role, not hers. The books stay the record of what happened — a role boundary the runtime holds, not a habit she keeps.
Anything irreversible: Money leaving the building, filings landing with an authority — anything that cannot be undone stops at you at every trust level. That is one of the three always-escalate rules, enforced at the API.
Spending over your tier: A payment over the tier you set is refused by the server and parked for you with its context. The refusal itself goes on the ledger.
The judgement calls: Write-offs, disputed charges, and what a messy transaction means are decisions, not entries. They arrive at your desk as escalations with the evidence attached.
Working a dead credential: If a connection or key stops working, she stands down rather than improvising. The stand-down lands on the ledger with its reason, and you get an email. The failure mode is stillness.

The limit

Why the books ship at the bottom of the ladder.

CMU’s TheAgentCompany benchmark (NeurIPS 2025) found finance and admin among the weakest areas for full autonomy — the best model completed roughly 30% of simulated office tasks end-to-end unaided. That finding is why this role ships at T1 and climbs on evidence instead of arriving autonomous: the draft-first ladder is the benchmark’s conclusion, turned into architecture.

Direct answers

The bookkeeping questions.

Does this replace my bookkeeper?

“Replace” takes a role here, never a person. If the role is unfilled — you doing the books at midnight — Lena holds it. If a person holds it, the ladder supports her drafting under them indefinitely; what changes and when stays your call.

Can she file my taxes?

No. A filing is irreversible, and irreversible acts stop at a human at every trust level. What she gives your accountant is clean, attributed books to file from — which is most of what a filing costs.

What does she cost to run?

Her share of the $39 is $0 — the fee covers the whole roster. Her inference runs on your keys at provider prices, frugally routable on the models page; the worked totals by profile are on the total-cost page.

Start where every entry waits for you.

Real books, drafted and parked, from the first session — and a dial that moves only on evidence you have watched.

Request an invite The whole roster