Sebee: The AI That Doesn't Just Answer — It Acts

Ask most ERPs' "AI feature" a question and, at best, it draws you a chart. Ask it to actually do something — chase an overdue invoice, draft the journal entry that clears a suspense account, flag a duplicate bill before it gets paid twice — and you hit the wall. It's a read-only novelty, a search box with a personality. Sebee, the AI CFO built into LedgerQ, is designed around the opposite idea: an assistant earns its keep when it can safely operate your books, not just describe them. Most "AI in accounting" is a chatbot bolted onto a dashboard; LedgerQ is further along, because Sebee proposes and — once a human approves — executes real financial work, inside the same controls your staff already live under.

In summary: Sebee is LedgerQ's AI CFO. It answers plain-English questions from your live ledger, then goes one step further: it drafts balanced journal entries, extracts vendor invoices, forecasts cash, and flags anomalies. Every action that touches money is staged as a proposal a human must approve — with typed confirmation and a second approver above set thresholds — so the AI never bypasses your controls. It acts, but only on your terms.

What makes an AI "act" instead of just answer?

The difference is whether the assistant can produce a change to your records, and whether that change is safe. Sebee starts where every assistant does — natural language. You ask about your business in plain English and it translates the question into a real query against your tenant's live data: "What was our cash position this morning?" · "Show me overdue invoices over ₱10,000." · "Which customers have the highest DSO?" Answers come back grounded in the actual numbers, and Sebee cites the figures it used rather than hand-waving.

But an answer is where most tools stop. Sebee keeps going. Tell it "book last night's fuel purchase to field-ops" and it doesn't just explain how — it stages a balanced journal entry as a proposal card, showing the accounts, the amounts, and its reasoning. Nothing posts. The proposal sits in a queue until a person reviews it and approves. That's the whole distinction between an assistant that talks and one that works: the talking one leaves you to open the journal-entry screen and retype everything; the working one hands you a finished, balanced entry to approve or reject in one click.

How does Sebee turn a question into an action?

Sebee — the AI CFO runs through a deliberate four-step loop for anything that changes the books. First it interprets intent, pulling the exact figures from your live ledger rather than a stale nightly snapshot. Then it drafts: for any change, it builds a balanced proposal and surfaces the accounts, amounts, and reasoning as an action card. Then you approve — with a typed confirmation for larger amounts and a second approver above a set threshold; reject it and it's gone, with nothing left on the books. Only then does it post, and it posts through the same validated services your accountants use, writing to LedgerQ's double-entry general ledger as a real, balanced, fully auditable transaction. It's never a black box that mutates rows behind the scenes.

That grounding matters more than it sounds. Because Sebee operates on the authoritative ledger — the same posted GL that every report reads from — every figure it quotes and every entry it drafts is a real balanced transaction, not a number scraped from a cache that may have drifted. (We wrote more about why a single authoritative ledger changes everything.)

Can an AI safely touch the general ledger?

This is the question a controller asks first, and rightly. Sebee's reach is bounded by design, not by good intentions:

• Reads run against an allowlist. Sebee can only query an explicitly allowlisted set of tables, all scoped to your tenant. It can't wander into data it was never meant to see, and the tenant isolation that protects every other query protects the AI's queries too.
• Writes are proposals, never direct. Anything that changes the books is staged as a proposal a human approves. Larger entries require a typed confirmation; above a higher threshold, a second approver has to sign off — the same segregation of duties your security controls enforce everywhere else.
• Spend is capped. AI usage is bounded, so an agent can't run up an unbounded bill.
• Everything is auditable. Who — or what — proposed an action, who approved it, and when, is all recorded. AI actions are first-class citizens in the audit trail, not an untracked side channel.

The governing principle is simple: an agent should never be able to do something a person in that role couldn't. Sebee inherits your controls; it doesn't route around them.

What can Sebee do while you're not watching?

Answering questions is reactive. The higher-value work is the monitoring you never have time to do yourself, and here Sebee runs on a schedule:

• Cash forecasting. Sebee projects your cash position forward from GL-derived cash, open AR and AP, recurring entries, and historical posting patterns — and runs what-if scenarios and customer-payment scoring, so you can see which receivables are likely to land late.
• Anomaly detection. It learns a rolling baseline for metrics like daily revenue, new orders, AR over 60 days, low stock, and cash position, then flags statistically significant deviations before they become month-end surprises.
• Scheduled CFO briefings. Set Sebee to run recurring briefings — a written summary of your numbers and anomalies, saved as a chat session and delivered to your configured channels, like email — so the state of the business is waiting in your inbox instead of trapped in a report you have to remember to run.
• Alert rules and agent mode. Define reactive rules (a metric crosses a threshold → notify), and turn on agent mode, where Sebee's specialist agents monitor your books and stage actions for review. The agents propose; a human approves; only then does anything execute.

None of this is a blank cheque. An agent that touches money still routes its proposal into the same approval workflow a person would, and every specialist agent's reach is scoped to what its role permits.

What about the boring, high-volume work?

A lot of Sebee's value is quiet drudgery removal. Drop a vendor's bill into the OCR inbox and Sebee extracts it, auto-matches the vendor, and runs duplicate and sister-company checks before a person confirms it — turning a pile of PDFs into a clean queue. From there, LedgerQ's procure-to-pay flow runs real three-way matching: each invoice line is checked against the purchase order and goods receipt for price and quantity within tolerance, and an out-of-tolerance invoice gets a payment block until it's cleared. The effect is that the mechanical, error-prone parts of AP — reading the invoice, catching the duplicate, matching the PO — happen automatically, and the human step shrinks to a confirmation.

So does Sebee replace the accountant?

No — and that's the point. Sebee is built on Anthropic's Claude, but it's wired to defer. It drafts the routine entries so your team reviews instead of retypes; it catches the anomaly your team didn't have time to look for; it forecasts the cash so nobody is surprised. What it never does is post unreviewed, reach data it shouldn't, or act outside the role it was given. A chatbot that only describes your business saves a few clicks. An assistant that can safely operate it — propose the entry, extract the invoice, chase the anomaly, all under human approval — changes what a small finance team can actually get done. That's the bet LedgerQ makes with Sebee: AI is most valuable not when it talks, but when it does the work and shows you exactly what it did.

Frequently asked questions

Does Sebee post journal entries on its own? No. Sebee drafts a balanced journal entry as a proposal, but nothing posts until a human approves it. Larger entries require a typed confirmation, and above a higher threshold a second approver must sign off. On approval, the entry posts through the same validated services your accountants use, so it's a real, auditable ledger transaction — never an opaque automation.

How is Sebee different from the AI chatbots other ERPs advertise? Most "AI in ERP" is read-only: it can answer a question or draw a chart, but it can't produce a change to your records. Sebee is further along — it proposes and, after human approval, executes real financial actions like drafting entries and extracting invoices, all inside LedgerQ's existing approval, permission, and audit controls.

Can Sebee see or change another company's data? No. Sebee's reads run against an allowlisted set of tables scoped to your tenant, and its writes are staged as proposals a human approves. The same tenant isolation and role-based access control that protect every other query in LedgerQ apply to the AI, so an agent can never reach data — or tools — outside its role.

What can Sebee do automatically on a schedule? Sebee can run cash-flow forecasts, learn a baseline and flag anomalies (unusual revenue, aging AR, low stock, cash swings), deliver scheduled written CFO briefings to channels like email, and — in agent mode — run specialist agents that monitor your books and stage actions for review. Anything that moves money is still proposed for human approval first.

Is there an audit trail for what the AI does? Yes. Every AI-proposed action records who or what proposed it, who approved it, and when. AI actions are first-class entries in the audit trail, and because approved entries post to the double-entry ledger through the standard services, they're as reviewable as any human-entered transaction.