An AI assistant that answers from your ledger, not its imagination

The problem

Generic chat assistants — ChatGPT, Copilot, Gemini — have no visibility into your general ledger. Ask one how many uncategorized transactions you closed last quarter and it will either decline or fabricate a plausible-sounding answer. Pasting a trial balance into a public chat window is worse: vendor names, TINs, payroll figures, and bank balances move into a third-party cloud you do not control, and you have no audit trail of what the model saw or what it produced.

The alternative — enterprise AI integrations bolted onto an ERP — assumes a budget and an IT team most small and mid-sized businesses do not have. The integrations are often shallow, the contracts long, and the residency questions unanswered. CPAs and controllers end up with two bad options: an unscoped public chatbot that should never see client data, or a six-figure enterprise pilot that solves a problem they do not yet have.

SMBs get squeezed in the middle. They want grounded answers about their own numbers, not paragraphs of generic accounting advice. They want a tool their bookkeeper can use to triage a close, not a productivity demo.

How it works

Grounded in your live books

The assistant does not summarize what it thinks your books contain. When you ask a question, it issues structured tool calls — list_uncategorized_transactions, summarize_period, list_open_close_periods, list_stale_reconciliations, vendor_1099_lookup, year_to_date_pl, list_vendors_missing_tin, transactions_for_vendor — against your live database. The model only speaks to numbers it retrieved. If a tool returns no data, the assistant says so. If a question falls outside the available tools, it declines rather than guess.

Read-only and role-aware

Every tool exposed to the assistant is read-only. It cannot post journal entries, recategorize transactions, void invoices, or modify vendor records. The assistant also inherits the signed-in user's role permissions: a bookkeeper sees what a bookkeeper can see, an owner sees what an owner can see, and the model has no path to escalate. Sensitive fields such as full TINs are gated by the same rules that govern the rest of the system.

Self-hosted option

Inference can run entirely on premises. The single-tower tier ships with an RTX 5090 and is sized for a small office; the rack tier scales up to dual H200s for firms running multiple concurrent users or larger context windows. In both cases the model weights, the prompts, the retrieved data, and the responses stay inside your network. For businesses without space or appetite for hardware, ATCS offers a dedicated managed instance — single-tenant, segregated, and contractually scoped to your firm.

Audit-logged

Every prompt, every tool call, every tool response, and every model reply is written to an append-only audit log. Your CPA or controller can review exactly what the assistant was asked, which tools it invoked, what those tools returned, and what the assistant said back. This is the credibility layer: you are not taking the model's word for anything, and neither is your auditor.

What you get

• Natural-language queries against your own ledger, returned with the underlying numbers and the period they cover
• Read-only tool access to uncategorized transactions, period close status, stale reconciliations, vendor 1099 status, and year-to-date P&L
• Role-scoped responses that respect existing user permissions
• Refusal to invent dollar amounts, balances, close dates, or vendor identifiers when no tool returns them
• Refusal to write, modify, or delete any record in your books
• Private inference on your own GPU hardware, with no third-party cloud dependency
• Optional dedicated managed instance for firms without on-prem infrastructure
• Append-only audit log of every prompt, tool call, tool response, and model reply
• Vendor TIN gap detection and 1099 readiness checks surfaced through conversational queries
• Period-close triage — list what is still open, stale, or unreconciled before your accountant logs in
• Citations back to the underlying transactions or report rows the answer was drawn from
• No training on your data, no telemetry to outside vendors, no prompt logging outside your tenant

FAQ

How do you prevent AI hallucination on financial data?

The assistant is not asked to recall numbers. It is given a fixed set of read-only tools and is instructed to refuse any answer that would require inventing a figure. If the underlying tool returns nothing, the assistant says so explicitly. Every reply that contains a number is traceable, through the audit log, to the tool call that produced it.

Does my data leave my network?

On the self-hosted tiers, no. Inference runs on your own RTX 5090 tower or H200 rack, and prompts, retrieved rows, and model responses stay inside your environment. On the managed tier, your instance is single-tenant and segregated, and the data residency terms are written into the contract.

Which AI model does this use?

The assistant is built on Anthropic's Claude family for managed deployments and on equivalent open-weight models for self-hosted deployments. The architecture — tool-grounded, read-only, audit-logged — is what makes the answers trustworthy, not the brand of the underlying model.

Can the assistant make changes to my books?

No. There are no write tools. The assistant can read, summarize, and explain, but it cannot post entries, recategorize transactions, modify vendors, or alter period status. Any change to the books still goes through a human user with the appropriate role.

What if I don't have my own GPU hardware?

Use the managed tier. ATCS hosts a dedicated, single-tenant instance for your firm with the same tool architecture and audit logging as the self-hosted version. You can migrate to on-prem hardware later without changing how your team interacts with the assistant.

What the assistant handles that an analyst shouldn't have to

Most owners can't justify a fractional CFO on retainer for the day-to-day "what does this number actually mean?" questions, so they go without — making decisions on stale exports or gut feel. The ATCS Business Assistant fills that gap. It answers the lookup questions instantly, around the clock, grounded in your actual ledger, and when you do retain a human advisor it frees them from the rote work that should never have been billed at advisor rates. The point isn't to replace strategic counsel; it's to stop paying strategic rates for data retrieval.

AI does this better

• Instant answers at 11pm. Ask "how many uncategorized transactions do I have?" and the assistant calls list_uncategorized_transactions and answers in seconds. Your analyst is asleep and won't email back until tomorrow afternoon.
• Cross-year vendor pulls in one shot. "Show me every transaction with this vendor for the year" runs transactions_for_vendor once. A human opens five filters in QuickBooks and exports a CSV.
• Period summaries without prep. summarize_period returns the latest closed-month numbers without anyone building a deck or pivot table for a question that didn't need one.
• Compliance hygiene checks on demand. vendor_1099_lookup and list_vendors_missing_tin surface 1099 risk before January, instead of waiting for an advisor to remember to ask.
• Reconciliation drift, caught early. list_stale_reconciliations flags accounts that have gone too long without a reconcile — a question a busy advisor will only think to ask once a quarter.
• Open vs. closed period clarity. list_open_close_periods answers "is March closed yet?" without a Slack thread or a phone tag round.
• Year-to-date snapshots, repeatable. year_to_date_pl returns the same answer the same way every time — no "which version of the spreadsheet did you use?"
• Honest refusals. The assistant is tool-grounded; it will not invent a balance or a dollar amount it cannot retrieve. Every prompt and tool call lands in an append-only audit log, scoped to your existing role permissions.

What only a human analyst should still do

• Recommend a strategic shift in pricing, product mix, or service lines. That requires market judgment and conversation with you, not a query.
• Translate board priorities into a chart of accounts redesign. Structural decisions are interpretive work.
• Defend a forecast in front of the bank or an investor. A human signs their name to that.
• Negotiate with vendors, landlords, or lenders. Relationships and leverage are not tool calls.
• Interpret a Business Health Snapshot score in context. The platform produces the 0-100 across ten signals; an advisor decides which signal matters most this quarter.
• Build the multi-year plan. Strategy is a human craft.

The trade is straightforward. Your analyst stops answering "what's our YTD revenue?" four times a week and starts answering "should we open a third location, and what does the chart of accounts need to look like if we do?" The assistant absorbs the lookup tier so the human tier gets the work it was actually hired for. That's a better use of both — and a better use of your money.

Where to next

To see how the assistant fits into a specific firm — number of users, transaction volume, on-prem versus managed inference — run the pricing calculator. For the hardware side, the infrastructure page covers single-tower and dual-H200 tiers and what a managed instance includes. For the data-handling side, see security & data residency.