How DeepLedger Works with QuickBooks Online

DeepLedger connects QuickBooks Online to an AI agent, then gives you a portal for the decisions that should not be buried in chat. The agent can read the books, use supported accounting tools, and record approved work. QuickBooks remains the ledger of record, and you remain responsible for judgment and final verification.

The shortest version is:

1. Connect a QuickBooks Online company to DeepLedger.
2. Ask the hosted AI agent, or a compatible AI client connected through MCP, to inspect or work on the books.
3. Review anything the agent cannot support from your request, QuickBooks history, or an approved policy.
4. Verify completed work in QuickBooks or in a dated report.

That final step matters. A confident chat answer is not proof that a transaction was recorded correctly.

The Four Parts of DeepLedger

1. QuickBooks Online is the ledger

Customers, vendors, accounts, transactions, and financial reports continue to live in QuickBooks Online. DeepLedger does not create a second general ledger and ask you to reconcile two versions of the books. Our AI general ledger explainer covers why that distinction matters when comparing ledger AI products.

DeepLedger stores the coordination around the work: tasks, task discussions, documents, bank-feed status, close runs, memory, and tool activity. When you need to confirm a balance or transaction, the final evidence is still the QuickBooks record or a report read from QuickBooks.

2. The MCP server is the tool layer

Model Context Protocol, or MCP, is a standard way for an AI application to connect to a server that exposes tools. DeepLedger's hosted MCP server turns supported QuickBooks reads and writes into named operations with structured inputs and results.

For example, the server can expose separate tools to:

• read the active company, accounts, vendors, customers, transactions, and reports;
• create or update supported bills, invoices, payments, expenses, deposits, journal entries, transfers, and related records;
• retrieve documents and unrecorded bank-feed items;
• create and complete review tasks;
• read or update durable company memory; and
• build a month-end close for review.

The current server registers 26 tools: 20 for QuickBooks Online and 6 for the DeepLedger workflow layer. That number describes the present product, not an MCP standard or a guarantee that every possible QuickBooks action is available.

MCP itself does not decide whether an entry is correct. It gives the AI agent a typed route to a system. DeepLedger's validation, company scoping, task rules, and accounting checks are application behavior built around that route.

3. The AI agent plans and calls the tools

There are two ways to use the tool layer:

• DeepLedger Agent: use the hosted AI agent in the DeepLedger portal.
• Bring your own AI client: connect a compatible client to the DeepLedger MCP server and work from that client's chat.

https://mcp.deepledger.ai/mcp

The client and model handle the conversation and decide which available tool to call. DeepLedger validates authentication and company access, checks tool inputs, sends supported requests to QuickBooks, and maintains the shared review state. This boundary is important because different clients can present confirmations, tool results, and errors differently.

If you connect an external client, follow the current setup guide. Do not assume an old menu label or a successful connector sign-in proves the correct QuickBooks company is active.

4. The portal is the review surface

The portal is where chat work becomes reviewable accounting work:

• Tasks shows questions and proposed treatments, who must act next, the discussion, the decision, and the QuickBooks transaction reference after recording.
• Bank Feed shows imported or connected bank activity and whether an item has been recorded.
• Documents holds receipts, invoices, and other evidence the agent can read.
• AI Memory shows company rules, policies, patterns, and lasting context that guide later work. You can add, edit, or remove entries.
• Month-End Close presents financial statements, checks, proposed entries, notes, and reviewer sign-off in the Close Sheet.
• Activity shows recent tool calls and outcomes as a worklog.

The portal does not make every agent suggestion correct. It makes the handoff, evidence, and final state easier to inspect.

A Bank Transaction from Question to Record

Suppose a bank-feed line says NORTHSTAR SUPPLY, but neither the memo nor the amount establishes whether it was office supplies, inventory, or equipment.

Step 1: Analyze the evidence

The agent should first confirm the active QuickBooks company, resolve the relevant vendor or account, inspect possible duplicates and open documents, and look at recent QuickBooks history. If a receipt is attached, it should read that too.

This is where access helps. The agent can inspect live accounting context instead of guessing from a merchant name alone.

Step 2: Decide whether authority exists

A write can be supported by an exact user request, an approved task, or sufficiently consistent QuickBooks history under the server's workflow rules. If the evidence is missing or conflicting, the correct result is a task, not a transaction.

The task should explain what is known, what remains uncertain, and what treatment the agent proposes. A vague confidence score is not evidence.

Step 3: You review the task

In Tasks, you can approve the proposed QuickBooks category, choose a different category, or send a note back. Approval moves the baton to the AI agent and changes the task to the recording stage. It does not silently post the transaction in the browser action itself.

When a categorized item is approved, the portal also writes the reviewer-confirmed payee treatment to company memory. That makes the decision visible and reusable, but it should still be reviewed if later facts differ.

Step 4: The agent records and closes the loop

The agent picks up the approved task, records the supported transaction in QuickBooks, and completes the task with the QuickBooks transaction ID. For direct bank-feed recordings that never needed a task, it marks the bank item with the resulting QuickBooks ID.

That recorded marker prevents the same feed item from appearing as unfinished work again. The server also refuses a conflicting marker if the bank item already points to a different QuickBooks transaction, because that conflict may signal a duplicate.

Step 5: Verify the result

Note the finished task's QuickBooks transaction ID and locate it in QuickBooks, fetch the transaction again, or run a dated report that includes it. Check the company, date, amount, account, payee, and document linkage.

The task status tells you the workflow reached its recorded state. QuickBooks tells you what actually reached the books.

What Is Enforced, and What Still Needs Judgment

Some tools add stronger hard stops. A journal entry must balance. An unapproved categorization task cannot be completed as recorded. A month-end close cannot move to review until all 16 checks have a status. A signed close is locked.

Supported voids are instructed to require explicit user confirmation. Elicitation-capable clients can also receive an in-tool prompt. The tool refuses an explicit decline, but unavailable or failed prompts fall back to the AI agent following that instruction.

Other boundaries cannot be automated safely. A receipt may be missing. A payment may belong to an open bill. A large purchase may require a capitalization policy. A client's description of an expense may conflict with its legal or tax treatment. In those cases, the useful result is a clear question with evidence, not a fast guess.

How to Read DeepLedger's Activity and Audit Evidence

The Activity page shows a compact summary of recent work: the tool name, time, success or failure, selected input fields, and the result message. The underlying activity record stores capped inputs and outputs, so large payloads are shortened. Activity insertion is deliberately separate from the accounting tool call, so an activity-write failure does not roll back an otherwise successful QuickBooks request.

That design keeps bookkeeping work from failing because a dashboard log did not save, but it means the Activity page should not be marketed as an infallible transcript. For important work, use three pieces of evidence together:

1. the task or request that authorized the action;
2. the returned QuickBooks transaction ID or report; and
3. the resulting record in QuickBooks Online.

For a close, also inspect unresolved notes, proposed adjusting entries, and the record that sign-off will freeze.

A Safe First Session

Use a read-first session to prove the connection before asking for a write:

1. Connect QuickBooks Online through the portal.
2. Ask the agent to list the available companies and name the active one.
3. Request the active company's profile, including home currency and report basis.
4. Run one report for fixed start and end dates, then compare it with the same report in QuickBooks.
5. Ask for a read-only list of a few recent transactions and note whether the result is paged or truncated.
6. Only then request one narrow, easy-to-verify piece of supported work.
7. Re-read the created or updated record and inspect the matching task or activity entry.

This checks identity, scope, dates, basis, completeness, and the write path separately. The QuickBooks MCP server guide explains the connection choices in more detail, while the write-access safety guide covers the review boundaries.

Current Limits to Know

• DeepLedger's QuickBooks tools target QuickBooks Online, not the QuickBooks Desktop product.
• A connected AI agent has access only through the tools and companies allowed by its authenticated DeepLedger identity, but those tools can include real write capability.
• Not every QuickBooks entity or correction has a tool. Some transaction types can be voided; others require a different supported correction or a handoff to a human in QuickBooks.
• Confirmation behavior differs by AI client. Never rely on a popup as the only approval control.
• Memory can preserve a reviewer-confirmed rule, but new evidence can still make an old rule inappropriate.
• The agent can organize evidence and execute supported steps. It cannot invent a missing receipt, determine undocumented business purpose, or replace professional judgment where the facts require it.

Frequently Asked Questions

What is DeepLedger? DeepLedger connects an AI agent to QuickBooks Online and adds a portal for review. QuickBooks remains the ledger of record. The agent can read reports and record supported transactions, while uncertain work becomes a task for you.

Does DeepLedger provide an AI agent? You can use DeepLedger's hosted AI agent in the portal or connect a compatible AI client through the DeepLedger MCP server. Both routes use the same QuickBooks and review tools, but the client controls the model and chat experience.

Can the AI agent write to QuickBooks? Yes. DeepLedger includes tools for supported bills, invoices, payments, expenses, deposits, journal entries, transfers, attachments, and other QuickBooks work. Access does not make a proposed treatment correct, so the agent must inspect the books first and route uncertainty to you.

What happens after I approve a task? Approval hands the task back to the AI agent at the recording stage. After a successful QuickBooks write, the agent completes the task with the QuickBooks transaction ID. Check that reference or the resulting report before treating the item as finished.

Does the activity page prove every accounting action succeeded? No. The activity page is a useful worklog, but it is not a substitute for the resulting QuickBooks record, the task's transaction reference, or a dated report. Large tool payloads can be shortened, and activity logging can fail separately from a tool call.

Start with the connection guide, use a read-only report to confirm the company, and keep the first write narrow enough to verify line by line.