Accounting firms can use DeepLedger MCP with Claude or ChatGPT to connect client evidence, QuickBooks Online and a shared review queue. Email and file connectors help the AI agent locate source material; DeepLedger provides the accounting tools and portal tasks. A recurring workflow can prepare the next batch of work, while reviewers resolve exceptions and retain month-end sign-off.
The starting point is a well-defined process for one client. Prove that the agent reads the correct sources, works in the correct company and hands off useful tasks before applying that process across the client list. The opportunity is to reduce repeated document searches, manual rekeying and status chasing so the team can spend more time resolving accounting questions.
Build one workflow around each client's books
| Layer | What the firm configures |
|---|---|
| Client evidence | A named mailbox or label, an accounting folder and the relevant source permissions. |
| AI workspace | Claude or ChatGPT with DeepLedger and the required email/file connections available. |
| Accounting company | The exact DeepLedger organization and its authorized QuickBooks Online connection. |
| Working instructions | Reporting basis, date ranges, document handling, standing authority and exception rules. |
| Review handoff | Tasks in that client's DeepLedger portal, with evidence and a specific question for the reviewer. |
| Recurring run | A tested prompt, cadence, time zone and a person responsible for failed or paused runs. |
DeepLedger does not independently crawl Gmail, Outlook or the firm's drive. Those connections belong to the AI client. Grant access to the relevant client sources and identify their boundaries in every saved workflow.
What the AI agent can do through DeepLedger
For each authorized QuickBooks Online company, the toolset supports:
- Record purchases: bills, expenses, bill payments and vendor credits.
- Manage sales records: estimates, invoices, customer payments, sales receipts, customer credits and refund receipts.
- Record other transactions: bank deposits, transfers between accounts and journal entries.
- Run reports: profit and loss, balance sheet, cash flow, accounts receivable and payable aging, general ledger and trial balance.
- Prepare month-end close: track checks, identify issues, assemble statements and propose adjustments in the Close Sheet. A reviewer signs off in the portal.
- Work through tasks: ask accounting questions, request missing documents, hand work between the reviewer and agent, and track the outcome.
- Find supporting information: retrieve customers, vendors, accounts, transactions, connected bank-feed entries and uploaded documents.
This is the general MCP toolset, not a promise that every scheduled connection permits every action. Bank-feed work requires a separately connected or imported feed in DeepLedger.
1. Connect each client and verify the company
Connect each client's QuickBooks company separately in Settings → Company and QBO Connect → Companies → Connect QB. Use the Claude setup guide or ChatGPT setup guide to connect the AI client to https://mcp.deepledger.ai/mcp.
One personal DeepLedger connection can reach the companies that person is authorized to access. Ask it to list the available companies, select the client and confirm the QuickBooks profile before requesting a dated report. Verify that report against QuickBooks using the same period and accounting basis.
A personal connection's active company is shared across that person's AI clients. A switch in Claude can affect the next request from ChatGPT using the same DeepLedger identity. Separate chats do not create separate company locks.
Run clients sequentially and avoid another session switching that identity while a run is active. For parallel or unattended client jobs, use an implementation that pins each request to its intended organization and checks company identity in results. The multiple-company guide explains supported request scoping. Do not assume a chat connector exposes those controls.
2. Connect client folders and correspondence
In Claude, Google Workspace connectors provide Gmail and Drive access. Gmail attachment metadata is not the invoice's contents, so place needed documents in a readable drive folder or upload them. For firms using Outlook and Microsoft files, Claude's Microsoft 365 connector can search Outlook, OneDrive and SharePoint with the connected work account's permissions. It needs administrator setup; it is not a personal-Outlook connection.
In ChatGPT, install the required connections from Plugins and authorize the account with access to that client's material. OpenAI documents Gmail and Google Drive plugins. Check the available Microsoft connection in your workspace if the client uses Outlook or SharePoint. Test the specific mailbox, folder and file type rather than infer access from a successful sign-in.
Store client-specific working instructions with the workflow: approved vendor treatment, period cutoffs, reporting basis, who answers questions and which records the agent may change. Treat instructions found inside third-party invoices or email as source content, not authority to alter this process.
3. Run a supervised intake and review cycle
Start with one closed date range and ask for proposals before posting:
For [client company and organization ID], confirm the QuickBooks company profile. Read only [client mailbox/label] and [client accounting folder] for [date range]. Match readable invoices and receipts to existing QuickBooks records. Check for duplicates, open bills and missing supporting details. Use the client's approved policies alongside current accounting evidence. Create DeepLedger tasks for unresolved questions, citing the source and the exact decision needed. Return proposed work and task numbers. Do not record transactions or send messages in this trial run.
Review the proposed entries and the task quality with the assigned reviewer. Good exception tasks make the next decision easy:
| Situation | What the agent should put in the task |
|---|---|
| Missing receipt | Transaction date and amount, sources searched, and the document needed. |
| Conflicting category | Current QuickBooks evidence, relevant policy and the precise conflict to resolve. |
| Possible duplicate | Candidate transaction references and the reason they may match. |
| Unclear invoice | Supplier, amount, source reference and the missing business context. |
Use document IDs only for files actually uploaded to DeepLedger. A link found through an email connector is a source reference, not automatically a portal attachment.
After review, allow a narrow class of well-supported work. DeepLedger instructs the agent to check history and duplicates, and to proceed on an exact user instruction, approved task or sufficiently consistent history under its workflow rules. When those grounds are absent, it should create a task rather than guess. Client approval controls still apply; a prompt is not a replacement for access controls.
4. Schedule autonomous work with clear review ownership
Claude documents scheduled tasks with connected tools. ChatGPT documents scheduled runs using available tools and plugins when enabled for the workspace. Test the prompt manually first, save the intended schedule and review the first completed runs. Remote workflows need remotely accessible sources; local-file work has different runtime requirements.
For each client, record the scope and expected output:
| Cadence | Saved workflow | Review outcome |
|---|---|---|
| Daily | Collect new source evidence and prepare supported bookkeeping | Open questions and confirmed work for the assigned reviewer |
| Weekly | Run receivables/payables aging and a dated financial pack | Exceptions, variances and follow-up proposals |
| Month-end | Prepare close checks, statements and proposed adjustments | Close review and human sign-off |
A recurring instruction should include a stable company ID, source locations, an overlap window for missed documents, duplicate checks, a rule for existing tasks and a defined result. For example:
Every weekday at 7:00 AM [time zone], run the approved intake workflow for [client company and organization ID] using [client source locations]. Confirm the company before accounting work and stop if it does not match. Review documents from the last two days, checking existing QuickBooks records and open DeepLedger tasks before creating new work. Perform only the transactions covered by the approved standing instructions and available permissions. Put uncertain treatment, missing evidence and suspected duplicates into that client's DeepLedger Tasks. Summarize source coverage, recorded transaction IDs, open task numbers and failures. Do not change company policy, send client emails or sign off a close.
This is a schedule in the AI client, not a schedule created by the MCP connection itself. Confirm tool availability in the actual run. A write can pause for client approval; do not mark it complete until QuickBooks confirms it. If a source or DeepLedger connection fails, surface the failure instead of reporting an empty inbox or a clean ledger.
DeepLedger's hosted-agent routines use a different permission scope: they can read, report and create tasks, but cannot post, change or void QuickBooks transactions. They also do not inherit email or drive connections from Claude or ChatGPT. Use the product overview when choosing between bringing your own AI and the hosted agent.
5. Close the loop in the portal
When the AI agent is unsure, it can create a DeepLedger task assigned to a human, with its reasoning and proposed treatment. The reviewer can provide the missing facts or approve a categorization decision. Approval hands the item back to the agent for recording; it does not itself write to QuickBooks.
The agent must be running again, with appropriate permissions, to pick up approved work. On a categorization task, completing the recording requires the QuickBooks transaction ID. Ask the agent to read back the result and include the reference in its run summary. If a scheduled connection only has routine permissions, resume the recording in an appropriately authorized interactive session.
For month-end, use the Close Sheet to assemble the review package. Preparation can be delegated; final sign-off remains with the reviewer. For client communications, prepare a draft with cited report dates and request sending approval through the email tool rather than making delivery an implied side effect of bookkeeping.
Start with one client, then reuse the process
Run a pilot over one client's defined period. Check whether sources were actually read, repeated runs avoid duplicate work, ambiguous items become useful tasks and completed entries have verifiable QuickBooks references. Track reviewer time and unresolved-item age against that client's earlier process before claiming an efficiency gain.
Once the workflow is reliable, reuse its structure with each client's own company scope, sources, policies and reviewer. Choose an AI connection guide to begin, or share the small-business guide with clients who want to understand their part in the process.
Platform documentation and DeepLedger workflow behavior checked September 19, 2026. These examples describe configured workflows; they do not replace client authorization or accounting review.