Expense reconciliation: automatically check invoices and bills once a week
💡 One-sentence summary: Use Grok Bot to set up your AI expense reconciliation—every week it automatically aggregates the expense system, matches receipts and invoices in the finance mailbox, flags policy exceptions, each with a policy citation, and follow-up drafts stop before approval.
This is the second lesson of Route A (workplace), using the same official six-step method as the previous lesson's daily briefing. This lesson corresponds to the Expense manager role, one of 8 roles on the official use-cases page Expense manager. The official positioning for it is: owns weekly expense reconciliation and follow-up on missing items—not "take a look at your expenses for you," but checking once a week, having someone follow up on missing items, and being able to explain the basis for every item.
The whole lesson takes about 10 minutes. The official guidance has two additional special requirements for this role, which this lesson treats as hard standards for AI finance reconciliation: every exception must have a policy citation, and totals must be traceable back to the source.
What you'll be able to do after this lesson
- Build an "expense manager" Bot: connect the expense system, finance mailbox, and finance sheets, and produce a weekly expense summary every week
- Have the Bot automatically match receipts and invoices from the finance mailbox, flag missing categories and policy exceptions—every exception with a citation of the original policy text
- Generate a follow-up draft for each colleague with missing items, stopping at the review point; it's only sent after you release it
- Schedule the reconciliation as a weekly routine: it runs on time in the cloud, and sending messages and modifying expenses always stay behind approval
Why use Grok Bot instead of a web-based AI?
| Capability | Web-based AI | Grok Bot |
|---|---|---|
| Can read your expense data | ❌ Export reports, copy and paste into the chat | ✅ Connects email and finance sheets to read them itself; logs into the expense system via the cloud computer browser |
| Remembers your expense policy | ❌ Re-paste the policy in every conversation | ✅ Give the policy documents to the Bot, write responsibilities into the description, and context accumulates across conversations |
| Reconciles on time every week | ❌ The web page won't come to you on a schedule | ✅ The routine triggers on a weekly schedule, runs in the cloud, and keeps running even with the laptop closed |
| Traceable exceptions | ❌ Says "possibly over the limit" based on the model's gut feeling | ✅ Officially required: every exception carries a policy citation, totals traceable back to the source |
| Chases colleagues and modifies expenses for you | ❌ Can only give text-based advice | ✅ Can do it, but follow-ups produce a draft card first and expense changes stay behind approval |
Your current predicament
- Every weekly reconciliation day: export from the expense system, dig through the finance mailbox for receipts and invoices, check totals in the sheets—the whole afternoon is gone
- Items missing invoices or categories require chasing people one by one, and after chasing you still have to double-check
- Which expense exceeded policy and on what basis is all judged from memory—scrambling through files when audit asks
When to use this technique
When you need to:
- Check the expense system, finance mailbox, and finance sheets once a week, confirming totals match and missing items have someone following up
- Compress "digging through the mailbox to match receipts" into "reading one exception list with policy citations"
- Colleagues with missing items automatically receive a reminder draft every week (sent only after you release each one)
And you don't want to:
- Manually export sheets, dig through the mailbox, and check items one by one
- Chase people one by one asking "where's your invoice," then double-check after chasing
🎒 Preparation before starting
- Finish reading 3.2 Skills and teaching—step 4 of this lesson stores the reconciliation process as a skill, and this is also the common prerequisite for all Route A scenarios
- Have a Bot you can assign work to: if you don't have one yet, first go to 1.5 Create your first Bot
- Think through the three data sources: the expense system (web address and login method), the finance mailbox (Gmail or Outlook), and the finance sheets (Google Sheets or Excel on OneDrive/SharePoint)
- Prepare an expense policy document (with limits, categories, and approval rules written clearly, sent to the Bot as the reconciliation standard)
- About 10 minutes, plus one round of human feedback after the first summary comes out
Follow along with me
Step 1: Write the expense manager's responsibilities into the Bot's description
Why: The first step of the official six-step method is to write the responsibilities, data sources, output format, and standing boundaries into the Bot's description—it relies on this text every week to understand its job.
How to do it: Create or open a Bot (sidebar + → Create a Bot), write the description like the following, adjusted to your actual situation:
You are my expense manager, responsible for weekly expense reconciliation and follow-up on missing items.
Data sources: my expense system, finance mailbox, finance sheets, and the expense policy document I provide.
Output format: weekly expense summary (totals traceable back to the source) + exception list (each item with a citation of the original policy text) + one follow-up draft per person responsible for a missing item.
Standing boundaries: only aggregate and draft; do not send messages or modify expense reports; explicitly state any data sources you cannot read, and never pad with old data.You should see: After saving the description, you can see this set of responsibilities on the Bot's detail page. From then on, no matter which conversation you assign work in, it works according to this positioning.
Step 2: Connect the expense system, email, and finance sheets
Why: The facts for reconciliation can only come from the data sources you connect—channels that Grok Bot cannot read cannot enter the summary.
How to do it: Handle each data source one by one; for details of the 5-step connection method, see 2.5 Connect apps:
- Finance mailbox: Search Gmail (or Outlook) in Connect apps and add it; when the Bot asks for Authorize / Authenticate, complete the login in the browser
- Finance spreadsheets: Google Sheets goes through Connect apps (use the Google Drive app to find files); for spreadsheets on OneDrive/SharePoint, once connected the Bot can pull the files onto its cloud computer to process them
- Expense system: If it appears in Connect apps, connect it directly; if there is no connector, have the Bot log into the web version using the browser on its cloud computer—login credentials are stored in Vault, and the Bot logs in on your behalf but cannot see the values; if the expense system provides an MCP server, you can also have the Bot add it by URL
- If authorization is interrupted, click Reconnect (on the app) or Reopen (on the connection card in the chat) to get back to the login page
What you should see: The mailbox and spreadsheet apps show Installed; the expense system can be opened in the Bot's browser and stays logged in. Note that Added ≠ logged in—when it shows Needs auth or Disconnected, you must click Authenticate (Authorize on mobile) to re-authorize for it to count. One more permission detail: for Google Sheets shared with you as Viewer, the Bot can only read them too—to write back the reconciliation status column, you need a spreadsheet you can edit.
💡 This is Grok Bot's unique value
A web-based AI can only look at the exported reports you paste to it; the Grok Bot has its own cloud computer and browser, so it can log into the web version of expense systems without connectors and check them itself. There are two paths for credential security: OAuth app tokens are stored in Cursor's connector backend, and the Bot only calls tools without touching login credentials; passwords for browser-based logins are stored in Vault, and the Bot likewise cannot see the values.
Step 3: Run your first AI expense reconciliation with the official starter prompt
Why: The official starter prompt deliberately stops at a read-only, no-modification review point—validate the quality of the summary first, then talk about automation.
How to do it: Send the expense policy document to the Bot along with a message (drag it into the input box as an attachment), then send this Chinese-localized version of the official Expense manager starter prompt:
Based on the expense system and the attached expense policy, generate this week's expense summary.
Match receipts from the finance mailbox, flag missing categories or policy exceptions, and draft a follow-up for the owner of each missing item.
Return the summary and drafts; do not send messages, and do not modify expenses.If you only want it to look at a certain app, type @ in the chat to attach the app to this task.
What you should see: The Bot reads each data source one by one and returns three things—a summary for the week, a flagged list of exceptions, and one follow-up draft per person with missing items; throughout, it sends no messages and touches no expense reports.
Step 4: Make every exception carry a policy citation, and reconcile totals back to the source
Why: The official guidance explicitly requires after the starter prompt: every exception must have a policy citation, and totals must be reconcilable back to the source—this is the hard standard that distinguishes AI expense reconciliation from "just having AI summarize things," and it is the foundation for being able to explain your basis during an audit.
How to do it: Check the Bot's output item by item: Does each exception cite the original policy text? Does the summary total match the numbers in the expense system? If it falls short, send it back to redo (using the "exception criteria feedback" magic prompt below saves effort), iterate until it can pass review without you watching, then lock in the workflow:
Save the complete expense reconciliation workflow you just did as a skill named "Weekly Reconciliation."
Requirements: Summarize this week's data from the expense system, and the total must match the numbers in the system;
Match receipts from the finance mailbox; attach an original policy citation to every missing item or exception;
Generate one follow-up draft for the owner of each missing item; do not send messages, and do not modify expenses.What you should see: The Bot confirms the skill has been saved. The official six-step method requires testing it with a second input—next week's reconciliation is a natural second test; after running it, check the citations and totals again.
Afterwards, ask for it by name in any conversation—just say "run the Weekly Reconciliation skill" to invoke it (@ is for Bots, groups, routines, and apps, not for skills). The private skill library is shared across all your Bots, so your other Bots can use it too. For saving details, see 3.2 Skills and teaching.
Step 5: Schedule the weekly reconciliation as a weekly routine
Why: The value of reconciliation is showing up at the same time every week—all you need to do is read the conclusions and release the follow-ups.
How to do it: Use one plain sentence to have the Bot create the routine (the official example uses exactly this phrasing):
Every Friday at 4:00, run the "Weekly Reconciliation" skill: summarize this week's expenses from the expense system,
match receipts from the finance mailbox, and post the exception list with policy citations and the follow-up drafts to this chat.
If the expense system or the mailbox cannot be connected, report the failure directly—do not pad the results with stale data.The Bot will first confirm 6 items with you: which Bot is responsible, the schedule and time zone, input sources, expected results, approval boundaries, and what to do when a data source is missing. Once created, open the routine and check When to run—it shows the saved schedule in plain language; if it differs from what you intended, click Edit to have the Bot change it.
What you should see: The routine appears in this Bot's routines list. Don't miss these three official details:
- A new routine only runs at the next occurrence: a "every Friday 16:00" created on Friday at 16:10 runs for the first time next Friday—to verify immediately, click Test
- The time zone follows Settings → Bot → Timezone (Auto-detect follows the local machine)—if the 4 o'clock seems wrong, check here first
- Routines run in the cloud, so closing your laptop doesn't affect them
For more scheduling details, see 3.3 Routines.
Step 6: Hold the approval line—follow-ups are drafted automatically, sending is not
Why: The last step of the official six-step method: external actions with consequences always stay behind approval.
How to do it: You don't need to do anything—just keep the defaults. Auto Review is on by default, outgoing emails produce a draft card first by default (Needs approval), and emails sent by routines are always saved as drafts waiting for you to click Send. "Modifying expenses" is explicitly forbidden in the official starter prompt—whether to delegate that, wait until the reconciliation has run stably for a few weeks.
What you should see: Each week at the scheduled time you receive a summary and a set of follow-up drafts, which only go out after you release them one by one. One detail that's easy to trip over: approval cards triggered by a routine expire if not handled within 30 minutes (ones initiated in the chat wait indefinitely)—so don't schedule actions that need your sign-off at times when you're offline. For approval rule details, see 3.1 Approvals and Auto Review.
📋 Magic Prompts
📝 Weekly Expense Reconciliation (Official Starter Prompt Structure Version)
Expected outcome: The Bot returns a three-part deliverable — a weekly summary, an exception list with policy citations, and one follow-up draft per person with missing items; the total matches the expense system.
## Role
You are my expense manager, reconciling my expense reports every week.
## Task
Summarize this week's expenses in the expense system, match receipts from the finance mailbox, flag missing categories and policy exceptions, and draft one follow-up per owner of missing items.
## Input
### Required
- Expense policy document: 【attached with the message; state limits, categories, and approval rules】
- Reconciliation period: 【e.g., this Monday through Friday】
### Optional
- Receipt scope: 【only match receipts from a certain sender or label?】
- Exception criteria: 【which cases must be flagged, e.g., over the limit, missing invoice?】
## Output Requirements
Three parts: ① weekly summary (total traced back to the source) ② exception list (each item with a citation of the original policy text) ③ follow-up drafts (one per owner of missing items, stating what is missing and how to fix it).
## Constraints
- ✅ Every exception includes a citation of the original policy text; explicitly state any data source you cannot read
- ✅ The summary total must match the figures in the expense system
- ❌ Do not send any messages
- ❌ Do not modify any expense reports📝 Missing-Item Follow-Up Drafts (Receipt Chase Version)
Expected outcome: One draft per colleague with missing items, clearly stating what is missing, which policy applies, how to fix it, and the deadline — you release each one individually.
## Role
You are my expense manager. You have just finished this week's reconciliation and are now drafting missing-item follow-ups for me.
## Task
Based on the exception list, draft one follow-up message per owner of missing items, asking them to submit the missing materials.
## Input
### Required
- Exception list: 【the missing-item entries flagged in this week's reconciliation】
- How to submit: 【which mailbox to send receipts to, deadline date】
### Optional
- Tone: 【e.g., polite but firm?】
- CC rules: 【which entries need to CC the direct manager or finance?】
## Output Requirements
Each draft includes: recipient, what materials are missing, the applicable policy clause, how to fix it, and the deadline.
## Constraints
- ✅ One message per person, no mass sends
- ✅ Cite the original policy text; no empty phrases like "per regulations"
- ❌ Do not send; save everything as drafts for my review
- ❌ Do not modify any expense reports🔧 Exception Criteria Feedback (Tuning the Judgment Standards)
Expected outcome: The Bot updates the exception judgment criteria based on your feedback and writes them into the skill; next week's exception list is noticeably more accurate — fewer false positives, and no real exceptions slip through.
## Role
You are my expense manager. You have just sent me this week's reconciliation results.
## Task
Based on my feedback, update the judgment criteria for "what counts as an exception and what counts as normal," and write them into your weekly reconciliation skill.
## Input
### Required
- Misjudged entries: 【list the false positives and false negatives, and state the correct judgment】
- Policy basis: 【the corresponding passages of the original policy text】
### Optional
- New points of attention: 【cases not flagged this week but to be flagged going forward?】
## Output Requirements
First restate the new criteria you have summarized; update the skill only after I confirm.
## Constraints
- ✅ Write the criteria into the skill, effective starting next week
- ✅ Only adjust the exception judgment criteria; do not change the output format or approval boundaries
- ❌ Do not omit real exceptions just to reduce reports
- ❌ Do not send any messagesCheckpoints ✅
- The mailbox and spreadsheet apps show Installed (not Needs auth / Disconnected) in Connect apps, and the expense system can be opened in the Bot's browser
- The total in the first summary matches the figures in the expense system, and every exception carries a citation of the original policy text
- No messages were sent and no expense reports were modified throughout
- The reconciliation workflow has been saved as a skill and can be invoked by name (e.g., "Run weekly reconciliation")
- The routine's When to run shows the schedule you want, and follow-up drafts are sent only when you click Send
Pitfall Reminders
❌ The receipt is clearly in the mailbox, but the Bot says it can't find a match
Check three places in order: ① If an app shows Needs auth / Disconnected, it isn't logged in — click Authenticate (on mobile it's Authorize); ② Gmail connects only one mailbox at a time — if receipts land in another mailbox, disconnect the current one first, then authorize that one; ③ If an attachment format can't be read, the official recommendation is to convert it to PDF, CSV, or plain text before reading.
❌ The exception list has no policy citations, just "possibly over the limit"
This means the hard standards weren't written into the instructions. Use the "Exception Criteria Feedback" magic prompt to bounce it back for a few rounds, and confirm that "every exception includes a citation of the original policy text" has been written into the skill — the official requirement for this role is policy citation on every exception; an exception without a citation is as good as unreported.
❌ The total doesn't match the expense system
Most likely a data source couldn't be reached and the Bot patched the numbers with stale data. When creating the routine, the Bot will ask "what to do if a data source is missing" — choose to report failure instead of letting it silently degrade; also hard-code into the routine instructions "if a connection fails, report the failure directly; do not patch with stale data."
❌ The routine runs too frequently and burns a week's usage in one day
Every routine run consumes usage, and the official docs explicitly warn: running once an hour or triggering in busy channels can burn a week's quota in a single day. Once a week is enough for weekly reconciliation; also don't pair a high-frequency trigger like "new email arrives" with a full reconciliation — run receipt archiving and reconciliation summary separately.
❌ No reconciliation arrived at 4 o'clock on Friday, so you thought the routine was broken
A new routine waits until the next scheduled time to run: a "every Friday 16:00" routine created on Friday 16:10 runs for the first time next Friday. To verify immediately, click Test. If it's still off, check the time zone: Settings → Bot → Timezone (Auto-detect follows the local machine).
Advanced Tips (Optional)
- Send receipts directly to the Bot's mailbox: Add an alias to the Bot's dedicated mailbox, e.g.,
你的名字+receipts@mail.grokbot.com(the +alias is just a label; mail lands in the same inbox). Have colleagues and vendors send receipts here, then set up an email-triggered routine so the Bot archives them on arrival — note that new email does not wake the Bot automatically; you must set a routine; filtering works only by sender or domain, not by subject. See 2.6 Bot's dedicated mailbox. - Archive the weekly reconciliation report as Pages: Have the Bot also organize each week's reconciliation results into a Pages page (officially positioned as readable pages like reports and dashboards), stored in the Bot's Library — at month's end, review four weeks of exception trends without digging through chat history. See 2.2 Files and reviewable results.
- Swapping the data sources gives a new scenario: With the same six steps, swap the expense system, email, and finance sheets for CRM, product usage, support tickets, and billing, and you get a ranked list of risk and expansion signals for A5 customer health.
Lesson Summary
- ✅ One "expense manager" Bot + three types of data sources (expense system, finance mailbox, finance sheets + expense policy document), producing a weekly AI expense reconciliation with policy citations
- ✅ Start with the official starter prompt: read-only with no modifications; a three-part deliverable of summary, exception list, and follow-up drafts, stopping at the review point
- ✅ Two hard standards: every exception carries a citation of the original policy text, and the total is traced back to the source — iterate until reviewable, save as a skill, and test with the second week's input
- ✅ Schedule it as a weekly routine: cloud execution arrives on time; sending messages and modifying expense reports always stay behind drafts and approvals
| Key item | Value of this lesson |
|---|---|
| Bot role | Expense manager — weekly reconciliation and missing-item follow-up |
| Connected apps | Expense system (connector or browser login), Gmail/Outlook, Google Sheets, or OneDrive/SharePoint |
| Reconciliation criteria | Your expense policy document |
| skill | "Weekly reconciliation" (name it yourself, call it by name, shared across all your Bots) |
| routine | Every Friday afternoon 4:00 (example); for the time zone, see Settings → Bot → Timezone |
| Approval line | Follow-up emails start as draft cards; emails sent by routines are always saved as drafts waiting for you to hit Send |
Next lesson preview
Next lesson A3 Sales outreach: still the same six steps, but swap the information sources for your CRM, intent sources, and email, letting the Bot research accounts and draft reviewable outreach for you. You will learn:
- Have the Bot research accounts in a CRM view, scoring them against your ideal customer profile and recent intent
- Pick up to three relevant contacts per account and draft emails and LinkedIn outreach following the writing-style samples you provide
- As always, stop at the review point: no sending, no adding people to sequences — schedule the overnight research routine only after the output is reliable