You text "dinner with client 85.50" from the parking lot, and by the time you reach your car the expense is categorized, matched to your card, submitted, and approved. No app to open, no form to fill, no receipt photo to chase later. That is what Expensify turned on in July 2026: expense management by plain-English text message, email, or in-app chat, plus AI agents that run multi-step approval workflows on their own. If you run a small business, the question is not whether the demo is impressive. It is how much manual entry is actually left, what the IRS still requires from you, and how to keep your books audit-ready when a chatbot does the typing.
What Expensify actually shipped
On July 1, 2026, Expensify expanded Concierge AI from a support assistant into an automation layer that takes action inside the product. You interact with it in natural language through email, text message, or the Expensify app, and it completes the work rather than only answering questions.
The new capabilities break down into five pieces:
Conversational analytics. Ask about spending by employee, category, merchant, workspace, or reporting period, and get back visual charts with explanations. Instead of exporting a CSV and building a pivot table, you ask "what did we spend on travel last month" and see the breakdown.
Chat-based expense management. Create, edit, categorize, tag, submit, approve, and reject expenses and reports by asking. Concierge can also mark expenses as cash and explain category limits when something looks out of policy.
Workspace administration. Configure policies, create and manage expense rules, add or remove members, troubleshoot account issues, and update settings by asking. For a small team without a dedicated admin, this removes the settings-panel scavenger hunt.
Cross-channel access. The same assistant works over email to a Concierge address, by text message to a short code for US numbers, and directly in the app. The channel is your choice; the underlying expense record is the same.
Autonomous workflow agents in beta. Define custom instructions for AI agents that perform routine expense management tasks on their own, with a complete audit trail of actions taken. An agent can submit, approve, reject, hold, and reroute reports, and take over as approver when given workspace admin rights. Each agent gets its own account and is added as a full-access copilot on its creator's account, so its actions are attributable rather than anonymous.
Underneath sits SmartScan, Expensify's receipt scanning technology, which extracts merchant, date, total, currency, and VAT amount, codes the expense, matches it to the corresponding card transaction when one exists, and flags prohibited purchases at the line-item level. A separate integration launched in June 2026 connects Expensify data to outside AI assistants over OAuth, so you can ask which reports need approval without exporting anything.
The company's framing is blunt: AI should eliminate work, not create another interface to learn. For a small team, the expense workflow collapses to a sentence you were going to type anyway.
How much manual entry is really left
The honest answer is 70 to 90 percent gone, with 10 to 30 percent still needing human eyes. That range comes from realistic small-team deployments, where the remainder is transactions the AI categorizes incorrectly: unusual vendors, split personal-and-business charges, and policy edge cases it has never seen. Automation amplifies whatever category mapping and spending policy you configure at setup, so a sloppy chart of accounts produces confidently wrong coding at scale.
The time savings behind that range are substantial. Manual expense tracking eats about 10 hours each month per employee. For a team handling 500 receipts a month, AI tools cut processing from roughly 16.7 hours to about 1.5 hours. Small businesses running AI expense workflows report saving around 4.3 hours per week and capturing 12 to 18 percent more deductible expenses than manual methods, largely because fewer receipts go missing. One 2025 case study found a 90 percent reduction in errors, a 40 percent cut in approval times, and a 25 percent drop in expense-related costs. Vendors publish benchmarks showing reimbursement cycles falling from 10 to 14 days down to 24 to 48 hours after switching to automated workflows. Finance teams that make the switch typically recover 60 to 80 percent of hours previously spent on data entry while cutting error rates from a manual average of 1 to 4 percent down to under 0.5 percent.
What remains is review work, not typing work. Run a weekly exceptions pass over low-confidence or unusual-merchant items, plus a monthly policy check for out-of-policy spend, duplicates, and personal charges coded as business. And keep your category map tight. Every vague category you allow ("miscellaneous," "general expense," "other") becomes a bucket the AI fills enthusiastically and your accountant has to empty at year end.
A common mistake is treating the AI's first month as representative. Accuracy climbs as it learns your vendors and your corrections. Budget 30 minutes a week for review in month one, 15 in month two, and under 10 once the mapping stabilizes. If accuracy does not improve, the problem is usually your policy, not the model: overlapping categories, unwritten exceptions, and approvers who rubber-stamp everything teach the agent that anything goes.
What the IRS still requires from you, even when AI does the typing
Automation changes who types, not what counts as proof. The operative standard is "adequate substantiation," the phrase running through IRS Publication 463. For each expense you must be able to show the amount, the date, the place or description of the purchase, and the business purpose. For meals, add the business relationship of the people involved. A vendor receipt alone rarely covers all of this, because it does not state why the expense was business-related. That "why" is the field AI is worst at inventing and auditors care about most.
Two thresholds matter. You generally need documentary evidence such as a receipt for any expense of $75 or more. Under $75, the IRS does not require you to keep a receipt, but you are still legally required to prove amount, date, place, and business purpose from your records. Lodging is the exception: hotel and lodging expenses always require a receipt, no matter how small. Transport fares and incidentals under $75 are the classic no-receipt cases, but "no receipt" never means "no record."
Timing matters too. Under an accountable plan, you must substantiate expenses within a reasonable period, which regulations treat as 60 days to provide records and 120 days to return any excess advance. Text-message capture helps here, because the message itself is timestamped at or near the time of the expense. A shoebox of receipts reconstructed in April is the opposite: amounts without dates, dates without purposes, and purposes you no longer remember.
The practical implication for AI workflows is to require business purpose at capture, not at review. Configure your tool so a text like "dinner 85.50" prompts back "who was there and what project was it for" before it submits. That one follow-up question is the difference between an adequate record and an $85.50 guess. Store the answer with the expense, not in a separate chat thread you will never find again. And keep the underlying receipt image even when the AI extracted everything correctly: extraction is convenience, the receipt is evidence.
If you reimburse employees, the same rules apply through your accountable plan. The employee accounts for each expense with receipts and a record of what it was for, you reimburse the substantiated amount, and anything unsubstantiated becomes wages. An agent that approves without checking business purpose does not make the reimbursement compliant; it makes noncompliance faster.
The AI-receipt fraud problem, and how to protect yourself
The same models that read receipts can now generate convincing fakes in seconds. AI-generated receipts accounted for around 14 percent of fraudulent documents submitted in 2025, compared to essentially none the prior year. Roughly 70 percent of chief financial officers believe employees are using AI to attempt to falsify travel expenses and receipts, and searches for AI receipt generators spiked 340 percent in a single quarter of 2025. Most bad expense reports are still honest mistakes rather than fraud: industry data puts non-compliant spend at 15 to 20 percent of reports, driven by duplicates, non-itemized receipts, and out-of-policy spending. But the fraud floor has risen, and your controls need to assume a fake receipt can look perfect.
Structural defenses beat detective ones. The most effective approach is to create fewer opportunities for a fraudulent receipt to enter the workflow at all:
Pay by corporate card wherever possible, and cross-reference. When the card feed and the submitted receipt disagree on amount, date, or merchant, hold the report. A fake receipt for a real card charge still has to match the bank's version of the truth.
Require itemized receipts, not just totals. A total-only slip is trivial to fabricate and useless for policy checks. Itemization forces the forger to invent a plausible basket of goods and gives your reviewer something to sanity-check against the venue.
Use pre-spending requests for large or unusual purchases. Approval before the money goes out turns "reimburse this surprising $2,400 charge" into "here is the $2,400 you already approved." Agents can enforce this automatically by holding any report above a threshold without a matching pre-approval.
Write the policy down, specifically. "Reasonable meals" is unenforceable; "client meals up to $75 per person, no alcohol without partner approval" is something both an agent and an auditor can apply. Well-defined policies with pre-spending controls are the single most cited prevention measure in current fraud guidance.
Go digital end to end. Paper receipts that get scanned once and discarded are the easiest to swap. A capture-to-ledger chain where the phone photo, the AI extraction, the card match, and the approval live in one record leaves far less room to substitute a file later.
For small teams, the 80/20 version is corporate cards plus itemized receipts plus a weekly review of anything the AI flagged or anything over your threshold. You do not need forensic image analysis; you need a second pair of eyes on the exceptions and a policy clear enough that the AI can enforce it the same way twice.
How to book text-message expenses in your ledger
Whether the expense arrives by text, email, or app, it still lands in the same places: an expense account for what you bought, and either a cash account or an employee-reimbursement liability for how it was paid. The AI's category is a suggestion; your chart of accounts is the decision. Map each app category to exactly one ledger account before you turn on automation, and resist the urge to create a new account per vendor. Ten stable expense accounts you reconcile monthly beat fifty precise ones you never review.
A typical small-business mapping looks like travel meals to one account, office supplies to another, and reimbursements through a liability that clears when you pay the employee. The Beancount documentation walks through this account structure in detail, and a dashboard like Fava turns the resulting postings into the same spend-by-category charts the chat assistant shows, except computed from your audited ledger rather than the app's feed.
Here is what a text-submitted client dinner looks like once it posts, followed by the reimbursement that clears it:
2026-01-01 open Expenses:Travel:Meals USD
2026-01-01 open Assets:Bank:Checking USD
2026-01-01 open Liabilities:Reimbursements:Employees USD
2026-07-02 * "Client dinner" "Text-submitted, agent-approved"
Expenses:Travel:Meals 85.50 USD
Liabilities:Reimbursements:Employees
2026-07-05 * "Employee reimbursement"
Liabilities:Reimbursements:Employees 85.50 USD
Assets:Bank:CheckingThe first transaction books the expense when it is incurred and holds the amount you owe in a reimbursement liability. The second clears that liability when cash leaves the bank. If the dinner went on a company card instead, the offset posts directly to the card liability rather than through reimbursements, and the card-statement reconciliation becomes your cross-check against the AI's extraction. Either way, the business purpose from the text thread belongs in the narration or a metadata field on the posting, because that is the piece the IRS asks for and the receipt does not contain.
One mapping decision deserves care: per diems and mileage. These are not receipt-based expenses; they are computed allowances. Book them to their own accounts rather than mixing them with receipted meals and fuel, so a reviewer can tell at a glance which lines should have receipts attached and which should have a mileage log or travel dates instead. When the AI agent handles mileage from a "drove 120 miles to client site" text, the log entry it creates is the substantiation, so confirm it records origin, destination, and business purpose rather than just the distance.
Getting from inbox chaos to audit-ready
If you are starting from a pile of card charges and a camera roll of receipts, set things up in this order:
1. Lock your categories first
Write down your ten or so expense accounts, define what goes in each with two examples and one exclusion, and map every app category to one account. Delete or disable the app's default categories you do not use. This one-hour exercise prevents a year of miscoding.
2. Require purpose at capture
Turn on the follow-up prompt for business purpose and, for meals, attendees. Make it required before submission, not optional at review. Your future self, reconstructing a February dinner in April, will thank you.
3. Set agent rules narrowly, then widen
Start workflow agents on the safest tasks: auto-approving recurring software under a fixed amount from known vendors, rerouting anything over threshold to you, and holding reports missing receipts or purpose. Review every agent action weekly for the first month. Only then let agents approve broader categories, and keep a hard rule that no agent approves its own creator's expenses above a nominal amount.
4. Reconcile monthly, not annually
Match the app's submitted totals to the card statements and the bank feed every month, and confirm the ledger carries the same expenses with the same purposes. Monthly reconciliation catches a miscoded vendor while you still remember the trip; annual reconciliation catches it while your accountant bills you to investigate.
5. Keep the audit trail readable
The agent's complete action log is only useful if a human can follow it. Periodically open a random approved report and check that you can answer who spent, what for, who approved, and where the receipt is, without opening a second system. If you cannot, tighten the narration template before you widen automation.
The throughline is that text-message expenses and workflow agents remove typing, not responsibility. Your policy decides what is allowed, your category map decides where it posts, your purpose prompt decides whether it is substantiated, and your monthly review decides whether any of it worked. Get those four right and the AI handles the rest. Your next useful step is the first one: write down your ten accounts and their definitions before you connect anything, because everything the agent learns afterward starts from that page.
