Expense Reports and Receipts
Nobody was hired to do expense reports. A program manager takes a photo of a gas receipt, forgets about it for six weeks, and then reconstructs a mileage log from memory on the last day of the quarter. The bookkeeper chases eleven people for missing documentation while the audit clock runs. The work is small, repetitive, unloved, and it carries real tax consequences when it goes wrong, which makes it one of the better places in a nonprofit to point AI at.
Reimbursement is one of those processes that looks trivial until you total up what it costs. A twenty-person organization might process a few hundred expense claims a year, each requiring a submission, a coding decision, an approval, a payment, and a filed receipt. None of those steps takes long. Together they consume days of finance staff time, they generate more back-and-forth email than almost any other routine process, and they are a persistent source of friction between finance and everyone else.
The stakes are higher than the dollar amounts suggest. If your reimbursement arrangement fails to meet the IRS requirements for an accountable plan, the payments become taxable wages. That means payroll tax exposure, corrected W-2s, and a conversation with staff about why their reimbursement for a tank of gas is now income. Very few nonprofits fail this test deliberately. They fail it by drifting: documentation gets thin, advances are never reconciled, and the substantiation that was supposed to happen within weeks happens the following spring, if at all.
The interesting thing about expense reporting is that almost every painful part of it is a transcription, matching, or reminder problem. Reading an amount off a photograph, matching a receipt to a card transaction, guessing which grant a flight belongs to, noticing that a claim is missing a business purpose, and asking someone politely for the fourth time to send in their documentation are all tasks that machines handle well and that humans handle grudgingly and inconsistently.
What follows covers what an accountable plan actually requires, why nonprofit reimbursement breaks in ways corporate expense systems do not anticipate, precisely where AI earns its place in the workflow, the decisions that have to stay with a person, how federally funded expenses change the calculus, and a practical way to introduce this without buying an enterprise platform your organization cannot justify.
What an Accountable Plan Actually Requires
Before deciding what to automate, it helps to be precise about what the process has to produce. Under the IRS rules for accountable plans, a reimbursement arrangement has to satisfy three conditions. The expense must have a business connection, meaning it was incurred in the course of performing services for the organization. It must be adequately substantiated within a reasonable period. And any amount advanced in excess of substantiated expenses must be returned within a reasonable period. Meet all three and the reimbursement is not wages. Miss any of them and it is.
Substantiation means specific facts, not a total. The record needs the amount, the date, the place, and the business purpose. For travel and similar categories the business relationship of anyone else present matters too. The IRS treats records created at or near the time of the expense as far stronger evidence than a reconstruction assembled later, which is exactly why the quarterly catch-up ritual is a problem even when the underlying expenses are entirely legitimate.
On receipts, the general rule is that documentary evidence is required for lodging and for any expense of $75 or more. Below that threshold, other adequate records can substantiate the expense, which is why a well-kept contemporaneous log matters. Many organizations set an internal receipt threshold well below the IRS floor, which is a reasonable policy choice, but it is worth knowing which parts of your policy are legal requirements and which are house rules, because staff comply better when they understand the difference.
For timing, the IRS safe harbor treats substantiation within 60 days of the expense and the return of excess advances within 120 days as reasonable. Organizations that adopt those windows explicitly, state them in the policy, and actually enforce them have a far easier answer for an auditor than organizations relying on a vague expectation of promptness. The safe harbor is not the only way to be reasonable, but it is the cheapest way to be obviously reasonable.
What every claim must capture
The substantiation floor, not the ideal
- Amount paid, in the currency actually paid
- Date the expense was incurred
- Place or vendor
- Business purpose stated in plain language
- Who else was present, and their relationship to the work
- Documentary evidence for lodging and expenses of $75 or more
Timing that keeps the plan accountable
Windows worth writing into policy
- Substantiation submitted within 60 days of the expense
- Unspent advances returned within 120 days
- Records created at or near the time, not reconstructed
- A stated consequence when the window passes
- The same rules applied to executives and to everyone else
Why Nonprofit Reimbursement Breaks Differently
Commercial expense systems assume a salesperson with a corporate card, a laptop, and a manager who reviews a queue. Nonprofit reimbursement has a different shape, and the mismatch explains why so many organizations abandoned an expense product and went back to a spreadsheet.
The first difference is the population. Claims come from salaried staff, hourly staff, part-time program workers, board members, and volunteers, and only some of those people have an organizational email address, let alone a login to a finance system. A process that requires everyone to authenticate into a portal quietly excludes the volunteer driver who spends the most on mileage.
The second is coding complexity. A single conference trip might split across two grants, one unrestricted fund, and three functional expense categories. Commercial systems ask which cost center an expense belongs to. Nonprofits need to answer which grant, which program, which functional classification, and whether the cost is allowable under the terms of the award, which is four questions wearing one coat. Getting this right at submission is far cheaper than untangling it during the close, a point we made in more depth when discussing how to structure a chart of accounts AI can code to.
The third is the dollar sizes. Many nonprofit claims are small, and small claims produce a bad economics problem: the staff time to process a $14 parking reimbursement exceeds the reimbursement. Organizations respond either by making the process so light that substantiation suffers, or by making it so heavy that people stop submitting, absorb the cost personally, and quietly resent it. Neither outcome is good, and the second one falls hardest on the lowest-paid staff.
The fourth is timing pressure that comes from outside. Grant reporting deadlines, fiscal year end, and audit fieldwork all create moments when finance suddenly needs six months of undocumented expenses reconciled at once. That scramble is where reconstructed mileage logs and invented business purposes come from, and it is the single most common way a well-intentioned accountable plan degrades into an unaccountable one.
The fifth is that the person approving is often the person who benefits. In a small organization the executive director approves the finance manager's expenses and the finance manager processes the executive director's. That is workable if the board or treasurer reviews executive expenses independently, and it is a genuine control weakness if nobody does. Sector fraud cases are full of ordinary expense reimbursement that nobody looked at closely for years.
Where AI Earns Its Place
The useful framing is that AI should reduce the effort of producing a complete claim and increase the chance that an incomplete one is caught immediately. It should not decide whether an expense is reimbursable. Everything below sits on the first side of that line.
Reading the receipt. Extracting vendor, date, total, tax, and line items from a photograph is mature technology now, including from crumpled thermal paper, foreign-language receipts, and photos taken at an angle in bad light. This removes the most tedious data entry in the process and, more importantly, it makes submitting a claim from a phone at the moment of purchase realistic. Contemporaneous records are a legal advantage, and the way to get them is to make same-day submission take fifteen seconds.
Suggesting the coding. Given a vendor, an amount, a date, and the submitter's role and current project assignments, a model can propose the account, the program, the grant, and the functional classification. The proposal is a draft that the submitter confirms and finance reviews. Even at imperfect accuracy this is a large improvement over a blank dropdown, because most people code expenses by guessing anyway, and a reasoned suggestion beats a guess. This pairs directly with the thinking in our guide to defensible functional expense allocation.
Checking completeness against policy. Before a claim is submitted, a check can confirm that a business purpose is present and specific rather than a single word, that a receipt is attached where the policy requires one, that the date falls within the submission window, that per diem or mileage rates match the current published figures, and that attendees are named for meals. Catching these at submission rather than at review eliminates most of the email round trips that make the process feel slow.
Matching receipts to transactions. If the organization uses cards, reconciling statement lines to submitted receipts is pure matching work, complicated by merchant name mangling, tips added after authorization, and split payments. This is exactly the class of fuzzy matching where automation is dramatically better than a person scrolling two windows, and it surfaces the specific transactions that have no documentation at all, which is the list finance actually needs.
Chasing what is missing. The unloved work of reminding people is well suited to automation, and automation does it without the social cost. A polite scheduled nudge listing exactly which claims are outstanding and what each one needs is more effective than a quarterly all-staff email, and it means the bookkeeper is not personally cast as the nag.
Surfacing patterns for review. Aggregated across a year, expense data answers questions nobody has time to ask by hand. Which programs are absorbing travel that was never budgeted. Whether a particular vendor is being used repeatedly without a contract. Whether claims cluster just under an approval threshold. These are questions for a human reviewer, and the value of the analysis is that it produces a short list rather than a verdict.
Preparing the audit sample. When fieldwork begins, the auditor selects a sample of transactions and asks for supporting documentation. Having every claim already linked to its receipt, its approval, and its general ledger entry turns a multi-day retrieval exercise into an export. Our guide to audit preparation with AI covers how that fits the wider fieldwork season.
Checks worth running before a claim is accepted
Each one prevents a later email
- Business purpose present and more specific than "meeting" or "travel"
- Receipt attached wherever policy or the $75 rule requires one
- Expense date inside the submission window, with the age flagged
- Mileage and per diem calculated at the current published rate
- Meal attendees named where more than the claimant was present
- Alcohol, gifts, and other restricted categories identified for routing
- Possible duplicate of a claim already submitted by anyone
- Outstanding advance that this claim should be applied against
What Has to Stay With a Person
The approval itself is the obvious line. Approving a reimbursement is an assertion that the expense served the organization's purposes, and that is a judgment about work, context, and reasonableness that a supervisor makes. An automated approval for anything under a threshold is a policy decision the board should make knowingly rather than a default the software arrives with.
Business purpose is the second. A model can tell you a purpose field is empty or unhelpfully vague. It cannot write the purpose, and it should not be allowed to, because a generated business purpose is a fabricated record. If staff start accepting a suggested purpose they did not think about, you have automated the production of documentation that says nothing true, which is worse than a blank field because it looks compliant.
Third, allowability under a grant agreement is a reading of a contract. A model can flag that an expense is coded to a federal award and that the category is one that frequently raises questions, which is genuinely useful. Whether the specific cost is allowable, allocable, and reasonable under that award is a determination for someone who has read the terms and who will defend it in a single audit.
Fourth, anything that looks like an accusation. Duplicate detection and threshold analysis produce candidates for review, and a flagged item is a question, never a finding. The failure mode here is a report that reads as though the system has identified misconduct, circulated to people who then treat a coincidence as evidence. A staff member wrongly implicated by an automated flag is a serious harm, and the same reasoning we applied to conflict of interest disclosures applies exactly here.
Fifth, the data question. Receipts contain locations, times, health-related purchases, and personal card details, and expense records show where staff were and who they met. That is sensitive information about employees, and routing it through a third-party service deserves the same scrutiny as any other personal data. Confirm what the vendor retains, whether submissions are used for model training, and who inside your organization can see claims other than their own. Our guide to privacy risk assessment for nonprofit AI projects covers how to run that evaluation.
Safe to automate
Transcription, matching, and reminders
- Extracting amounts and dates from receipt images
- Proposing account, program, and grant codes for confirmation
- Checking claims for missing required fields
- Matching card transactions to submitted documentation
- Reminding people about outstanding claims and advances
Keep with a person
Judgment, authorship, and accusation
- Approving a claim, at any dollar amount
- Writing or suggesting the business purpose text
- Determining allowability under a grant or federal award
- Concluding that a flagged pattern indicates wrongdoing
- Deciding to pay a claim that arrived outside the policy window
When the Expense Is Charged to a Grant
Reimbursement gets meaningfully harder the moment a cost lands on a restricted or federal award, because two different rulebooks now apply to the same receipt. The IRS cares whether the payment is wages. The funder cares whether the cost is allowable under the award, whether it was properly allocated, and whether the documentation supports the charge. A claim can satisfy one and fail the other.
Federal awards bring specific expectations about documentation, cost principles, and consistent treatment of similar costs. Travel charged to a federal award generally needs to follow the organization's own written travel policy, which means an organization without a real policy has a problem before anyone opens a receipt. Categories that reliably attract questions include entertainment, alcohol, first-class travel, and anything that looks like lobbying, and these are worth flagging automatically at submission so the conversation happens early.
Allocation across multiple awards is the other recurring difficulty. A staff member who attends a conference serving three programs cannot charge the whole flight to whichever grant has room. The split has to reflect actual benefit and follow a documented, consistently applied method. Automation helps by applying the organization's chosen method uniformly and recording how each split was derived, which is the part that usually exists only in someone's head. The related discipline of tracking restricted funds and donor intent depends on the same underlying records.
One practical warning. Do not let a suggested grant code become the de facto allocation decision. If the system proposes a split and everyone accepts it because it appears first, the organization has adopted an allocation method it never chose and cannot explain. The suggestion is a convenience. The method is a policy, it should be written down, and the written version is what an auditor will ask to see.
A Rollout That Fits a Small Finance Team
The temptation is to buy a platform. For many organizations that is the wrong first move, because the platform imposes a workflow before you know which parts of your workflow are actually broken. A cheaper sequence gets most of the benefit and tells you what to buy if you eventually buy anything.
Start by fixing the policy, because automation makes an unclear policy fail faster. Write down the submission window, the receipt threshold, the mileage and per diem rates you use, the categories that require pre-approval, who approves whom, and what happens when a claim arrives late. If your policy does not currently answer those questions, no tool will.
Then reduce submission friction before adding review sophistication. The largest single gain available to most nonprofits is making it easy to submit a complete claim on the day of the expense from a phone. Receipt extraction plus a short structured form gets you there. Everything downstream improves when the inputs arrive on time and complete.
Add the completeness checks next, and let them run before submission rather than after. A check that tells the claimant what is missing while they still have the receipt in their hand is worth ten checks that tell the bookkeeper three weeks later.
Add matching and reporting last. Card reconciliation and pattern analysis are valuable, but they matter most once the flow of complete, timely claims exists. Introducing them first produces a lot of alerts about a mess that a better front end would have prevented. If you are already using AI in the close, this connects naturally to the work described in our guide to bookkeeping and month-end close.
Finally, handle volunteers and board members deliberately. They are often the people with the most out-of-pocket costs and the least access to internal systems, and many of them never claim at all. A simple emailed or texted submission path that does not require an account, combined with a clear statement that unreimbursed expenses may be deductible if they choose not to claim, treats them properly. It also stops the organization from quietly relying on volunteers absorbing costs it should be paying.
A workable sequence
Policy first, tooling second
- Rewrite the reimbursement policy so every rule has a number attached
- Make same-day mobile submission the easiest path available
- Run completeness checks at submission, not at review
- Automate reminders so no person has to chase colleagues
- Add card matching once claims arrive complete and on time
- Give volunteers and board members a path with no login
- Route executive expenses to independent review by the treasurer
Conclusion
Expense reimbursement is unglamorous, and that is precisely why it is worth automating well. The process is high-frequency, low-value per transaction, and legally consequential when it degrades. Every hour a finance manager spends reading a photograph of a receipt is an hour not spent on the work that actually requires their expertise, and every reconstructed mileage log is a small weakening of the organization's accountable plan.
The gains are concentrated at the front of the process. Making it trivially easy to submit a complete claim on the day of the expense fixes more problems than any amount of downstream review sophistication, because it produces the contemporaneous records the IRS prefers and eliminates the quarterly scramble that generates weak documentation. Extraction, suggested coding, and completeness checking are all in service of that single goal.
Keep the boundary firm on the other side. Automation reads, matches, checks, and reminds. People approve, write the business purpose, determine allowability, and decide what a flagged pattern means. Draw that line explicitly in your policy, tell staff where it sits, and you get a faster process that is also more defensible than the one you have now, which is not a trade-off you get offered often.
Make Reimbursement Take Minutes, Not Weekends
We help nonprofit finance teams put AI where it removes tedium and keep it away from the decisions that have to stay human.