Start expense AI where the rules already exist

Start where the rules already exist

Receipts pile up. Approvals stall. Exceptions sit in inboxes that are already full. Expense is repetitive, data-rich, and you can say what "correct" looks like. That is why it is a better first AI job than a vague transformation program.

A simple loop: submit, check against policy and vendor lists, pay what is clean, flag the rest with a plain-language reason. A person still decides the odd ones. After the system is right for a while, you can loosen review on the boring categories.

Pick one bucket, travel or software. Time approvals now. Time them again in 30 days. That comparison is the only business case you need before you expand.

Fraud is pattern-spotting, not an accusation

Duplicates across reports. Vendors not on the list. Weekend charges when nobody was traveling. Round numbers with no itemization. Amounts that hover just under the threshold, again. The model surfaces those. A person looks. Clean history, at least a year, makes the baseline real. Gaps in the file become gaps in the flags.

Most shops have spend data. Few have it in one place. Ask the pile: which departments run hot, which vendors crept up, where policy breaks cluster, what next quarter looks like from the trend. A language model with a structured export can draft that summary. Incomplete input still produces an answer. It just will not be one you should trust.

Four weeks, then another category

Weeks 1-2: export 90 days, name the three slowest approval steps. Weeks 3-4: run one of those through a tool you already have, compare speed and misses. Month 2: a second category if the first held. Month 3: an AI spend summary vs the one your team already writes.

Track hours, errors, cost per transaction. Keep humans in until the numbers say otherwise. ERP feature, Make, Zapier, Claude, ChatGPT. Start with the task that eats the most hours this month.