Catch billing leaks and rogue SaaS spend with an AI expense review

A cancellation-and-renegotiation list built from twelve months of transactions, surfacing every duplicate tool, silent price increase, and charge nobody remembers signing up for.

What you'll have when you're done

A ranked list of software and subscription spend you can act on this week: the duplicate tools (two CRMs, two file-storage apps), the prices that crept up while nobody was looking, and the recurring charges your team cannot explain. You hand it to whoever owns the renewals, and a meaningful slice of your software budget comes back. Most companies waste 25 to 40 percent of their SaaS spend; this is how you find your share of it.

The spend that leaks is the spend nobody is looking at

Here is the pattern in almost every growing company. Someone expensed a tool eighteen months ago for a project that ended. The annual renewal auto-charges every year. The vendor raised the price 15 percent and sent an email nobody opened. Multiply that by every tool, every team, every card, and you get a number that would genuinely annoy you if you ever saw it in one place. I have run this on my own companies and been annoyed every time. One pass turned up a tool we had been paying for since before a reorg that nobody had logged into in over a year. It was not a rounding error, it was real money, every month, for nothing. Studies put average SaaS waste at a quarter to forty percent of spend, with companies carrying around seven-and-a-half duplicate licenses on average.

You never see it in one place because it is scattered across statements and nobody's full-time job is to comb them. That comb is exactly what an AI does well: hand it a year of transactions and it groups every recurring charge by vendor, spots the duplicates, and flags the increases, in the time it takes to read this paragraph. There is a nice irony worth naming, the AI token bill itself is now a line creeping onto company P&Ls, so this same review will catch your AI spend too.

What you need first

• Twelve months of transaction data, exported from your spend platform (Ramp, Brex), your card statements, or your accounting system's vendor report.
• A Claude Project on a business plan. Spend data can include vendor and account detail; keep it off personal accounts.
• A willingness to ask the team. The AI finds the charges; confirming whether a tool is actually unused is a quick human check (covered below).

Step-by-step

Step 1Export a year of transactions

Pull twelve months of charges as a flat file. If you are on a spend platform like Ramp or Brex, export the vendor-level transaction history; otherwise use your card statements or your accounting system's vendor report. A full year matters because the worst offenders bill annually, the once-a-year charge is the one that hides best.

Step 2Ask for the four-part audit

Drop the export into your finance Project and prompt:

Review this 12 months of transactions and return a software/subscription audit:
1. Recurring charges grouped by vendor, with annual run-rate per vendor.
2. Likely duplicate tools (e.g., two CRMs, two storage tools, two design tools): group them.
3. Price increases: any vendor charging more than the same period last year, with the % jump.
4. Charges I might not recognize: recurring vendors that aren't obviously core software.
Rank the output by annual dollar size. Do not assume a tool is unused; flag it as "verify usage."

Here is the shape of what comes back, illustrative, for a 40-person company, ranked by annual run-rate:

Duplicates (act now)

• Two CRMs: HubSpot ($28K/yr) and a legacy Pipedrive ($4.8K/yr) still billing → kill Pipedrive, save $4,800
• Two storage tools: Dropbox Business ($9K/yr) and Box ($6K/yr) → consolidate

Price creep (renegotiate)

• Notion: $14K this year vs $9K last → +56%, mostly seat growth; audit seats
• Datadog: $31K vs $24K → +29% with no plan change

Verify usage (ask the team)

• Loom · $3,600/yr · 12 seats, invoice unchanged for 14 months
• Webflow · $2,400/yr · one project, unclear owner

Don't recognize

• "CLOUDXYZ INC" · $890/mo recurring since last March · no obvious owner

Step 3Separate "definitely cut" from "ask the team"

The output splits into two piles. The definitely-cut pile is duplicates and price creep you can act on directly: two tools doing the same job, a vendor that quietly raised rates. The ask-the-team pile is the charges the AI flagged as possibly unused, because the AI can only see that you are paying for a seat, not whether anyone is using it. Send that shortlist to the relevant team leads with one question: "do we still need this?"

Walk the example above through to dollars, because that is the part that makes a CEO actually do this. The definitely-cut pile, the duplicate Pipedrive plus consolidating Box and Dropbox down to one, is about $9,000 a year you can reclaim this week. The ask-the-team pile comes back with Loom genuinely dead (nobody had opened it in months), another $3,600. And the unrecognized $890-a-month charge turns out to be a sandbox a former contractor spun up and never closed: $10,680 a year for nothing. Total found in one twenty-minute pass, a little over $23,000 of annual run-rate, against a software bill the CEO had assumed was "about right." None of it was hidden on purpose. It was just scattered across statements where no single person ever saw it in one place.

Step 4Turn it into a renewal calendar

For the tools you keep, ask the model to build a renewal calendar from the charge dates: which renewals land in which month, and which are annual (your renegotiation leverage points). Now the silent auto-renewal becomes a scheduled decision instead of a surprise.

Step 5Re-run it quarterly

SaaS sprawl is not a one-time cleanup; it grows back. Put a quarterly 20-minute review on the calendar. Each pass catches the new tools that crept in and the renewals coming up, and keeps the waste from re-accumulating.

Make it yours. If you are on a spend platform like Ramp or Brex, the export is clean and the audit is sharp out of the gate. If you are not, the spend is spread across two or three cards and a PayPal account, and the win is just as big but the prep takes a few more minutes: export each source separately, label which is which, and tell the model they are different cards for the same company so it does not treat the same vendor on two cards as a duplicate. Companies without a spend platform usually have more waste, not less, precisely because the spend was never visible in one place to begin with.

How you'll know it's working

A real dollar number comes back the first time you run it, and you can act on a chunk of it within a week. The longer-term tell is that "wait, we pay for what?" moments stop happening, because the quarterly review catches them before they compound. Track annual software spend per head; if it flattens or drops while headcount grows, the review is working.

When it breaks

• It flags charges as duplicates that are not. Often annual-versus-monthly billing of the same vendor showing up oddly, or two legitimately different tools. Tell it which vendors bill annually.
• It calls a tool "unused." It cannot see usage, only charges. That is why those go to the team as "verify," never straight to "cancel." Cutting a tool the AI guessed was idle is how you break something someone depended on.
• A "missing month" gets flagged. Usually an annual subscription that only charges once. Not a leak, just a cadence the model misread.
• The numbers do not match your statement. Re-export clean detail; formatted statements with subtotals confuse the grouping.
• It groups two tools as duplicates that you run on purpose. Sometimes the overlap is intentional (a CRM for sales, a separate help desk for support). Tell it which overlaps are deliberate so it stops re-flagging them every quarter.
• The annual run-rate looks wrong for a tool you just adopted. A subscription that started mid-year shows only a partial year of charges, and the model may total what it sees instead of annualizing. Tell it the start date, or ask it to annualize from the current monthly rate rather than summing the year.

Where this fits in your harness

This is the finance workflow with the fastest, most concrete payback, real dollars, usually in the first run. It pairs naturally with closing your books faster (clean vendor data makes the audit sharper) and feeds the cash-flow forecast, since killed subscriptions change your outflows. It also sets up the governance conversation: a surprising amount of "rogue" SaaS is shadow AI and unsanctioned tools your team adopted on their own.