Bill audits for businesses

Use a strong model, because this pipeline involves judgment calls and a weak model makes expensive mistakes with real bills and real money.

After that, give it the keys for the job. In chat, tell the agent which accounts it's working with and paste the credentials when it asks: a Google Places API key from Google Cloud Console with the places API enabled and a budget cap on it, a postcard API account with Lob, PostGrid, or Stannp, and a Vercel AI Gateway key for Jev, which the next step covers. Once a client signs you'll also set up an email inbox where their bills get forwarded. On Grok bot you can also log into a website once in the bot's browser and it stays logged in.

Jev is a decision model from typesafe. The agent does the reading and the writing, and Jev is the part that decides. You give it a situation, like one bill with every line on it and the contract it should match, and ask it questions with fixed answers. It picks an answer and gives you a probability for how sure it is, from 0 to 1, and the agent only flags a line when Jev is sure.

You get Jev through Vercel AI Gateway. Make a Vercel account, open AI gateway from the dashboard, create an API key, and paste it into the agent when it asks. The model name is typesafe-ai/jev, and the agent calls it the same way it would call any other model. It costs almost nothing to run, when I tested all three of the results-based services in this series, 472 Jev decisions came to $0.0735 in total.

A Jev question for one line on a bill looks like this:

The rulebook is the list of overcharge types, each with what the line gets compared against. This is the one I used:

One Jev call covers a whole bill. For every line it answers whether it's an overcharge and which type, and for the bill as a whole whether it's worth disputing and roughly how much it saves. A line only gets flagged when Jev is more than 0.8 sure, anything under that goes to your review list, and the agent adds the flagged lines into a monthly and a yearly number.

Comparing against the published fee table is what keeps the flags honest. In my test there was a $170.08 fuel surcharge that looked high but matched the waste company's published table, and Jev left it alone, and an environmental fee that looked wrong wasn't wrong either. So ask each client for their contracts early, and have the agent find the provider's published fee pages.

When a bill has flagged lines, the agent writes the letter to the provider. It's plain and specific: the account number, the bill date, each flagged line with the amount, what the contract says instead, and what you're asking for, which is a credit for the past charges and the line removed going forward. The letter goes out in the business's name, signed by the owner, and the agent keeps a copy with the bill.

Any business with trucks, a storefront, or a lot of card payments has these bills, so HVAC companies, plumbers, restaurants, auto shops, and retail stores all work. Restaurants and retail are good for card processing audits because they run a lot of card payments, and trades with trucks and yards are good for waste and fuel surcharges. Filter for a rating of 4.0 or above, at least 10 reviews, and a phone and website listed.

You don't have their real bills yet, so the agent makes one mock bill for their trade, like a card processing statement for a restaurant or a waste bill for an HVAC company, with a few overcharges planted in it. It runs the bill through Jev and builds a page showing every line with the flagged ones marked, what type of overcharge each one is, and the monthly and yearly total at the bottom.

Put their name and logo at the top, mark it clearly as mock data, and end with one line: send us one bill and we'll audit it for real, free.

Postcard APIs like Lob, PostGrid, and Stannp mail real printed cards from an API call for a dollar or two each. The front is a screenshot of the demo page with the yearly number showing, and the back says "we found the overcharges on a bill like [business name]'s. Scan to see them." With the QR to the demo page and your phone number.

The agent watches the demo links and messages you when one is opened. An owner who looked at the flagged lines is usually already thinking about their own bill, so call them that day and ask which bill they'd want checked first. The answer to how you get paid is a share of what you save them, and nothing if you find nothing.

When an owner sends a real bill, which is usually a pdf from the provider's portal, the agent reads it, asks for the contract if it doesn't have one, finds the provider's published fee table, runs it through Jev, and shows you the flagged lines and the drafted letter. Ask for the last three months of the same bill if they can get it, because it shows when a fee first appeared and how long it's been charged.

When I tested this on 20 bills for an HVAC company, 274 lines in total, run twice, Jev got 273 of 274 lines right, didn't flag a single correct line, and found $1,837.33 a month that could be recovered, which is $22,047.96 a year, the same to the cent on both runs. The one it missed was a $5.95 duplicate that Jev named correctly but only at 0.60, so it went to the review list instead of being flagged.

Once a business signs, give them an email address to forward their bills to, or connect the inbox the bills already land in. The agent reads each new bill when it arrives, pulls the contract terms you saved for that provider, asks Jev, and puts the flagged lines and a drafted dispute letter in front of you. After you approve, the letter goes to the provider, and the agent tracks every credit and refund that comes back.

At the end of each month the agent sends you and the owner a statement listing every overcharge found, what the provider credited back, and what's still open, and you bill your share from that.

Every new client is the same skill with their own contracts saved, so one agent runs all of them from one inbox. It also stacks with guides 8 and 9, because a business that's letting overcharges sit on their bills usually has unpaid invoices and lost disputes too, and all three get paid out of money the owner didn't know they had.

Once every step above works on its own, this is the whole pipeline in one message. Paste it in, fill in the brackets, and the agent saves it as one skill you can rerun on a new city with a single line.