How Fuse On replaced manual data entry for 100,000 utility bills a year

‍

The Challenge

Fuse On pays utility bills on behalf of property management companies, most of them running single-family rentals, with a smaller book of multifamily and commercial properties. About 100,000 bills a year still arrive by mail. Staff open the envelopes and scan each bill, and until Fuse On rebuilt its platform the next step was a person reading that scan.

Todd Plambeck at Fuse On describes the job without the title on it. “We call them bill processing specialists. Data entry clerk is effectively what they were,” he says. “They would look at the PDF image and manually extract the data from the PDF, put it into the system. Then once they had completed that, they would approve the bill for payment, we would pay the bill, and move on to the next bill.”

That model grows with the client list. A new property manager means more mail, and somebody has to key it. “If you process more bills, you need more people to process more bills,” Todd says. “More headcount, more expense.”

He had tried to automate the same work before. At his previous company, where he spent ten years on commercial utility accounts for customers like railroads, the team tested several OCR products and dropped all of them.

We had tried several OCR solutions, and of course this was in the 2010, 2015 era, and most of them were, what’s the best word I can come up with? Horrible. Mainly because it was all manual and there was no AI involvement at all. So we abandoned that.

Fuse On came back to the problem last year as part of that rebuild, moving from a legacy app on a Mongo database to a new web application on a relational one. The plan was to parse each bill and load its fields straight into the new database, with enough data points per bill that Fuse On’s own system could cross-check them and approve payment without anyone opening the PDF. “Here’s the bill. All the data points match up. Pay the bill,” Todd says. “Not quite instantaneously, but pretty damn close to it.”

‍

The Solution

Two of Todd’s colleagues chose Sensible. He wasn’t part of that decision, and he had doubts about the part of the setup that asks for a configuration per bill layout.

Quite frankly, I was rather dubious of Sensible, that we had to make templates. Just throw the document at AI and let it figure it out itself. The problem there is, depending [on] whatever AI model you’re using, it would apply reasoning, and it would pontificate and make all these esoteric judgments about things.

He came round once he looked at where Fuse On’s mail comes from. Around 80% of its bills come from about 50 utilities, the large electric and gas companies plus a few big city governments billing for water and sewer, and a bill from any one of them comes in the same layout month after month. That made the configurations, which Todd now calls “the things that I thought were foolish,” a good fit for most of the mail. “Having the ability to define [what] the bill looks like every month, you can say, ‘This is what it looks like. This is where the data[’s] at. Get it, bring it back,’” he says. “And it’s extremely consistent data retrieval.”

When a utility redesigns its bill, the fix stays inside that utility’s configuration. “If they move something, add something that we want, we can go in, alter the template to accommodate for any changes, and it works nearly 100% every time going forward.”

No sorting before the API call

Fuse On doesn’t sort the mail before it reaches Sensible. Each scan goes in as it is, and Sensible works out which configuration it belongs to. “We have, like, a bill from Reliant, an electricity company, and we have bills from Atmos, we’re just sending those bills in, and then Sensible is able to determine that, oh, this is an Atmos bill, this is a Reliant bill,” Todd says. “It uses, what’s the word? Fingerprints.”

Fuse On does run a couple of small parsing engines of its own, for mail that carries no bill data at all: a warning that a water meter is spinning like there’s a leak, a “Welcome to Georgia Natural Gas” letter, the minutes of a city council meeting. Fuse On skims those and passes the information on to the client, and Todd says that is where “we almost need the reasoning.” It is also slow. “Even doing that very cursory glance at the documents takes it between five and eight seconds.”

Whereas the Sensible parsing is, I haven’t actually timed it, usually I count stuff in milliseconds, is extremely fast. We just fire documents at it, and they come back within a second or two, if that.

One schema across 1,000 vendors

The 100 configurations in production cover the high-volume utilities. Bills from the remaining vendors go to a general LLM-based configuration that returns the same schema, so Fuse On’s application gets the same fields whichever path a bill took.

We have 100 templates, but we have hundreds and hundreds of vendors. Over 1,000 vendors actually right now that are running through Sensible. Those that do not have a predefined template use, and I’m assuming it’s an AI-powered model, and even those come back faster than mine, so I’ll have to figure out how that works.

The vendor list runs from very large to very small. “One of the biggest vendors we have might have several thousand bills a month,” Todd says. “Whereas we might have a town of Slippery Rock, Missouri, where we get one bill every month, and that’s all. And it’s able to consume those and return the data to us.”

Working with the Sensible team

Sensible’s managed services team builds and maintains Fuse On’s configurations, and the review happens in a shared Slack channel. “We generally communicate over Slack. He’ll say, ‘Hey, I’ve got these templates done if you want to approve them,’” Todd says. “I go in, look at them and approve them.”

Corrections travel the same way. A utility bill can print several addresses, and a configuration would sometimes return the wrong one. “We’re like, ‘No, no, no, that’s not the right [one]. Pull this address.’ And once we define that this is the data we need, he’s able to quickly change it,” Todd says. “The team we’ve worked with has been just great. [They’re] very responsive for our requests.”

The schema itself changed once Fuse On saw real output. Todd calls the first version “kind of esoteric, what I think I want.” The team added fields, dropped some that turned out to be redundant, and in one case asked for a change to all 100 configurations at once, because late fees show up on bills in several forms.

Hey, we really need to pull if there’s a late fee on a bill, and it may come in half a dozen different flavors. If the word late fee or past due shows up on the bill, then we need to pull it too. So then [Sensible’s engineer] went in to all 100 templates and made those changes and updated them all.

“It’s pretty simple to make those changes and get the results we want back,” he says.

‍

The Results

Bills now go to Sensible as soon as they are scanned into Fuse On’s document storage. For the first stretch Todd held the pipeline back, because he had the rate limit wrong.

Originally I thought you said that I could send five a minute. So I was throttling pretty severely. And then [Sensible’s engineer] said, “Oh, no, no, it’s five a second.” So we are firing as soon as we can get them scanned and loaded into our Supabase doc storage, they immediately fire over there.

On accuracy, the misses he sees start with the paper. “We get some bills that are literally postcards that came from a dot matrix printer, and some of those are pretty horrendous,” Todd says. A bad scan of one of those can cost a character. “Sometimes Sensible will miss, like, a dash in [an] account number. Nothing critical. But for the most part, all the data points are spot on.”

Asked what still worries him, he says most of his concerns “have gone away now that we’re actually pumping bills through it.”

Speed and accuracy were obviously concerns, and both of those concerns proved to be unfounded. So it’s working great.

Fuse On’s contract covers 200 configurations and it has used 100. Todd has already raised the rest with his COO: “We’ve got another 100. Maybe we need to move further down to that list of what the higher volume vendors are.” After that come the bills Fuse On held back at the start, where a single PDF carries several properties and several accounts. “With the confidence that we’re getting,” he says, “I can certainly see us sending some of those.”