AI Heard Every Customer Call. Then the Company Fixed Why People Were Calling.
Happen, a $1 billion bank, opened AI to nearly all its people on one guarded path, put it to work on every customer call, and fixed the reasons people called. Operations handled 26% more loan volume with 20% fewer people, mostly by not refilling jobs. Here is how they did it, the traps they stepped around, and why every team needs a custom AI of its own.
The Operator
Name & Title
Scott Sanborn, Chief Executive Officer
Company
Happen, Inc.
Ticker
Nasdaq: HAPN
Revenue
$998.8 million (FY2025)
Headquarters
San Francisco, CA
Years in Role
10 years (CEO since 2016)
Industry
Digital banking and consumer lending
Founded
2006 (as LendingClub)
Public / Private
Public (Nasdaq: HAPN)
The 60-second version
Before you buy another AI tool for a team, find out why every customer contacted that team last month. Then fix the reasons you can.
Happen, a $1 billion bank, got AI working across nearly the whole company in about a year. It built one guarded path to the AI models, capped what each person could spend, and freed its own people to apply it. The biggest results came where the AI was fitted to one team’s real work: scoring every customer call so the company could fix why people called, writing the notes after each call, and checking loan documents in minutes instead of a day. In the third quarter of 2026, operations handled 26% more loan volume than a year earlier with 20% fewer people, credited to AI and automation together. AI use doubled over the summer while the AI bill went down.
The idea to take home: every team needs a custom AI for that team, tailored to how it works, the way a suit is tailored to the person who wears it.
In this issue
- What happened (3 min)
- How they did it (9 min)
- Questions most are afraid to ask (2 min)
- The Craft of AI read (3 min)
- The Workbench: one hour, no software (1 min)
- AI Patterns: what 21 issues show (1 min)
What happened
At the end of October, Scott Sanborn, Happen’s chief executive, told investors what a year of AI work had done to his company’s operations. “Together with our other AI tools, we supported 26% higher loan volume this quarter with 20% lower operations headcount and delivered another quarter of record efficiency,” he said. At the efficiency of a year earlier, “we would have had over 200 additional heads in operations to process the volume.”
Happen is a digital bank in San Francisco. It started in 2006 as LendingClub, took its new name in June 2026, and brought in $998.8 million in revenue in 2025 with 1,075 employees. Most of its customers come for a personal loan, many of them to pay off credit card debt, and more of them now use its savings and checking accounts too.
The pressure on the operations team was growth. Loan originations rose 33% in 2025 and kept climbing in 2026. More loans mean more phone calls, more applications that need a person to check something, and more paperwork after every call. The routine part of lending was already automated; the work left over was the part that needed people, and that part was growing with the business.
The usual answer would have been to hire. Happen tried something else. It opened AI to the whole company through one controlled path, and by July about 90% of employees were using it regularly. It had AI listen to every customer call and score it, so the company knew why each person called and could fix the cause. It had AI write the notes agents used to type after each call. It gave the simplest calls, like a balance check, to an AI agent named Penny. And it pointed AI at the slowest exceptions in lending, the applications where a person had to check a document by hand.
Sanborn described what the listening did on a fintech podcast in October: “because AI monitors all the calls and all the issues, we’re able to identify things that are causing questions and friction and move upstream and fix them. So we just don’t get calls.”
The people side moved slowly. As staff left operations, managers asked whether AI could cover the gap. “We certainly are seeing as onesie, twosies when people leave, there’s an increasing sense of we don’t really need to backfill that role because we can cover the gap with AI,” Sanborn told analysts. The company kept hiring elsewhere for new products, and its total headcount rose during 2025.
The bill went the other way from what most leaders fear. “AI usage roughly doubled between July and September,” Sanborn said. “Yet, by ensuring employees are using the right tool for the job, we were actually able to lower our total AI spend.”
How they did it
What was broken
Happen was no stranger to automation. It has used machine learning to decide on loans for close to two decades, and in April 2025 Sanborn said new members got a loan process that “requires no human intervention 86% of the time.” Its call center had used older forms of AI for years. Lukasz Strozek, then the chief technology officer, told investors in November 2025: “We’ve been using neural networks in our models and natural language processing in our call center for years, but it’s really with generative and agentic AI that we’ve seen an explosion in internal uses of AI at LendingClub.”
The company also owned its own software, which mattered later. Its credit team could change lending rules without waiting on engineers. “We’ve architected our system so that credit analysts can fully self-serve the changes, eliminating engineering bottlenecks,” Strozek said. They made over 630 changes in 2023 and 2024 “without engineering’s involvement.”
What remained broken was the work that still needed a person, and it sat in three places.
The calls. Quality checks in the call center were done the way most companies still do them. “The old way you did QA in a call center was you sampled a few calls against a checklist,” Sanborn said in April 2026. A supervisor heard a handful of calls and coached from those. Everything outside the sample went unheard, so nobody could say with confidence why most customers had called.
The paperwork. Every call ended with notes an agent had to write up before taking the next one. In a growing bank, the typing added up.
The exceptions. Nine in ten loans went through with no person involved. The tenth often stopped for a document. Steve Mattics, the chief lending officer, described the cases: “a self-employed borrower where we have to verify their income or a recent mover where their address doesn’t match what’s on the credit bureau.” A person checked those documents by hand, and it took more than a day. Every day a good applicant waited was a day they could take a loan somewhere else.
What they changed, step by step
1. They built one guarded path to AI before they asked anyone to use it. Happen built one controlled way in. Sanborn describes “a gateway in place that allows guardrailed, monitored, controlled, measured access to the models. Those models are connected to our internal data sources and tools.” On the October call he said the company had “put the infrastructure, training, controls, and governance in place to enable safe, model-agnostic, and cost-controlled connectivity to our internal tools and data.” Model-agnostic means the company can switch between AI models from different makers without rebuilding anything.
The trap this avoided is the one most leaders worry about first: people pasting company data into whatever AI tool they found on their own, with nobody able to see it. Because everything ran through one path, the company could monitor and measure how AI was used. Then it encouraged everyone to use it. By July about 90% of employees used it regularly, for work that Sanborn listed as running “from the mundane, like drafting emails or creating presentations, to more complex tasks like building and evaluating financial models, conducting compliance reviews, developing marketing campaigns, and dramatically reducing the time it takes to onboard new partners.”
2. They capped the bill from day one. AI is billed by how much you use it, and a whole company using it can run up a bill nobody planned for. “We were concerned about token costs before it was cool to be concerned,” Sanborn told FinAi News in July. Happen set a cap on how much each employee could use and steered people to the right tool for each job. That is how use could double while total spend fell.
3. They freed their own people to apply it. Sanborn’s approach is to “identify the internal champion and kind of create capacity for them. Backfill them, if you will, so that they’ve got the ability to focus more on how to apply it in our environment.” He named engineering, compliance and audit among the places doing this. By October the company had 38 AI champions “representing each of the business areas.”
4. They had AI listen to every call, then went after the causes. This is the move that changed the most. AI now hears and scores every member support call. “You know exactly why everybody called,” Sanborn said. The quality team’s job changed with it. “The job becomes much more analytical, and also much more incumbent on them not to give individual coaching, but to actually identify real customer bottlenecks and friction points.” Each one they find is a reason to call that the company can remove.
Sanborn was honest that this cost effort before it saved any: “there are places where near term it creates more work in order to create less work down the road.” Listening to every call produces a pile of findings, and someone has to fix each one.
5. They took the typing off the agents, and measured it the careful way. AI now writes the notes after each call and helps agents during it. Before they called it a success, they measured it. The company compared the 12 weeks before the after-call AI went live in August 2025 with the roughly 12 weeks after. By the second quarter of 2026, Sanborn said, “AI servicing tools have contributed to a 65% reduction in after-call work and a 10% reduction in average call time.”
6. They gave an agent the simple calls, and only those. Penny, Happen’s AI member service agent, takes routine requests. “Like when somebody just wants to check their balance, we don’t need a person for that,” Sanborn said. He says Penny resolves 30% more calls than Happen’s older system did. Everything that is not simple still goes to a person.
7. They piloted AI on the slow exceptions, then widened it. For the documents that used to take a day, Happen added AI that checks the document is what was asked for, checks that it is real, pulls out the data and feeds the decision. “From what used to take over a day to now we can do in minutes and seconds,” Mattics told investors in November 2025. It started as a pilot. “Based on the pilot we’re running right now that we’re about to roll out,” Mattics said, the company expected about 9,000 more loans a year from faster checks. The finance chief, Drew LaBenne, said in January that the tool went “into test in Q4” and was “expanding across the board.” Happen has not yet reported how many extra loans it produced.
8. They wrote down what AI may not do yet. Some of Happen’s newer agents, for fraud and for writing code, are still in testing, the coding ones “with a human in the loop.” The company has a team set aside to think about outside AI agents acting for customers, and for now it keeps them out of customer accounts. “Where possible, we are blocking today until things like identity, authority, permission and liability are worked out,” Sanborn told American Banker in November. “If somebody gets a loan and then says, ‘Oh, I didn’t mean to get the loan, that agent did it for me,’ who’s at fault here?”
9. They let departures do the shrinking. Happen announced no cuts tied to AI. When someone left operations, managers asked whether the gap could be covered with AI, and more often the answer was yes. Sanborn framed the choice: “we could either do the same amount of things with less people, or we can keep the same amount of people and do more.” Where the company needed new skills, it hired, including a team for a home equity product planned for next year.
Most of the people who made this work are not named anywhere: the call-center associates who worked through the change, the quality team that turned call scores into fixes, the lending staff who handle the exceptions, and the 38 champions who took on a second job inside their first.
What it produced
26% more loan volume, 20% fewer operations staff. Third quarter of 2026 against the same quarter of 2025. The company credits Penny, call scoring and “our other AI tools” together, and LaBenne credits “our automation, efficiency, and AI initiatives.” Compensation expense rose 6% while loan originations rose 26%.
Over 200 jobs that did not need filling. Sanborn’s estimate of the operations staff it would have needed at last year’s efficiency.
65% less after-call work, 10% shorter calls. Second quarter of 2026, when the call center ran with 10% fewer staff than a year earlier while loan volume grew nearly 30%.
Document checks in minutes instead of more than a day, now rolling out across the business.
Over 90% of loans issued with no person involved, up from 86% for new members in early 2025. Applications for debt consolidation loans also take nearly 60% less time to submit.
AI use doubled while the AI bill fell, July to September 2026.
About 90% of employees using AI regularly, on one path the company controls.
Some of this is still early. I don’t know how far Happen’s call volume has fallen. The company has not published that number, a customer satisfaction figure, or what share of calls Penny takes. Fewer calls also has more than one cause: Mark Elliot, the chief customer officer, said in November 2025 that the mobile app had brought “a 20% decline in calls from our members who are digitally active.” The 20% headcount figure is operations only; total employees rose from 1,002 to 1,075 during 2025 as the company staffed new products. And the gains show up in profit more than revenue: third-quarter revenue was slightly lower than a year earlier, mostly because of an accounting change and rising rates, while pre-tax income hit a record $76.8 million.
What I do know: the operations results are real, measured over defined periods and repeated quarter after quarter. Sanborn’s own summary in July was that “In some cases, the results have been profound,” and that the company is “still in the early innings.”
Questions most are afraid to ask
Did people lose their jobs? Fewer people work in Happen’s operations, and most of that came from jobs left empty when people left. Operations headcount was 20% lower than a year earlier. The company as a whole grew, from 1,002 employees at the end of 2024 to 1,075 at the end of 2025, because it was hiring for new products. Happen has not said either way whether anyone was laid off. LaBenne also declined to promise that AI savings will cover the new hiring: “I don’t want to commit that it’s going to offset that growth yet at this point.”
Can’t we just buy everyone an AI tool and be done? You can, and in a sense that is where Happen started: one path to the models for everyone. But the general tool is a suit off the rack. It fits most people most of the way. Happen’s large results came where AI was fitted to one team’s actual work: call scoring for the quality team, after-call notes for the agents, document checks for lending. Each was built for one job, connected to the company’s own data and tools.
What did it cost, and will the bill run away? The total is not public; Happen has not disclosed its AI spend. It has said it capped each employee’s usage, matched tools to tasks, and saw usage roughly double from July to September 2026 while total spend went down. The bigger cost was effort. Listening to every call made more work before it saved any.
What happens when the AI is wrong with a customer? Happen keeps its agent to the simple calls, sends the rest to people, and scores every support call, so a problem shows up in the data before it shows up in a complaint letter. Agents for fraud and coding are still in testing, the coding ones with a person in the loop. Its annual report names the risk plainly: AI in lending “can implicate potential fair lending issues.”
We’re not a tech company. Can a company our size do this? Happen owns its software and has its own engineers, which most mid-sized companies don’t. The management choices need neither: one approved path to AI, a cap per person, one insider per team with time to apply it, a habit of finding out why customers call before automating the answer, and a written list of what AI may not do yet. And building software for one team’s work is getting cheaper fast, because AI now writes much of the code.
The Craft of AI read
My read
Most companies that bring AI to customer service ask how to answer calls more cheaply. That is the wrong question. Happen asked why people were calling at all, used AI to find out on every call, and removed the causes. The cheapest customer call is the one nobody needs to make. Klarna, in our 700-agent story, led with an agent that answered customers and later brought people back.
The lasting lesson is in where Happen’s results came from. The company-wide path gave everyone access. My read is that the big numbers came from AI tailored to one team’s work, on data the company already owned, in a company that had freed insiders to do that tailoring. Generalize that, and every team needs a custom AI for that team.
Think of the AI tools you can buy today, the general assistants and the setups builders use on top of them, as suits off the rack. They are good suits. They fit most people about 75% of the way (my rule of thumb), and your team spends the next year tugging at the sleeves by hand. A custom AI for that team is the tailored suit: cut to how operations, finance or marketing actually works, connected to your own data and rules, and owned by you.
What changed is the price of the tailoring. Custom software used to mean a large team working for years, so mid-sized companies bought what was on the rack and lived with the fit. Now AI writes much of the code, and the cost of building a tool for one team’s work is falling toward zero. The tailored suit is becoming the affordable one.
A custom AI for that team. Our read, for your Monday meeting.
- Built around this team’s real work, and owned by an insider from the team with time freed to run it.
- Runs on the company’s one approved path to the AI models, then tailored to this department, the way a suit is cut for the person wearing it.
- Used for company work, never personal AI accounts.
- Reaches only the data and systems you have approved, and someone reviews how it is used.
- A usage cap for each person, so the bill cannot surprise you.
- No decision that binds the company, and no customer matter that can go wrong, until a named owner says the controls are in place.
- If one use goes wrong, turn that use off.
What could have gone wrong, and what they did
Everyone using AI anywhere, with company data: Happen built the guarded path, the training and the controls before it encouraged anyone to use them, and capped usage per person.
An AI agent doing something a customer didn’t mean: Happen blocks outside agents from customer accounts until the questions of identity, permission and liability are settled, and keeps its own agent on simple calls.
The Workbench
Do this tomorrow. One hour, two people: whoever answers your customers (calls, tickets, email or a shared inbox) and whoever owns the cost of answering them. A tailor measures before cutting. This is the measuring.
Pull 50. Take the last 50 customer contacts in order. Don’t pick the interesting ones. Write down, in plain words, why each person got in touch.
Rank the reasons. Group them, count them, and write the top five on a whiteboard with their counts.
Mark each reason. For each one, ask whether you could remove the cause: a confusing letter, a missing status update, a form that asks the same thing twice, a policy nobody explained. Mark it “remove,” “answer simply” (the balance-check kind of question) or “needs a person.”
Put a number on it. Take the contacts you marked “remove,” scale them up to a month, and multiply by the minutes each one takes. Write that number on the board. It is the time your team spends answering questions your own business is causing.
If you want to map the workflow behind your top reason before any AI touches it, our free guide, The Claude Opus Workflow Interrogator, walks you through it.
AI Patterns
What the companies in 21 issues of The Brief have in common. Running counts, updated each issue.
- AI layoffs reported: none of 21. Of the 10 issues that say what happened to the people, 5 moved them to new work, 3 stopped refilling jobs as people left (Happen joins Ariat and Klarna), and 2 removed work without saying how.
- Started with one use, one site or one step: 13 of 21. Five went broad from day one, and 2 of those pulled back. Happen is the fifth; it built the guarded path and the spending caps before it opened AI to everyone.
- Fixed or rebuilt the process before the AI went in: 10 of 21. Happen automated its routine loan work first instead.
- Pointed AI at back-office work before customers: 15 of 21.
- Owned by the chief executive, founder or owner: 17 of 20 rollout stories.
- Hit a real setback in the work: 4 of 21, and all 4 kept going.
Want this done for you?
You just found the five reasons your customers contact you and put a number on the ones you could remove.
My Ground-Up Workshop is the rigorous version of the same exercise, run on your business with the small group of your people closest to that workflow. Every AI strategy you’ve been sold starts with software. This one starts with the people who do the work.
One to two days in person, then a synthesis week, and you walk out with a plan: a ground-truth map of how your core workflow actually runs, the one where a rebuild would release real operating margin, a target operating model with AI built in from the ground up, and a 30-60-90 day plan your own team can start on. I do a small number of these a quarter, for $20,000.
I’m a Partner at geniant. Since 2003, our craftspeople have done one thing: understand how work actually happens, design how it should happen, and build software that works the way people do. When a workshop gives you the roadmap to execute or turns into a build, they’re who does it: one senior-led team, layered on top of the systems you already run, value in weeks rather than months.
If you already know where you want to start and have an AI prototype in mind, my team can build it with you: a working version, layered on the systems you already run, so you can see the value before you commit to more.
— Grant K. Baldwin