AI in the Wild: How 3 Teams Turned AI Into Real Operational Wins

The three projects have little in common on the surface. But a theme ran through all of them: none of these teams treated AI as something to set and then walk away from. Every system kept a human checkpoint, even after the manual work disappeared.

Freeing Up Frontline Focus With AI

Hernandez's team started by studying what call center agents actually did during a customer call, and it was more than solving the customer's problem. Agents were also taking notes and trying to tag the conversation with the correct contact reason — out of more than 1,500 possible categories — all while the call was still happening.

"We had over fifteen hundred different variations of contact reasons that they should select from," Hernandez said. "And knowing which one was the right one was always difficult." The process was manual, inconsistent and error-prone, and it pulled attention away from the member on the line.

The team built a classifier that used an LLM to review call transcriptions and apply the contact reason automatically. That work led to a second, unplanned project: Hernandez's team was now spending significant time verifying the classifier wasn't hallucinating and was performing as expected. So they built a second AI pipeline just to audit the first one. That system "groups the recurring failures, it catches the errors, it also does its testing and presents the findings to my team" instead of requiring a person to review every case.

Getting there wasn't simple. Hernandez said every model decision comes down to three levers: speed, accuracy and cost. A smaller, cheaper model produced fast results but weaker accuracy; a larger model that could handle the complexity meant a higher price tag. Once prompt engineering hit a ceiling, the team turned to retrieval-based approaches and batch processing to manage cost. They also had to build and continuously maintain a "golden data set" — a reliable source of truth to measure the classifier's accuracy against, even as the underlying contact reasons kept changing.

Buy-in came from questioning the taxonomy itself, not just the model. "Sometimes we think like you could slap AI on every solution and that's going to be the end all be all," Hernandez said. "But good data in is good data out. And if you have bad data in, you're going to get bad data out." That led the team to simplify their 1,500-plus contact-reason list rather than asking the model to work around it.

The results included call-handling savings and improved after-call work, but Hernandez was clear about what mattered most: giving time back to agents so they could focus on the member, not the paperwork. "Every time we're automating tasks, we didn't really remove the checkpoint, we moved it," he said. "AI is a multiplier for clearing busy work, but the human in the loop is what makes that multiplier actually trustworthy."

That balance shows up directly in how the audit pipeline works today. Instead of trusting the system to push changes straight to production, it generates a weekly report flagging where regressions or test failures occurred, so Hernandez's team knows exactly where to focus. "That's where we're striking the balance," he said, "trying to leverage LLM as a judge and different capabilities to be able to expand and take on more without introducing any risk to the business."

Automating Supply Chain Planning With Agentic AI

Ghonge's team works with thousands of suppliers and partners, and the data needed to forecast demand for each of them was scattered across CRM software, project management tools and Excel sheets — with no consistency in how it was labeled from one system to the next. "There was no control over the data because the data pipelines are coming from different softwares, with different owners and users," he said.

Reconciling all of it manually took Ghonge 16 to 20 hours a week, at a point when the business was still relatively small. "That was crazy for me," he said. So he started experimenting, first with an Excel prototype to test whether the approach could work. It helped, but not completely. That's when he turned to agentic AI, building the system himself with a background in dashboarding and product requirements but no formal engineering training.

Adoption started small — Ghonge was the only user — and grew through informal channels: business review calls, lunch-and-learns, word of mouth. Four months later, the platform had more than 75 users and thousands of views.

The results were substantial. What used to take 20 hours now takes two, an 80% reduction in time to generate a demand forecast, alongside a 40% reduction in forecasting error. "The biggest one is like 80% reduction in time required to generate demand and a high accuracy," Ghonge said. "We also saw about a 40% reduction in the forecasting [error]." The system also gave the team visibility they never had before — the ability to break down exactly which supplier attributes were driving thousands of monthly changes, rather than guessing.

Trust didn't come from the AI label. It came from transparency. When project management data exported as 18,000 rows that actually represented just 600 to 700 real tickets, Ghonge's system parsed and consolidated that mess into a clean table users could review directly, rather than digging through the source tool ticket by ticket. "Instead of going to the tool every time for every ticket, now they're able to easily see everything in one system," he said. Users could also pull up any previous week and compare it against the current one, watching forecasting errors trend down over time — proof the system was actually working, not just automating.

Ghonge also credited AI with something he hadn't originally planned for: documentation. Once he started asking the AI to track his own deployments and time spent, he was able to quantify his own build process after the fact — down to the number of prompts it took to get there.

Building an AI-Powered Digital Operations Engine

Banakh's team set out to reimagine the entire vehicle sales journey — how vehicles are brought to market, valued, merchandised, matched with buyers, transacted, fulfilled and transported — as one connected digital process instead of the industry's traditional model: a customer bringing a car to a brick-and-mortar auction once a week and hoping it sells that day.

"We've recently built a full intelligent digital end-to-end workflow for one of our core partners, including: vehicle data intake, inspection, valuation, merchandising, buyer discovery, transaction, fulfillment," Banakh said. "It's really all morphed into one continuous process."

The hardest part wasn't the technology. It was the industry itself. "The human factor is notoriously digital adverse in this industry," Banakh said. "I still see dealers walking around with pen and paper for documentation." That meant the project required more than a process change — it needed what Banakh called a paradigm shift in core behaviors, which is easy to get wrong.

With a tight timeline, Banakh's team used AI to simulate a range of stakeholder personalities and talk through how the workflow should come together. "We ended up basically co-creating the draft of the workflow," she said. "And then that was brought to our partner for validation. And it really expedited conversations" — replacing slow, one-by-one validation meetings with a draft the partner could react to directly.

The project is still in early pilot, so Banakh could only share limited numbers, but the top-line result stood out: cycle time has already dropped by two-thirds, before the system has even been fully optimized. "The cycle time is reduced by two-thirds right now, and we haven't even optimized fully," she said, adding that reductions in cost, asset depreciation and human labor are expected to follow.

Getting skeptical stakeholders on board didn't require a hard pitch. "You speak to the rationale and ROI, and true business people are converted when they see the result," Banakh said. For dealers who prefer to keep doing things the old way, that's fine too — the system is built for partners who want to scale nationally, not replace every workflow that already works for someone.

The Common Thread

Different industries, different tools, different scales — but the same underlying discipline. Hernandez's team built a second AI system just to keep the first one honest. Ghonge built visibility into his platform so users could verify the data themselves instead of taking the output on faith. Banakh's team kept a "warm handoff every step of the way," with a program manager accountable for catching anomalies human review would otherwise miss.

None of these teams got buy-in by pitching AI as a concept. They got it by shipping something small, proving it worked and letting the results make the case. As Banakh put it, "true business people are converted when they see the result." That's the pattern behind all three case studies: less about the AI itself, and more about the discipline of building trust into the system alongside it.

Want more insights like these? Sign up for a free DPM account to hear from more experts like these and be the first to know about upcoming events.