Why 95% of AI Projects Fail to Deliver Positive ROI
It’s increasingly common that articles about generative AI warn us about its negative impact on businesses, what pitfalls we should look out for, and what not to do, hence my personal focus on AI readiness and what must be considered as a foundation to deliver value. I wanted to focus rather on the positives, why they are positive, and what makes case studies like these successful.
How do we Measure Success?
Too often in our 25 years of working in knowledge management and AI, we have had customers focussing solely on time and budget. A reasonable consideration as this has a direct correlation to budget and resource usage, but almost never was user adoption seen as a measure of success.
How often have we had user adoption as a success metric, or have accurate Objectives and Key Results (OKRs) per deliverable that we can measure for efficacy? I can tell you with absolute certainty that user adoption is the most critical metric, because if the solution is not adopted, then the budget spent on the project was most likely wasted, leading to aimless execution.
So, why is this relevant? Because when we see that 95% of Generative AI projects fail to yield a positive ROI right now, we need a laser focus on user adoption understanding why Generative AI is being implemented in your business. Because we truly believe that the psychology behind tech adoption needs a laser focus and understanding.
United Wholesale Mortgage—The Case Study
United Wholesale Mortgage (UWM) is America’s largest home mortgage lender trading publicly on the New York Stock Exchange (NYSE). They originate mortgage loans through a wholesale channel, setting up a nationwide network of brokers. UWM has invested heavily in technology, including an AI partnership with Google Cloud, and broker-facing tools so successful, it led to a nearly 30% market share by mid-2025. Leadership is targeting growth well beyond that.
UWM deployed AI tools, powered by Google’s Vertex AI, Gemini, and BigQuery, with an absolute focus on one high-impact use case: accelerating the mortgage underwriting process. Rather than trying to “AI everything,” they targeted the bottleneck.
But here’s the critical distinction: UWM didn’t just deploy AI internally, they built an entire AI ecosystem designed to make their broker partners more competitive. In April 2025, when UWM partnered with Google Cloud, they integrated Gemini Flash 1.5 into their proprietary technology. According to a report by Nick Manes on Crain’s Detroit Business 1, this allowed underwriters to increase their capacity from 6 loans per day up to 14, more than doubling their throughput! Successful on any metrics.
But the real story is what they built for brokers.
What they actually built
Over the course of 2024 and 2025, UWM rolled out a suite of AI tools that fundamentally changed how independent mortgage brokers operate. An incredible feat, but it takes hard work, knowing your exact use cases, and the metric for success of each.
The three systems put in place leveraging AI each had a use case
- ChatUWM: Launched in May 2024 as an AI-powered search engine grounded in UWM’s own knowledge base. Within months, it had over 25,000 active external users generating around 3,000 prompts per day.2 Brokers can upload a credit report and proof of income, and the system reads the credit report to calculate income, ask follow-up questions, and immediately recommend products that fit the borrower’s specific situation. Think about what that does: it frees up conversations with account executives to focus on business growth and strategy instead of routine questions.
- Loan Estimate Optimizer (LEO): Announced at UWM LIVE in May 2025 to all brokers in the network. This AI powered tool gives brokers an interface to drag and drop a competitor’s loan estimate into UWM’s platform, LEO then scans, extrapolates and analyses it line by line, identifying gaps and opportunities like reduced title fees, appraisal credits, or basis point improvements.3
CEO Mat Ishbia described it as having the most intelligent loan officer in America review an estimate in seconds. This is a direct competitive weapon helping brokers beat retail lenders on price in real time.
- MIA: An AI voice bot that acts as a personal loan officer assistant. Now, this doesn’t always work, especially depending on country and culture, but it did for UWM.
It helps with consumer retention by making outbound calls to check on clients, fielding questions 24/7, scheduling appointments, and taking messages. What is so important in this instance is that it can be customised with each broker’s own branding, so it feels like the broker’s assistant, not UWM’s. Psychologically, this has a big impact. MIA tackles the biggest weakness brokers have traditionally had against retail lenders: the inability to maintain ongoing client contact at scale.
Beyond these tools, UWM also launched Lead Pipeline (AI-driven alerts when past clients list their home or have their credit pulled), Action IQ (a consolidated notification centre for daily tasks), and an Easy Qualifier redesign that lets brokers create comparison breakdowns presenting up to three different products and rates.4
This is an entire ecosystem of AI components that are contextually bound, have measured success, and massive adoption.
Why? Because it reduced effort, created convenience, and helped create more success for the users.
The psychology, why adoption worked
This is where it gets interesting and ties directly back to user adoption as the most critical success metric. UWM’s approach addresses several psychological dynamics that typically kill AI adoption, as determined by a 2025 MIT Nanda report 5 showing that 95% of enterprise Gen AI pilots fail to deliver measurable Profit & Loss impact.
Let’s break down the psychology here:
- Reduced cognitive load: Most enterprise AI tools add complexity to a user’s workflow. UWM did the opposite. The psychological experience is one of relief. This aligns with cognitive load theory (CLT) 6 by removing the mental overhead of routine tasks, brokers have more capacity for the high-value relationship work.
- Agency and identity preservation. This is the most important psychological dynamic. Independent brokers chose this career because they value autonomy, being their own boss, being the trusted advisor. AI that threatens that identity gets rejected. UWM designed tools that amplify the broker’s expertise rather than replacing it. LEO doesn’t tell the broker what to recommend, it gives the broker ammunition to win a deal. MIA carries the broker’s branding, not UWM’s. The broker’s self-concept shifts from “person who looks up guidelines” to “strategic advisor who wins deals.”
- The endowment effect. LEO is psychologically brilliant because it makes the broker feel like they possess a secret advantage. The endowment effect means that once brokers feel they “own” this advantage, they’re deeply reluctant to give it up. This ensures that they retain and adopt the solution to stay ahead and retain this advantage.7
- Social proof and FOMO at scale. UWM announced these tools at a live event with 6,000+ attendees, a very clever tactic. Brokers saw their peers getting excited about the same tools. This creates powerful social proof: “Everyone is adopting this, I need to as well.” The explicit market share targets create a shared mission narrative for everyone. This is classic in-group identity building, brokers aren’t just using a tool, they’re part of a movement that’s winning.
- The path of least resistance, redesigned. In a recent article, I discussed the “path of least resistance” and how people default to it, they go back to the way they’ve always worked if the system in place creates additional work. UWM’s approach flips this. The path of least resistance is now the AI-augmented path, not the manual one.
The Net Effect
UWM hasn’t just deployed AI, they’ve redesigned the psychological experience of being an independent mortgage broker. The broker feels more capable, more competitive, more connected to their clients, and more aligned with a winning team. That’s why the adoption works: it doesn’t feel like a technology change, it feels like a professional upgrade.
This is the exact opposite of what happens when organisations deploy AI tools that add friction, threaten identity, or optimise for the company’s efficiency at the expense of the user’s experience. It’s why people are afraid AI will take their jobs, they need readiness and tools that enhance, not threaten.
So, when we ask what “getting AI right” looks like, UWM gives us a blueprint:
- Target a real bottleneck,
- Design for adoption (the end user’s psychology),
- Amplify identity rather than threatening it,
- Make AI the path of least resistance, and
- Build trust through transparency.
Do that, and you on the right trajectory to adoption.
That’s the lesson. User adoption isn’t a downstream metric, it’s THE metric! And the psychology behind it should be driving your AI strategy from day one. We need to spend more time from inception asking the question, what is the true measure of success? Delivery the enabler? Or its use to meet the desired objectives.
Reference List
Challapally, Aditya, Chris Pease, Ramesh Raskar, and Pradyumna Chari. The GenAI Divide: State of AI in Business 2025. MIT NANDA, 2025. https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf.
Kleimann, James, and Flávia Furlan Nunes. ‘UWM’s New AI Tools Analyze Rivals’ Offers, Virtually Assist Borrowers’. LendingLife, Mortgage, Origination, Technology. HousingWire, 15 May 2025. https://www.housingwire.com/articles/uwm-ai-tools-leo-mia-offer-analysis-virtual-borrower-assistance/.
Manes, Nick. ‘As Competitors Get in the Game, UWM Goes “all in” on Artificial Intelligence’. Technology. Crain’s Detroit Business, 9 April 2025. https://www.crainsdetroit.com/technology/united-wholesale-mortgage-goes-all-ai.
Nunes, Flávia Furlan. ‘UWM Beefs up AI Chatbot with Instant Feedback on Products and Comp’. LendingLife, Mortgage, Technology. HousingWire, 10 March 2026. https://www.housingwire.com/articles/uwm-ai-chatbot-provices-product-and-comp-info/.
Paas, Fred, and Jeroen J. G. van Merriënboer. ‘Cognitive-Load Theory: Methods to Manage Working Memory Load in the Learning of Complex Tasks’. Current Directions in Psychological Science 29, no. 4 (2020): 394–98. https://doi.org/10.1177/0963721420922183.
Reb, Jochen, and Terry Connolly. ‘Possession, Feelings of Ownership and the Endowment Effect’. Judgment and Decision Making 2, no. 2 (2007): 107–14. https://doi.org/10.1017/S1930297500000085.
Sweller, John, Jeroen J. G. van Merriënboer, and Fred Paas. ‘Cognitive Architecture and Instructional Design: 20 Years Later’. Educational Psychology Review 31, no. 2 (2019): 261–92. https://doi.org/10.1007/s10648-019-09465-5.
Sweller, John, Jeroen J. G. van Merrienboer, and Fred G. W. C. Paas. ‘Cognitive Architecture and Instructional Design’. Educational Psychology Review 10, no. 3 (1998): 251–96. https://doi.org/10.1023/A:1022193728205.
UWM. ‘UWM Introduces Suite of AI-Powered Solutions’. UWM.Com, 14 May 2025. https://www.uwm.com/media-alert-may-14-2025.
By Philippe Morin | Head of Advisory and Commercial at Cyber Dexterity