AI Adoption Lessons: What Early Movers Got Right and Wrong

Four research-backed lessons for boards and executives from companies that went in early on AI.

Introductory Note and Some Personal AI Learning

My social media feed today included this Thread (image above).

It got me thinking: 'is this true - is there massive AI buyer remorse out there?' Also, 'what are the lessons - good and bad - from AI early adopters that might help everybody getting into AI?'

To help me get good answers I went to Claude (with a little help from ChatGPT).

I have no qualms about going to AI for research on topics like this - using AI is consistent with the mission of my site and almost every time I use it I learn something new - often something worth sharing.

What you should know in this case: some important answers Claude initially returned to me were not true - they were fabricated and couldn't be verified. Claude even admitted they were made up.

So Claude and I had a serious chat - basically, 'dude, you need to get your shit together; you need to do proper research and you need to help me think, not make me feel good.' I also changed the AI thinking model I was using to a more intense level (recommended by Claude).

Several iterations later I am satisfied that the examples and the lessons below are legitimate, verifiable - and helpful learning for Boards and executive leaders.

The top line on company AI adoption: there is buyer remorse, success and key learning

As researched and verified by AI (Claude and ChatGPT), here are four key lessons that emerge from companies that have gone in big-time on AI. Interestingly, none of the lessons are about the technology itself.

Lesson 1: AI Replaces Volume, Not Judgment

AI performs well on high-volume, predictable, rule-bounded tasks. It struggles with variability, nuance, and the judgment calls experienced employees make automatically. Several well-resourced companies learned this lesson the hard way.

In 2023, Klarna replaced approximately 700 customer service employees with an AI chatbot built with OpenAI. The bot handled the volume — reportedly managing two-thirds of all customer service requests. Resolution times dropped from 11 minutes to under 2 minutes.

For a while, the financial projections looked strong. Then customer satisfaction dropped as complex interactions, emotionally charged cases, and multi-step problem resolution consistently overwhelmed the system.

By May 2025, Klarna's CEO Sebastian Siemiatkowski told Bloomberg the company was reversing course and rehiring human agents. "From a brand perspective, a company perspective, I just think it's so critical that you are clear to your customer that there will always be a human if you want," he said. By September 2025, Business Insider reported Klarna was reassigning internal workers to customer support roles.

McDonald's ended its IBM drive-thru voice AI pilot in July 2024 after three years of testing across more than 100 restaurants. Social media had documented repeated ordering errors — the system misheard customers, added unwanted items, and confused adjacent lane orders. In a memo to franchisees obtained by CNBC, McDonald's chief restaurant officer wrote: "While there have been successes to date, we feel there is an opportunity to explore voice ordering solutions more broadly."

Starbucks scrapped its AI-powered inventory management tool after just nine months. The system, deployed across more than 11,000 North American stores, repeatedly miscounted and mislabeled items. Reuters reported it failed to identify bottles on shelves. A Starbucks promotional video captured a peppermint syrup bottle going unregistered while the system scanned bottles on either side of it. Baristas returned to counting inventory by hand.

A major Pizza Hut franchisee, Chaac Pizza Northeast, filed a lawsuit in May 2026 alleging that Pizza Hut's mandatory AI delivery management system — called Dragontail — caused over $100 million in lost business and enterprise value across its 111 locations. According to the lawsuit, delivery times went from under 30 minutes for 90% of orders to over 45 minutes for half. Year-over-year revenue growth in New York inverted from over 10% to approximately negative 10%.

The case is pending; Pizza Hut has declined to comment on the litigation.

The pattern across all four cases is consistent: AI deployed to replace human judgment — rather than to augment it — produced results that required expensive reversals.

The lesson that holds up: AI is a force multiplier for capable people. It is not a substitute for judgment and deploying it that way has been costly.

Sources: Klarna — Bloomberg, May 2025; Entrepreneur, January 2026; Business Insider, September 2025; CX Dive, September 2025. McDonald's — CNBC, June 2024; Restaurant Business, June 2024; AP, June 2024. Starbucks — Fortune, May 2026; Reuters, May 2026; Engadget, May 2026. Pizza Hut — Fortune, May 2026; Futurism, May 2026; Tom's Hardware, May 2026.

Lesson 2: Successful AI Adoptions Solved Real Business Problems

Goldman Sachs published a research note in March 2026 analyzing AI's productivity impact across corporate America. Their core finding: companies that targeted AI at specific, well-defined tasks reported a median productivity gain of approximately 30%.

At the economy-wide level — where AI is deployed without that specificity — Goldman's senior economist Ronnie Walker wrote, "We still do not find a meaningful relationship between productivity and AI adoption."

Four companies illustrate what problem-first deployment looks like in practice.

In 2018 Bank of America asked this question: how do we handle billions of routine customer interactions without proportional headcount growth?

The question drove the development of Erica, their AI-powered virtual assistant. As of August 2025, Erica has logged over 3 billion client interactions. Nearly 50 million customers have used it. Erica now handles 2 million consumer interactions daily — saving the bank the equivalent of 11,000 staffers' daily work, according to Holly O'Neill, president of consumer banking. The internal employee version has cut IT service desk calls by more than 50%, and over 90% of BofA's workforce now uses AI tools in their daily work.

Walmart started with two specific operational problems — delivery route inefficiency and product catalog accuracy — and deployed AI against each one separately. AI-powered route optimization eliminated 30 million unnecessary delivery miles and avoided 94 million pounds of CO2 emissions. The project won the Franz Edelman Award in 2023, and Walmart has since commercialized the technology as a SaaS product for other businesses. On the catalog side, GenAI improved over 850 million product data points — a task CEO Doug McMillon said on an August 2024 earnings call would have required 100 times the headcount using manual processes.

Duolingo asked a content scaling question: how do we launch courses in dozens of new markets without proportional headcount growth in content creation? AI-generated content allowed the company to introduce nearly 150 new language courses for different geographies in Q1 2025 alone.

That efficiency — alongside AI-driven personalization of the learning experience — helped Duolingo expand its adjusted EBITDA margin to 25.7% in 2024 while growing bookings 32% year over year. CEO Luis von Ahn described the approach in the company's Q1 2025 shareholder letter: the testing infrastructure explores hundreds of ideas at once, measures their impact rigorously, and lets the best ones shape the product.

Each of these companies had a measurable problem before they had an AI solution. That sequencing is the common thread — and it explains why Gartner predicted that at least 30% of generative AI projects would be abandoned after the proof-of-concept stage by the end of 2025. The primary reasons Gartner cited — poor data quality, escalating costs, and unclear business value — are all symptoms of organizations that started with AI and went looking for a problem to attach it to.

The question that separates the two groups: What specific problem are we solving, and how will we know if we solved it?

Sources: Goldman Sachs research note, March 2026 (via Yahoo Finance); Bank of America press releases February 2025 and August 2025; SEC filing FY2025; The Financial Brand, January 2026 (Holly O'Neill quote); Walmart — AI News, December 2025; Supply Chain Dive, October 2025; CEO Doug McMillon earnings call, August 2024; Duolingo SEC filings Q4 FY2024 and Q1 FY2025 shareholder letters; Gartner press release, July 2024.

Lesson 3: AI Scales Bad Data, it Doesn't Fix It

When S&P Global Market Intelligence surveyed over 1,000 enterprises in 2025, 42% had abandoned most of their AI initiatives — up from 17% just one year earlier. The average organization scrapped 46% of its AI proof-of-concepts before they reached production.

Rationale for abandoning AI initiatives varies. Informatica's CDO Insights survey found data quality and readiness cited by 43% of organizations as the primary obstacle. BCG found 74% pointing to data governance and accessibility. BARC's independent research found data quality issues more than doubled as the top obstacle year over year — from 19% in 2024 to 44% in 2025. The common thread across all of them: the infrastructure underneath the AI was not ready for what was being asked of it.

Wipro's 2025 State of Data4AI report, cited by the World Economic Forum, puts a number on the gap: only 14% of business leaders say their data is mature enough to support AI at scale, yet 79% say AI is essential to their company's future. More than half of business leaders in the same study acknowledged using inaccurate or inconsistent data to guide key decisions — before AI was even in the picture.

"Mature data" means something specific: data that is consistent in format across systems, accessible to the tools that need it, governed so people know who owns it and how it can be used, and clean enough that an AI model drawing on it produces reliable outputs rather than confident-sounding errors. AI does not fix bad data. It scales it.

Aviva, the UK's largest general insurer, demonstrates what getting this right before deployment looks like. Working with McKinsey's QuantumBlack unit, Aviva deployed over 80 AI models across its motor claims operation. A team of more than 50 data scientists, engineers, business leaders, and change professionals built the data infrastructure and trained claims teams before the AI tools went live. Crucially, Aviva built a "double helix" approach — the claims journey can switch seamlessly between digital and human interaction, always optimizing for both business and customer outcomes.

The results, which were disclosed publicly to investors: liability assessment time for complex cases fell by 23 days, claims routing accuracy improved by 30%, customer complaints dropped by 65%, and net promoter scores climbed more than sevenfold. Aviva told investors the transformation saved more than £60 million (approximately $82 million) in 2024.

That is an infrastructure-readiness and organizational-preparation story in which AI was the final step — not the starting point.

The board question: Before the next AI investment is approved, can management demonstrate that the underlying data is consistent, accessible, and governed — and that the cost, security, and skills infrastructure can support it? If the answer to any of those is unclear, the investment is premature.

Sources: S&P Global Market Intelligence 2025 (survey of 1,006 respondents, North America and Europe; reported by CIO Dive, March 2025, and Fortune, June 2025); Informatica CDO Insights 2025; BARC AI Survey 2025 (via Bigeye); Wipro State of Data4AI 2025 (via World Economic Forum, October 2025); BCG "Where's the Value in AI?" report, October 2024; McKinsey case study: "Aviva: Rewiring the insurance claims journey with AI"; Insurance Business, April 2026.

Lesson 4: Leadership Commitment Is A Clear Predictor of AI Results

McKinsey's 2025 survey identified a small group of "AI high performers" — roughly 6% of respondents — generating more than 5% of EBIT from AI. What separates them from the other 94% is not budget, tools or industry - it's leadership.

Nearly half of high-performing organizations report that senior leaders show clear ownership and long-term commitment to AI — role-modeling usage, protecting budgets, and repeatedly sponsoring initiatives. Among everyone else, that number drops to 16%.

Here is what senior leadership ownership actually looks like in practice, at three companies where it has produced measurable results.

JPMorgan Chase: Structure, accountability, and CEO-level sponsorship.

Jamie Dimon created a dedicated AI and data function that reports directly to him. That structural decision — not delegating AI to a technology department, but making it a CEO-level function — set the tone for everything that followed. JPMorgan's annual technology budget reached $18 billion in 2025, approximately 9.5% of the firm's total revenue.

In July 2025, Dimon held a four-day executive retreat focused specifically on AI. According to a person who attended, discussion topics included how AI-driven changes would be adopted across the bank's 317,000-person workforce and how automation would affect the apprenticeship model in investment banking. At the February 2026 investor day, six C-suite leaders — CEO, CFO, Consumer Banking CEO, Asset Management CEO, and two Commercial and Investment Bank co-CEOs — each presented AI as embedded in their own business units. AI was framed as an operational capability influencing lending decisions, pricing models, customer interactions, and infrastructure modernization — not as a technology initiative sitting in one department.

Dimon's approach to workforce impact is specific. Rather than layoffs, the bank built what Dimon publicly called "huge redeployment plans" — shifting displaced employees into new roles. Overall headcount has held steady at approximately 318,500 while operations roles declined 4% and support functions declined 2%, offset by a 4% expansion in client-facing and revenue-generating teams.

The measurable results: operations teams now handle 6% more accounts per employee, fraud-related costs per unit have fallen 11%, and software engineer productivity has climbed 10%. The LLM Suite — JPMorgan's proprietary AI platform — is now used by over 200,000 employees, and American Banker awarded it 2025 Innovation of the Year.

Walmart: CEO-led planning with structural investment in leadership roles.

In July 2025, CEO Doug McMillon announced two newly created executive positions: EVP of AI Acceleration (filled by Daniel Danker, hired from Instacart) and EVP of AI Platforms, reporting to Walmart's Global CTO. McMillon announced these hires personally, on LinkedIn, with the message: "AI is changing how we work."

McMillon has been publicly specific about workforce planning. At Harvard Business Review's Future of Business event in November 2025, he said: "Every job we've got is going to change in some way — whether it's getting the shopping carts off the parking lot, or the way our technologists work, or certainly the way leadership roles change." Internally, Walmart executives have been tracking which job categories will increase, decrease, or hold steady — planning meetings that acknowledge uncertainty rather than promising a painless transition.

Every AI project at Walmart is evaluated against measurable ROI, with continuous monitoring for model accuracy — described as a "top-down approach, championed by CEO Doug McMillon." The company's Walmart Academies program trains employees on AI tools through both physical classrooms and virtual reality technology, and a partnership with OpenAI focuses specifically on upskilling the existing workforce.

Unilever: Leaders using the tools themselves.

Unilever's Personal Care division took a different approach: senior leaders personally demonstrated AI tools in their own work, visibly using them in front of teams. The result was 75% of office-based Personal Care employees becoming regular users of AI productivity tools — an adoption rate most organizations struggle to reach.

A pilot called Smart Briefing, tested by the Closeup marketing team, showed what that adoption produced: a 14% improvement in brief quality, a 26% boost in team satisfaction, and up to 58% time savings. These are modest, specific numbers — not transformational claims — and they came from a division where leadership modeled the behavior rather than mandating it.

What the three examples have in common:

Each organization made structural decisions — not just statements — about AI. JPMorgan created a reporting line to the CEO. Walmart created new executive roles and tracking systems. Unilever's leaders used the tools themselves. In each case, the results are measurable, the leadership actions are specific, and the approach treated AI as an operational priority rather than a technology project to delegate.

The board question: Who in this organization is accountable for AI outcomes — not AI activity, but outcomes? What structural decisions have been made to support that accountability? If the answer to either is unclear, that is the first thing to fix.

Sources: McKinsey State of AI 2025 (survey of 1,993 participants across 105 countries); JPMorgan Chase — SEC filing FY2025, CNBC exclusive September 2025 and February 2026, PYMNTS February 2026, Crowdfund Insider March 2026, Klover.ai analysis July 2025; Walmart — CIO Dive July 2025, Business Chief November 2025, Harvard Business Review November 2025, Fox Business September 2025, Klover.ai analysis July 2025; Unilever — Unilever.com November 2025.

What All Four Lessons Have in Common

There's no surprises here. Every one of these lessons comes back to organizational clarity: senior leaders who own AI outcomes, clear problem definition before tool selection, clean data foundations, and an honest understanding of where AI adds value and where it doesn't.

BCG found that only 5% of organizations are creating substantial AI value at scale — while 88% claim to be using AI. That is a leadership and organizational discipline gap, not a technology gap. The tools are widely available.

The discipline to deploy them well is still being developed.

Research sources and verification notes:

All statistics in this post are sourced to primary research reports or verified company disclosures. Where company results are cited, they come from the company's own press releases, SEC filings, investor disclosures, or direct executive statements to identified journalists. The Pizza Hut figure ($100M) is an allegation in pending litigation, not a confirmed loss, and is identified as such in the text.

McKinsey State of AI 2025 (survey of 1,993 participants, 105 countries) | BCG "Where's the Value in AI?" October 2024 and "The Widening AI Value Gap" September 2025 | Goldman Sachs research note, March 2026 | Wipro State of Data4AI 2025 (via World Economic Forum) | Gartner press release, July 2024 | Bank of America press releases and SEC filing FY2025 | JPMorgan Chase SEC filing FY2025 and CNBC interview, September 2025 | McKinsey/QuantumBlack Aviva case study | Fortune, Reuters, Bloomberg, CNBC, Entrepreneur, Business Insider, AP (individual citations noted in text)

Jim Crocker is an AI governance consultant and board director. He writes about what boards and senior executives need to know about AI at jimcrockerai.com.