The best AI lab in business is not for sale—you already own it

Instead of spending billions to buy an AI transformation, grow one instead.

A version of this article was originally featured in Fast Company.

The last time every client wanted to talk about the same thing at the same time, it was Y2K. It is happening again with AI, but with a key difference. This time, the money is outpacing the understanding.

As boards set their 2027 budgets, one question will follow AI spending into nearly every conversation: Why isn’t it showing up in our results? A Harvard Business Review article, “How to Move from AI Experimentation to AI Transformation,” gave a name to something many executives already sense: the “micro-productivity trap.” Companies capture small, isolated gains while the larger transformation those gains were supposed to produce never materializes.

Everyone will agree AI spending needs to show a return. Far fewer will agree on why it has not yet or how to solve this challenge.

In the absence of measurable results, the instinct is to keep spending on new tools and platforms. In about ten weeks last spring, between late April and early July, the major AI labs and their partners committed roughly $10 billion to ventures that place their engineers inside client companies. When a promising enterprise technology disappoints, the reflex is to reach for one of the oldest fixes in the playbook: Write a bigger check to someone else. Buy a more capable platform. Bring in more outside engineers. Look for the one agentic solution that can transform an entire operation.

Yet Gartner projects that more than 40% of agentic AI projects will be canceled by the end of 2027, while MIT research found that roughly 95% of generative AI pilots produced no measurable return. Intelligence platform AI4SP shows a similar trend: 76% of enterprise AI initiatives struggle to show ROI. The answer cannot simply be to buy more of what companies have struggled to turn into value.

Here is what I believe the market has wrong. The dominant enterprise AI model of buying a large, centralized solution and expecting it to transform the business is ineffective and incomplete for driving AI transformation. Most firms treat AI as a technology purchase. We treat it as a people transformation. Hand the best tools in the world to people who have no reason to care, and you have bought expensive software nobody uses. To truly unlock transformative AI, leadership needs a different approach: Give employees the access, skills, guardrails, and incentives to do it — and then create a system that can recognize, reward, and scale what works.

The care that drives AI innovation inside a company has to be grown, and you do that with your own people. While outside technology providers have an important role to play in supplying the models, infrastructure, and technical expertise, what they cannot supply is the intimate knowledge of how work gets done inside a company: where time disappears, where information gets trapped, which handoffs routinely fail, which judgments matter most, and which seemingly minor task creates problems everywhere else. That knowledge resides with the people doing the work.

Before we spent big on anything new at BDO, we listened. Working with our technology partner AI4SP, we ran an anonymous diagnostic across our firm, benchmarked against more than 538,000 workplace assessments spanning 17 industries. The results reshaped many of my assumptions about our AI usage. Eighty-seven percent of our people were already using AI, most of them on their own initiative, and more than 11,000 ideas came back on how BDO professionals could use AI to improve their work and processes.

Almost nobody asked for another platform. Rather, they asked for help with the work that sits between them and providing more direct, better service to their clients. They wanted AI that helps take on the many small things that rob people of time that could be spent on people development, client relationships, and deep work. 

The findings from our internal survey fundamentally changed our approach. Our people were already identifying applications, testing tools, and teaching themselves. The internal innovation engine existed. We just needed a formal program to empower our people and harness their capabilities and talent.

So, instead of searching for panaceas, we started focusing on empowering our colleagues to become the innovators. Our approach is neither a top-down mandate nor a purely grassroots initiative. At BDO, we call it “guided grassroots”. We gave our people access, training, and guardrails, then built the support needed to turn isolated experimentation into an organizational capability. 

One core element of this strategy was to create a network of AI champions embedded across the firm. These champions are helping colleagues identify useful applications, experiment within our guardrails, evaluate the output, and share successful approaches with other teams. 

The champions gave us a way to connect and harness bottom-up experimentation across the firm, but there was still a risk that experimentation would remain isolated at a company of more than 14,000 professionals. To combat this, we made the experimentation visible across the firm through a dashboard that showcased working agents, who built them, who else is adopting them, and the value they return. 

Before our first formal training session had even taken place, the dashboard showed that more than 600 of our people had already built over 1,200 working agents. 

Two weeks after training an initial group of champions, the number of agents built had increased by roughly 80%, and the number of builders had grown by more than 50%. Within 90 days, the number of agents built and the number of people building agents and automations had increased by over 500%. Just as important, employees were creating agents their colleagues valued. Nearly one in five agents was being adopted more broadly as people saw meaningful time savings in their work. From that early activity, we also identified a dozen agents with the potential to scale firm-wide, and those champions began to train others.

Behind those numbers are people I could not be prouder of.

When our brand language changed, updating nearly 2,000 design files used to take 20 people close to a month. Emma, one of our creative professionals who will tell you she is not a programmer, used AI to shape an automation that updates the designs in about a day. It only took one person. The other 19 professionals are still on that team. The month they used to lose to rework now goes back into the creative work they were hired to do.

Melani, an AI engineer in our forensics practice, is helping transform how investigations, compliance, and discovery work are delivered. By designing AI agents for forensic workflows and developing client offerings around those capabilities, Melani represents a powerful success story: AI not only improving productivity but winning new work.

In our Tax practice, Cassie, a principal, saw an opportunity to apply AI to some of the manual steps involved in preparing tax returns. She shared the idea with senior tax manager Yvonne, who built Tara, an AI-powered quality check agent, in just two days. Agent Tara searches PDFs, checks figures against workpapers, traces sources, and drafts notes, leaving the tax professional to apply their expertise and judgment. 

Within weeks, 45 colleagues had begun using it, stating that its outputs were as good as their own. Our IT team scaled it, and a couple of weeks later it was being used by over 300 professionals. As I look forward, I’m anticipating $10 million in projected added capacity in the processing of 100,000+ tax forms a year. 

Our dashboard is full of hundreds more stories just like these. 

Finding meaningful ways to help people effectively use AI is the first challenge. Keeping what people build active is another. AI4SP’s research finds that without the internal capacity to continually maintain and improve agents, more than 60% of users stop engaging with them within three months. Today, 64% of what our people have built remains in active use each month. The rest includes experiments that served their purpose and agents that failed to find a lasting audience. Both outcomes matter in bringing us one step closer to true AI transformation. Seeing what persists and what dies off also allows us to understand where employees are finding genuine value, recognize problems shared across the firm, and direct further investment toward the applications with the greatest potential to scale.

To be clear, we are not suggesting uncontrolled adoption. This approach requires clear leadership and strong governance. At each stage, leaders need oversight that determines which tools employees can use, what information those tools can access, and how outputs must be verified. A human must always own and sign off on the work. 

Employee-led innovation is the operating model I believe companies now need. Give employees room to discover where AI belongs. Build a network that can spread what they learn. Support it with an enablement team that removes roadblocks and ensures it is connected to strategy and business outcomes. Then measure what persists, what creates value, and what deserves to scale.

Outside providers will remain essential sources of tools, infrastructure, and technical expertise. But they equip the lab. They are not the lab. They cannot replicate the knowledge within your workforce or create the internal capacity to turn that knowledge into a continuing stream of valuable applications. The best AI lab in the business? You already own it.

As boards ask why AI spending is not yet showing up in results, they should examine more than what their companies have purchased. They should ask whether they have built an organization capable of discovering what AI is actually good for and then scaling what they find.