AI Transformation Begins Before Tech: Why 90% of Companies Fail to Realize Value from AI Investments

Why AI transformation begins before the tech

AI investment has expanded rapidly, with organizations introducing copilots, agents, and generative AI across functions. Yet the business transformation many leaders anticipated can remain difficult to identify. A 2026 analysis found that nearly 40% of companies tracking AI cost savings achieved less than 10%, while 90% planned to increase their budgets. Additionally, 38% of finance leaders and 39% of CEOs considered it too early to determine whether AI was delivering value. For Alina Kukarina, co-founder of Deeply Human Innovation, these figures point toward a broader leadership question. Her experience across digital transformation, software, and management training has led her to examine how organizations make decisions before technology enters the picture. She suggests that the natural reaction to disappointing results can be to examine the technology, the model, or employee adoption. A deeper issue can sit earlier in the process. Organizations may begin implementation before defining precisely what they are trying to improve. That gap can be described as a "thinking gap." Businesses typically develop financial plans, implementation schedules, and technology roadmaps, while the structured thinking that connects those elements can receive less attention. Kukarina says, "The starting question often becomes, 'Where can we use AI?' A more useful starting point would be to ask, 'What are we trying to improve, and why?'" The distinction matters because technology can accelerate an existing workflow with remarkable efficiency. If that workflow contains unnecessary steps, unclear ownership, weak data, or decisions that depend heavily on human judgment, automation can amplify issues that arise at scale. Kukarina points to a simple principle. "Process evaluation should precede AI evaluation. Organizations need to understand how work happens, where value is created or lost, and where human judgment remains essential before assigning technology a role," she states. In her approach, she considers five connected elements: Problem, People, Process, Technology, and Outcome. The problem establishes the purpose. People reveal who is affected and where judgment, expertise, and trust matter. Meanwhile, the process shows how work currently happens. Technology identifies which technology to choose and whether AI will provide meaningful assistance. The outcome defines the business and human results leadership expects to improve and needs to be defined from the outset. This sequence also creates a stronger basis for leadership decisions, including financials. AI initiatives carry costs involving software, infrastructure, training, governance, integration, and potential mistakes. Their value can extend across customer satisfaction, employee experience, service quality, and operational performance. Kukarina argues that leaders benefit from examining those dimensions together. For example, an AI-generated response may increase speed while influencing how a customer perceives the organization, while an AI-generated employee development plan may affect trust and motivation. Success metrics then deserve the same scrutiny as implementation plans. License counts, user numbers, and token consumption can describe activity and cost, while business impact requires a wider lens. Kukarina points to time saved, retention, employee satisfaction, reputation, and the quality of customer interactions as examples of indicators that can reveal whether technology is contributing meaningful value. Research on AI transformation supports the importance of this broader view. Around 70% of potential AI value sits within core functions such as sales, marketing, manufacturing, supply chain, and pricing, areas where workflow redesign can have substantial implications. Another study found that workflow redesign had the strongest relationship with EBIT impact among 25 organizational attributes studied. People therefore become part of the implementation equation from the beginning. Kukarina's work at Deeply Human Innovation includes strategic advisory, intelligence and research, innovation programs, and ecosystem-building designed to connect technological ambition with human and organizational considerations. Her experience suggests employee reactions can provide valuable operational information. Resistance may reveal accumulated change fatigue, unclear responsibilities, insufficient preparation, or practical issues that executive planning has missed. This perspective also informs her "Well-Being Compass," built around four principles: proactive thinking, purpose-driven decisions, humanity-centric design, and adaptability. The framework encourages leaders to consider scenarios in which models, markets, regulations, workforce expectations, or business conditions change. It also invites them to challenge assumptions such as universal data readiness or the expectation that every new AI capability belongs somewhere in the organization. Her philosophy extends beyond AI implementation. Deeply Human Innovation's purpose is to embed humanity-centric thinking into the tools, teams, and systems shaping the digital world. That perspective places organizational decisions within a wider social context, where business choices can influence employees, customers, communities, and future generations. "The quality of our tools matters, and so does the quality of the world those tools help us create," Kukarina remarks. For leaders, that expands the AI conversation from deployment to responsibility. The strategic advantage may come from developing the discipline to decide where intelligence belongs, how people participate, and which outcomes deserve investment. Essentially, technology can accelerate execution, but strategic judgment remains a human responsibility. For Kukarina, that judgment begins before implementation, with the questions leaders choose to ask.

What we learned about how to achieve ROI with AI - context, craft and culture

Not surprising -- 85% of knowledge workers now use Artificial Intelligence (AI). Quite surprising -- only six percent of executives can point to clear, organization-wide return on investment (ROI). That gap is the whole problem, and as token usage climbs and licensing bills multiply, it's growing. When transformation stalls, many IT leaders instinctually blame the models or go hunting for a sharper algorithm. But as 'customer zero', rewiring how more than 13,000 Atlassians plan, decide, and deliver, we've hit on a deeper truth -- context, craft and culture, not models, are your competitive advantage. AI transformation is a leadership, culture, and architecture problem. If you want to close that six percent gap, you have to rethink what enterprise context actually means, how you capture it, and how you lead your people through the shift. We laid out the full playbook -- including our biggest missteps -- in our new report, Leading with Context: Lessons from Atlassian's AI Journey. De-faking 'context' - thinking beyond static data repositories You've probably noticed that 'context' has quickly become a vendor buzzword. But we're casting too wide a net. Enterprise tech providers are defining context as plugging a Large Language Model (LLM) into a static, historical library of documents. But that isn't really context -- real institutional memory isn't just what was built or what the policy states. Your team (and your AI agents) rely on the lived history of how decisions were actually made -- the trade-offs weighed in a meeting, the rationale behind a pivot, the links between engineering tasks and corporate Objectives and Key Results (OKRs), and the real-time velocity of cross-functional teams. This constantly evolving context layer is the beating heart of your business. At Atlassian, the context layer at the core of our platform is called the Teamwork Graph and it grounds our AI strategy. In our evaluation of AI models across 10 complex scenarios, we found that grounding them in this context yielded 44% more accurate responses and decreased token consumption by 48%, while maintaining the same speed. The mechanisms capturing context by default Expecting busy workers to manually document every decision just to supply an AI is unrealistic -- relying on individual discipline to capture context is a losing strategy. Instead, context must be embedded by default through a handful of structural mechanisms: * Open by default -- Atlassians work in shared, searchable digital spaces, rather than personal drives or locked private channels. This way, AI can reason across team boundaries -- and our human teammates can always cross-reference what their peers are doing. * Work linked to strategy -- every project is explicitly tied to centralized, rank-ordered goals. This lets agents understand trade-offs and prioritize based on business strategy rather than chat volume. * Automated institutional capture -- we use AI note-takers to transcribe back-and-forth conversations, map whiteboard sticky notes directly into shared tasks, and publish decision records in our shared Confluence workspace, where these documents live. When you build agent-ready data as a byproduct of everyday work, you can watch your context compound in real time. Where we learned tough lessons about our craft Authentic transformation requires admitting when something fails. Transparency prevented us from doubling down on bad bets. * We paused two of our 14 initial use cases. Of 14 AI use cases we scoped, we shipped 12 Minimum Viable Products (MVPs) and scaled 10. The two we paused taught us the most -- one because it required excessive engineering overhead relative to its value, another because our sales motion evolved, and we decided to reallocate the resources toward higher-value work. * We redefined 'superusers' -- twice. We started by measuring Weekly Active Users (WAUs), which only tracked logins. Then we set a company-wide bar of 40 AI interactions a week. That was a mistake, because it penalized non-technical roles (like legal) and rewarded sheer token volume. We ultimately moved to a function-specific metric, which we still utilize -- roughly the top 10% of users redesigning workflows within their own craft. * We hit an unexpected bottleneck. As AI increased our engineers' output, pull request (PR) cycle times improved by 30.8%. But reviewer fatigue spiked, and our 'ease of release' and 'build and test' scores dropped under the load. We had to introduce agentic gating and AI code-quality checks downstream, so faster generation didn't break our human evaluation layer. It quickly became clear that measuring tokens or lines of code is a vanity exercise. What matters for lasting impact is measuring the outcome, not the activity. Lead by demo, not mandate One lesson stands out from my 20 years in HR -- cultural transformation cannot be forced by executive decree. A stigma begins to form when organizations push for AI adoption without first establishing cultural alignment. Recent research from Atlassian's Teamwork Lab found that colleagues rated AI users as 10x lazier than non-users for the exact same quality of work. That penalty tracks directly with culture -- 37% of employees in neutral cultures perceived AI users as lazy, versus 11% to 13% in cultures that actively celebrate AI use. To kill the stigma for good, leadership has to model the right behavior: * Lead by demo. When Atlassian leaders live-demo how they use AI (think: screen-sharing their own decision-making or setbacks with AI), their teams are 4x more likely to make AI part of daily work. * Empower the superusers. A single sustained AI superuser lifts their entire team's performance by 56%. Find that top 10% of craft innovators in your org, and give them the space to uplevel their peers. * Bring tech and people leaders together. We expanded my role from Chief People Officer to Chief People & AI Enablement Officer, integrating IT, Data Science, HR, and Customer Engineering under one roof. To my team, it felt like the obvious next step, as the bottleneck to AI ROI isn't tool deployment, but human trust, skill building, and behavioral change. One Monday morning action - start with tasks, not titles If you want to move into the six percent bracket, your first move shouldn't be rewriting job descriptions or replacing titles. Odds are, that will create panic, resistance, and paralysis. Instead, start small. Map a single high-value workflow and break it down by tasks. Sit down with your functional leads and evaluate every task against two questions: * Where can AI execute the high-volume, repetitive work? * Where do we need humans to step in with high-stakes judgment? When we applied this process to customer support, we automated high-volume ticket routing, cut resolution times from days to minutes, and enabled 38% of issues to be resolved entirely by AI. More importantly, it freed up our support engineers to apply their judgment to complex enterprise challenges. Rapidly, we watched customer satisfaction improve. We still have much more to figure out, but analyzing our mistakes has been an insightful practice. As we move forward, we won't be chasing the perfect AI model, but rather focusing on institutional context, cultural alignment, and crafting real workflows. That redirection transforms AI activity into lasting enterprise impact.

Becoming AI-Native Requires the Right People More Than the Right Technology. Here's the Shift Leaders Need to Make.

Companies must understand the growing importance of human judgment, with human-in-the-loop processes ensuring AI outputs remain accurate, strategic and trustworthy. Becoming AI-native requires more than adopting new tools; it requires companies to transform their people, processes and workflows so AI can drive meaningful business impact while preserving human judgment and creativity. Not long ago, the corporate trend was digital transformation. Now the movement is about becoming AI-native. The change requires new technologies, but what is missing in most cases is the organizational transformation needed to make any of it work. The companies that become AI-native will not be the ones with the longest list of AI tools. They will be the ones that redesign their roles, operating models and workflows to support AI-enabled work. AI adoption is a people challenge first There is a persistent misconception that access to AI technology equals transformation. Executives assume that when their teams have AI tools, they will naturally weave them into everyday work. In reality, adoption fragments in predictable ways. Some employees experiment on their own with tools like Claude or ChatGPT. Others are unsure where AI fits into their role and worry about accuracy, brand risk or being seen as cutting corners. Without structure, adoption becomes scattered and inconsistent. Effective adoption requires change management. Marketing leaders have been through this before with social media, mobile marketing and marketing automation, each requiring new skills, new workflows and a deliberate organizational response. AI is different because of the speed of change. Models, capabilities and best practices move on a quarterly cycle. What worked six months ago is already outdated. That means AI can't be treated as a side project. To become AI-native, it has to sit at the center of how the function operates. "Figure it out yourself" does not scale In most organizations, AI enters as shadow technology. Employees experiment with public tools and automate small tasks. Experimentation is valuable, and I encourage it on my own team, but it is not enough to create an AI-native function on its own. As AI applications become more complex, self-guided adoption breaks down. New models are released, new agents create new possibilities and new risks, and what individual employees discover may not get shared, scaled or governed. The result is a function that looks more productive in patches but is no more capable as a whole. Becoming AI-native means cultivating a structured culture of experimentation, where teams can test new approaches within guardrails for responsible use. They also need shared awareness of data privacy, accuracy, brand voice and human review -- the failure modes that quietly erode trust if left unchecked. The need for embedded AI guidance As AI becomes part of everyday operations, new roles are needed to guide adoption. Hybrid positions like AI Forward Deployed Engineers are emerging to bridge the gap between the technology and how it actually gets used inside a function. Marketing teams understand the customer, the campaign mechanics and the brand. Technical teams understand the tools. Neither side, on its own, can reliably translate a real marketing problem into a well-designed AI workflow. That gap is where embedded AI specialists earn their place. In my own function, I have hired multiple AI automation specialists whose explicit remit is to work alongside marketers who are not yet AI-native, identify repetitive processes, build custom automations and AI agents, and coach the team on how to use AI in their day-to-day work. The hires that have made the biggest difference are not pure engineers and not pure marketers. They are operators who can build, ship and explain. What I have learned is that the role is less about technology than translation. The marketers know the work, but often do not know what to ask AI to do. The specialists know the tools, but they need the context to apply them well. Pairing the two has done more to move my function toward AI-native operations than any single tool rollout. Human judgment becomes more valuable As organizations become AI-native, AI shifts from supporting individual solutions to powering shared, repeatable systems, and the leverage compounds. In a marketing context, that might look like a shared AI infrastructure for content creation, SEO, reporting and campaign operations, built once and reused by every team rather than rebuilt by every individual. The paradox of AI is that as execution gets easier, the role of the human-in-the-loop (HITL) becomes more important, not less. It is not a limitation to design around but the mechanism that makes AI workflows trustworthy enough to scale. AI can accelerate content production, automate reporting and reduce operational bottlenecks. But it cannot own strategy, positioning, customer understanding, partnerships or accountability. Every meaningful AI workflow we have built in marketing has a HITL checkpoint somewhere in it, often more than one. As AI democratizes production, differentiation shifts to the quality of the idea, the clarity of the strategy and the strength of the customer relationship. The companies that get HITL right do not see it as a brake on AI productivity. We see it as the thing that lets us move faster, because we trust what we are shipping. AI-native is never finished Becoming AI-native is not a one-time project with a final milestone. The half-life of any "best practice" is getting shorter, and the functions that thrive will treat AI as a continuous operating discipline. The goal isn't to replace people. It's to eliminate repetitive work, increase leverage and give teams more time for creativity, strategy and customer impact. AI-native companies will not be defined by how many tools they adopt. They will be defined by how well they help their people adapt.