AI Adoption Is Not the Problem
Why capability, not adoption, is the new competitive advantage
By Donovan Crewe · · Enterprise AI, AI Strategy, Developer Productivity, Engineering Leadership, AI-Native Engineering
Over the past two years, the industry has been obsessed with AI adoption.
How many developers are using AI? How many lines of code are AI generated? Which model performs best? Every quarterly report seems to include another statistic demonstrating that AI has become embedded within software engineering. Organisations have invested heavily in enterprise licences, rolled out coding assistants across engineering teams and trained developers on the latest AI capabilities. By almost every measurable indicator, AI adoption has been a success.
But I believe we're now measuring the wrong thing.
Adoption was always the first challenge. It was never the end goal.
The organisations that will lead the next decade won't be distinguished by whether their engineers use AI. Almost everyone will. They will be distinguished by how effectively their engineers collaborate with AI, how quickly they improve, and how systematically that knowledge spreads throughout the organisation.
Capability, not adoption, is becoming the new competitive advantage.
There is a trend emerging that reflects this shift. More organisations are hiring Forward Deployed Engineers (FDEs). Originally popularised by companies building AI products, these engineers work directly with customers to understand problems, rapidly prototype solutions and demonstrate what modern AI systems are capable of. They reduce the time between an idea and a working solution, accelerate customer adoption and provide an invaluable feedback loop between engineering and the real world.
It's a role that makes perfect sense.
When introducing a transformational technology, organisations need people who can bridge the gap between possibility and practical application. FDEs have proven remarkably effective at doing exactly that.
What I find more interesting, however, is the capability we haven't built.
We've become increasingly good at helping customers adopt AI.
We're still remarkably poor at helping our own engineering organisations become better at using it.
Those are not the same problem.
Giving every developer access to Claude, ChatGPT or GitHub Copilot is a technology rollout. Building an organisation that becomes measurably better at engineering with AI every month is an organisational capability. One can be completed in a quarter. The other becomes part of how the company operates.
If we look back at previous shifts in software engineering, this pattern should feel familiar.
Agile didn't become successful because organisations purchased Jira. Cloud wasn't transformative because companies opened AWS accounts. Kubernetes didn't magically create platform engineering. Every major shift created entirely new organisational capabilities alongside the technology itself.
Agile introduced Scrum Masters and Agile Coaches. DevOps gave rise to Platform Engineering, Site Reliability Engineering and Developer Experience teams. These functions weren't responsible for delivering customer features. Their responsibility was far broader. They continuously improved how every engineer delivered software.
AI is following exactly the same trajectory.
Today, most organisations have invested in AI tooling. Engineers have access to frontier models, coding assistants and increasingly sophisticated agentic workflows. What happens next, however, is largely left to chance.
One engineer discovers a workflow that dramatically improves architectural planning. Another builds a collection of specialised agents that eliminate hours of repetitive work. Someone else learns how to provide richer context, reducing prompt iterations by half. A team quietly develops an exceptional pattern for orchestrating multiple agents across an entire delivery lifecycle.
Most of those discoveries never travel beyond the people who made them.
The knowledge lives inside Slack conversations, personal prompt libraries or individual habits. Teams improve independently while the organisation itself barely changes. Every engineer is learning, but the organisation isn't.
I believe this is the capability most enterprises are currently missing.
We have invested heavily in AI adoption.
We have invested very little in AI capability.
That distinction becomes even more important as AI-native engineering matures. Success is no longer determined by whether someone knows how to ask a good question. Engineers now need to understand context engineering, persistent project state, agent orchestration, model selection, evaluation strategies, MCP ecosystems and an increasingly complex collection of AI-native practices. Left unmanaged, every engineer develops their own way of working, some highly effective, others significantly less so.
The problem isn't that engineers aren't learning.
It's that organisations have no systematic way of learning from their engineers.
This raises an obvious question.
How do you scale continuous improvement across hundreds or thousands of engineers?
A Principal Engineer can mentor a team. An Engineering Manager can coach a handful of people. Communities of Practice help spread knowledge, but they rely on engineers voluntarily sharing what they've discovered. None of these mechanisms can continuously observe how an entire engineering organisation works with AI, identify emerging best practices, understand where people struggle and provide personalised guidance to every engineer.
That simply doesn't scale.
I believe this is where a new discipline begins to emerge.
Not AI adoption.
Not AI governance.
AI Capability Engineering.
The purpose of AI Capability Engineering isn't to manage AI. It's to continuously improve how engineers work with it.
Imagine an organisation that can identify where developers experience friction when working with AI. Not through surveys or anecdotal feedback, but by understanding behavioural patterns across real engineering work. It recognises when engineers repeatedly ask questions already answered in project documentation. It detects when insufficient context leads to unnecessary iterations. It identifies ineffective model selection, inefficient workflows and patterns that consistently produce poor outcomes.
More importantly, it doesn't stop there.
It also identifies what excellence looks like.
Within every organisation there are engineers quietly achieving extraordinary results. They orchestrate specialised agents effectively. They structure context in ways that consistently produce better outcomes. They use project memory, reusable workflows and proven engineering patterns that allow AI to operate at its full potential.
Today, those behaviours remain hidden.
Tomorrow, they become organisational assets.
At Lumenalta we've been exploring this idea through AtlusInsights. Rather than measuring how often engineers use AI, it focuses on understanding how effectively they collaborate with it. By observing anonymised engineering patterns, comparing behaviours against an internally developed capability rubric and identifying both friction and excellence, the platform provides personalised recommendations that help engineers become more agentic over time.
What makes this approach particularly powerful is that there is no single path to becoming an AI-Native Engineer. One developer may consistently provide rich context but struggle with agent orchestration, another may repeatedly ask AI questions that have already been answered in project documentation. Someone else may rely on inefficient workflows or choose models that are poorly suited to the task at hand. Generic training programmes inevitably target the average engineer. Personalised coaching meets each engineer where they are, helping them develop the specific capabilities that will have the greatest impact on their effectiveness.
Delivering that level of coaching manually simply doesn't scale. No engineering leader can continuously observe hundreds of developers, understand each person's strengths, identify friction points and provide timely, actionable feedback. AI changes that equation. It enables personalised coaching at organisational scale, ensuring every engineer receives the guidance they need to improve, rather than the same guidance everyone else receives.
Equally important, it learns from the organisation itself.
When highly agentic engineers consistently demonstrate successful workflows, those patterns can be identified, generalised and shared. The goal isn't to copy individuals. It's to understand why certain approaches consistently produce better outcomes and make those practices available to everyone else.
Instead of every engineer learning independently, the organisation begins learning collectively.
That, I believe, is the missing capability.
We've spent the last two years deploying AI.
The next decade will be defined by organisations that continuously improve alongside it.
This also builds on a broader trend I've written about before. AI systems become more valuable when they accumulate persistent state instead of repeatedly reconstructing knowledge. Engineering organisations should operate the same way. Every successful workflow, every effective prompting strategy, every orchestration pattern and every lesson learned should become part of the organisation's collective intelligence. Improvement shouldn't depend on who happens to sit next to whom or which Slack channel someone follows.
Capability should compound.
This is ultimately why I believe the highest-performing AI organisations won't be built by hiring better engineers.
They'll be built by systematically helping every engineer become more agentic.
The first wave of enterprise AI was about giving engineers access to intelligence.
The second wave will be about helping engineers continuously develop it.
The organisations that recognise that distinction won't simply adopt AI faster than everyone else.
They'll learn faster than everyone else.
And in the age of AI-native engineering, the organisation that learns fastest will almost certainly become the organisation that delivers best.