About the author
Harvey Fijucek
Chief Commercial Officer
TL;DR
AI transformation is an operating model change. It shows up in who owns the decision when an AI agent proposes something, which workflows get redesigned first, and whether the plan survives contact with an engineering team. The technology works. The organizations that get value are the ones that change around it. Here is how to move past the pilot phase, with real examples, and what it looked like when we ran the same change on ourselves.
The models were the easy part
A widely cited 2025 study1 found that 95% of generative AI pilots produced no measurable financial return. AI skeptics love that figure. The finding that matters lies beneath it: the failures came from poor workflow integration and misaligned incentives. The models were rarely the problem.
The 2026 research 2, 3 is blunter. Across dozens of enterprise deployments that reached scale and delivered real value, the difference between success and failure was almost never the model. It was the organization.
Here is what that looks like in practice. One fintech pointed an AI coding agent at millions of lines of legacy code and finished a migration in weeks. A major bank ran a comparable customer-support program and still measures it in years. Same models, same kind of use case, different organizations. Organizational context matters more than the technology.
AI works. The real question for a C-suite is whether the organization has changed enough to use it.
What is AI transformation?
AI transformation is a change to your operating model. Roles, decision rights, governance, and the daily mechanics of how work moves through review. Add AI to an existing workflow, and you get a marginal gain in efficiency. Rebuild the workflow around AI, and you get a different kind of organization. The real work is deciding which tasks still need human judgment and which don’t. That line gets redrawn across every role, every workflow, every handoff.
Three shifts are worth holding in mind:
- From isolated use cases to connected systems, where customer experience, operations, R&D, and planning reinforce each other.
- From episodic initiatives to continuous processes that sense signals, decide, and learn in real time.
- From task automation to human value creation, with people focused on judgment, orchestration, and accountability.
One principle sits underneath all three: humans move from in-the-loop to in-the-lead. As agents move from analysis into execution, someone has to own the decision, set the autonomy thresholds, and stay accountable for the outcome. We built that rule into how we work and call it AI proposes, human disposes. An agent produces a plan, an implementation, or a draft. A person decides whether to move forward and is accountable if it doesn’t hold up.
How to tell your effort has stalled
Before touching a roadmap, start with an honest look at where you actually are.
A few signs show up again and again in companies that adopted AI without laying the foundations:
- AI tools are live across teams, but nobody can say who owns the decisions those tools inform.
- Pilots keep launching. Few reach production. Nobody has asked why.
- The AI roadmap exists as a slide deck. The engineering team has never seen it.
- Different teams have quietly landed on different rules about what AI is allowed to touch.
- Leadership can point to new tools, but not to one workflow that genuinely changed.
One or two of these are normal growing pains. More than that, it means the tools arrived before the operating decisions did. That order rarely fixes itself.
What the companies that crossed over did differently
Success often comes down to where you place the human. Two operating patterns emerge:
- Approval models: A person signs off on almost everything the AI produces.
- Escalation models: The AI handles the bulk of the volume, and humans review only the exceptions.
The data is clear: escalation models delivered a 71% median productivity gain, compared to just 30% for approval models4.
We saw this play out with a recent client – a leading product data platform- where analysts parsed specs by hand, taking an hour per document. By deploying a system using retrieval-augmented generation (RAG) and vector search, we cut processing time to ten minutes per document and achieved up to 30x faster speeds on heavy files. This allowed analysts to shift from manual grind to high-value quality assurance. Critically, their feedback loop pushed the system’s accuracy from 70% to 85%. The AI carried the volume, while the humans maintained the judgment.
Fix the process before you point AI at it. AI amplifies whatever process it lands on.
A translation services firm 5 learned this the hard way. Its first attempt at AI recruiting failed because the team expected the model to fix a broken hiring process on its own. The second attempt started by fixing the process, then applying AI to it. Time to screen a role dropped from three hours to three minutes. Intake efficiency rose 83%. Same company, same goal, one working process in between.
This is the step most transformations skip. Read the business before touching the roadmap: which outcomes actually matter, what state your data is in, and which quiet dependencies a change could break.
We built a discipline around getting that order right and call it BDI™ – short for Business, Data, Interactions. It is the set of questions that must be answered honestly before a plan is written, so it reflects what is true rather than what looks good in a deck. The advantage is in orchestration, not the model. For a large share of deployments, the choice of foundation model was fully interchangeable. The durable edge came from the orchestration layer: how agents connect to systems, how work routes, how oversight is wired in. That is engineering, and it is where the real transformation happens.
Organization-wide AI transformation: what effective sponsorship looks like
Effective sponsorship shows up in what a leader does. The sponsors who drive results steer actively: weekly check-ins, blockers cleared as they appear. That’s the baseline. Organization-wide transformation asks for more.
The strongest sponsors tie AI adoption to the company’s OKRs, so success is measured against real business outcomes. They sit between the technical and business teams and get both sides to co-own the rollout. And they treat early pilots as experiments, giving teams room to fail, iterate, and learn.
One more pattern is worth naming. Every successful project was iterative. Start with one workflow, get it working, add the next. None of them planned the whole program up front and delivered it in one shot.
We ran this on ourselves first
We are an AI-native engineering partner for mid- to large enterprises, so the first AI transformation we ran was within Notch. AI now runs through every role and every stage of delivery. Engineers, analysts, designers, and project managers direct agentic workflows. We have 100 AI Champions across every discipline, backed by a VP of AI – an industry veteran of 15 years. Role boundaries became more permeable, and more people contribute to more of the work, so problems surface earlier and fewer are hidden in silos.
Capability was the bottleneck we cared about most. Upskilling usually depends on senior availability, and seniors are the people you least want pulled off client projects. So we built an internal certification system on Sovera AI™, our own agentic platform: scenario-based tasks, real-time evaluation, three skill levels, guided feedback. Eight specialized agents coordinate the experience. Engineering upskilling was accelerated by 50%. You can read more about it here.
Governance holds it all together. Every agent action maps to a control level, and irreversible actions always require a human gate. We built that into a framework we call NotchForge™. It is the governance layer that carries a transformation from the plan into the delivery work, so it doesn’t quietly disappear the first time a real deadline shows up or a team member shifts to a new project.
Where this leaves you
AI transformation has a starting point, a first milestone, and a steady operating state you settle into. Getting there is the project. Staying there is what compounds.
Six months is the honest checkpoint. By then, a real transformation has changed at least one workflow, not just handed a team a new tool. Ownership is sharper than it was in month one. And governance is something engineers do without thinking about it, not a document nobody has opened since it was written. Six months is long enough for a stalled program to still look busy. It is also long enough that it can no longer hide behind “we’re still early.”
Get the operating model right – decisions, ownership, and governance built for the way AI actually works – and the transformation is what your organization becomes. The tools follow from there.
Want to see what the operating model change looks like in practice? The frameworks behind this piece, NotchForge™ and Sovera AI™, sit alongside our strategy, agentic systems, custom software, and model work in one place. See how we engineer AI transformation, end-to-end. Visit AI/Impact™.
- 1MIT NANDA initiative, “State of AI in Business” (2025) ↩
- 2World Economic Forum, “Organizational Transformation in the Age of AI: How Organizations Maximize AI’s Potential” (March 2026). ↩
- 3, 4, 5 “The Enterprise AI Playbook: Lessons from 51 Successful Deployments,” Stanford Digital Economy Lab (April 2026). ↩ ↩ ↩
