AI-Native Operating Model: Rearchitecting vs. Layering AI on Top
AI-native operating model: rearchitect or layer AI on top?
A European mid-cap should rearchitect its operating model around AI for the two or three workflows that define its P&L, and layer AI on top only for peripheral, low-stakes tasks. Layering a copilot onto an unchanged process delivers convenience; rearchitecting the process around what AI can now do is what moves cost, cycle time, and error rates. The decision is not “AI yes or no.” It is “do we change how the work flows, or just who holds the mouse.”
That distinction explains most of the gap between AI ambition and AI results. When MIT’s Project NANDA studied enterprise deployments, it found the divide between winners and everyone else was driven by approach, not by model quality, regulation, or talent. The organizations extracting value were not the ones with the best models bolted onto old workflows. They were the ones that redesigned the workflow so the system could learn, adapt, and act.
Adoption is not what large European companies lack. Eurostat counts 55.03% of large EU enterprises (250 or more employees) using at least one AI technology in 2025. Across all enterprises with 10 or more employees, the Nordic shares lead the EU: 42.0% in Denmark, 37.8% in Finland and 35.0% in Sweden. The most common use was analysing written language, at 11.8%. Those figures count tools in use. They say nothing about whether a single workflow was redesigned around them, which is the question that decides whether the P&L moves.
For a mid-cap, roughly EUR 250M to 2B in revenue and below the Fortune 500, the choice carries more weight than it does for a global group. There is no internal AI lab to absorb a failed redesign, and the same small leadership team owns both the bet and the explanation to the CEO. That is why the decision has to be made workflow by workflow, with a number attached.
Key takeaways
- Layering AI on top means adding assistants and copilots to existing processes. It is fast, low-risk, and produces marginal gains. It rarely changes the P&L.
- Rearchitecting means redesigning the workflow, decision rights, data ownership, controls, and roles so AI does the work, with humans governing exceptions. It is harder and produces structural gains.
- The MIT NANDA report found the AI value gap is driven by approach, not technology: most pilots showed adoption without transformation.
- Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027 (CDO Magazine on the forecast) due to unclear value, cost, and inadequate risk controls: a rearchitecting-discipline problem, not a model problem.
- Choose per workflow, not per company: rearchitect where the work is core, repetitive, and measurable; layer where it is peripheral.
- The target operating model for a rearchitected workflow specifies five things with named owners: the workflow, decision rights, data ownership, controls, and roles.
- EU AI Act high-risk obligations now apply from 2 December 2027 (Annex III) and 2 August 2028 (Annex I) (European Commission). A workflow redesigned today will be running when they land.
What “layering AI on top” actually means
Layering is the default path because it is the path of least resistance. You take a process that already exists, a claims queue, a support desk, a monthly close, and you add an AI assistant inside it. The human still owns the workflow. The AI drafts, suggests, summarizes, or retrieves. The org chart, the handoffs, the systems of record, and the decision rights all stay the same.
This is real value and worth doing in many places. A copilot that saves an analyst twenty minutes per report is a genuine gain. But layering has a ceiling: because the process is unchanged, the improvement is bounded by how much time a human saves on their existing steps. You are optimizing the old operating model, not replacing it.
The risk is mistaking layering for transformation. Gartner has warned about “agent washing”: rebranding assistants and rule-based automation as agentic AI without substantive change. When Gartner analyzed why agentic projects get canceled (CDO Magazine), the causes were escalating costs, unclear business value, and inadequate risk controls: symptoms of bolting capability onto processes never designed to use it.
The AI-native target operating model: what it has to specify
A rearchitected workflow redesigns the work around what AI can now do. The unit of design is the workflow, not the tool. Instead of asking “where can we add an assistant,” you ask “if AI handled the core of this work, how would the process, the data, the controls, and the human role be different?” (For the company-level definition and the test that separates AI-native from AI-enabled, see what an AI-native enterprise is.)
The target operating model is the written answer to that question, one per workflow. Read as an AI operating model framework, it has five parts, and each needs an owner in the business, not only in IT.
| Element | What layering leaves in place | What the target operating model specifies | Owner |
|---|---|---|---|
| Workflow | Same steps, human-gated; AI drafts inside them | Which steps AI executes, which are removed, where exceptions route | Process owner in the business line |
| Decision rights | Implicit: a human decides everything | Per decision: AI decides, AI recommends, or a human approves, with the thresholds that switch between them | Business owner, signed off by risk |
| Data ownership | Data stays where it was; the assistant reads what it can reach | Which sources the workflow reads and writes, who owns each, where it is processed and stored, and what the system retains as it learns | A named owner per data source |
| Controls and audit | Whatever the tool vendor logs | Policy checks inside execution, an evidence record per action, monitoring thresholds that trigger review | Risk and compliance, with engineering |
| Roles | Unchanged job descriptions | Who reviews exceptions, who holds oversight authority, who maintains the system after handover | Line manager and HR, with the works council where one exists |
The workflow row is where steps get reordered or removed. Work that was sequential and gated by a person becomes parallel and executed by AI, and people handle the exceptions rather than every case. The decision-rights row is where control and speed are traded on purpose rather than by accident. How to set the approval gates and escalation paths is covered in human oversight for AI agents.
Data ownership is the row most designs leave thin. The system needs clean, accessible context to act, so data plumbing, retrieval, and state management become first-class work. NANDA identified the inability to retain feedback and improve over time as a core barrier; a rearchitected model designs for learning. That makes ownership the real question: if what the workflow learns sits inside a vendor’s model, the company rents its own process knowledge back. Keeping it in your own stack, model-agnostic and processed in the EU where your obligations require it, is what owning your intelligence means in practice.
Controls belong inside execution, not appended after it. What each action’s evidence record has to contain is set out in audit trails for AI agents.
The roles row has to be decided, not discovered. In Germany, the works council has a right of co-determination over “the introduction and use of technical devices designed to monitor the behaviour or performance of the employees” (Works Constitution Act, section 87(1) no. 6). A mid-cap with German operations that redesigns a workflow around a system logging every action should plan that conversation into the timeline, not meet it at go-live.
The tell that the model is real is measurability. A rearchitected workflow changes a number a CFO already tracks, cost per claim, days to close, resolution time, error rate, not a soft “productivity” estimate.
Where the EU AI Act lands in the model
If a rearchitected workflow falls into one of the high-risk use cases in Annex III of the EU AI Act, the controls and roles rows stop being design preferences. The Commission lists employment, education and critical infrastructure among those areas. Article 14 requires high-risk systems to be designed so natural persons can effectively oversee them while in use, and Article 26 requires deployers to assign that oversight to people with the necessary competence, training and authority. A mid-cap that runs the system is the deployer, so this lands on its org chart, not only on the vendor’s.
The dates moved this year. Regulation (EU) 2026/1744, the Digital Omnibus on AI, set the application date for Annex III high-risk systems to 2 December 2027 and for high-risk AI in products covered by Annex I to 2 August 2028 (European Commission). A workflow redesigned in 2026 will be in production when those dates arrive. Designing the oversight roles now costs a line in the target operating model; retrofitting them later costs a redesign.
How to decide: rearchitect vs. layer, per workflow
The mistake is treating this as a company-wide philosophy. It is a per-workflow decision. Use four questions.
1. Is this workflow core to the P&L? If the process directly drives cost, revenue, or risk at scale, it earns a rearchitecting conversation. If it is peripheral, layer an assistant and move on.
2. Is the work repetitive and high-volume? Rearchitecting pays off when the same shaped work happens thousands of times. One-off, bespoke work rarely justifies the redesign cost.
3. Can you measure the outcome today? If you cannot state the current cost, cycle time, or error rate, you cannot prove the redesign worked, and you will become part of Gartner’s cancellation statistic. Instrument first.
4. Can you govern the autonomy? If the workflow touches regulated decisions, you must be able to enforce policy and audit every action. If you cannot yet, constrain the scope until you can.
| Layer AI on top | Rearchitect the operating model | |
|---|---|---|
| Best for | Peripheral, low-stakes, varied work | Core, repetitive, measurable work |
| Human role | Owns the workflow, AI assists | Governs exceptions, AI executes |
| Effort & risk | Low | High |
| Upside | Marginal, bounded by time saved | Structural, changes the P&L |
| Failure mode | Mistaking convenience for transformation | Under-scoping governance and measurement |
A pragmatic sequencing: layer broadly to build fluency and surface where the real friction lives, then rearchitect the two or three workflows where the numbers justify it. Trying to rearchitect everything at once is how budgets and credibility evaporate.
Why rearchitecting keeps failing, and how to de-risk it
Rearchitecting fails for organizational reasons more often than technical ones. The usual pattern is the pilot-to-production gap the whole industry now names: a promising proof of concept that cannot survive contact with real data, real edge cases, real compliance requirements, and real users. We wrote more on this in why enterprise AI pilots fail.
Three practices separate the projects that ship from the ones that get canceled:
- Instrument before you build. Establish the baseline metric the redesign is supposed to move. No baseline, no proof, no budget renewal.
- Design governance in from day one. Human oversight and auditability are cheaper to build than to retrofit, and regulated buyers will not deploy without them.
- Embed engineers where the work happens. The blockers to rearchitecting, messy codebases, undocumented tribal knowledge, compliance edges, live inside the company, not in the model. This is why forward-deployed engineering has become the AI go-to-market of record. Per First Round Review, FDE job listings rose sharply through 2025 as AI companies raced to deploy inside client workflows. We break down the role in what a forward-deployed engineer is.
The through-line: a rearchitected operating model is not something you buy, it is something you build inside the constraints of a specific business. That is deliberate, measurable, and governed work: the opposite of hype-driven experimentation.
Frequently asked questions
What is an AI-native operating model? An AI-native operating model is a way of running work in which AI executes the core of a process and humans govern exceptions, rather than AI merely assisting humans who still run the process manually. It redesigns the workflow, decision rights, data ownership, controls, and roles around what AI can now do, so improvements are structural rather than marginal.
What does an AI-native target operating model have to specify? Five things, per workflow: which steps AI executes and where exceptions route; the decision rights (what AI decides, what it recommends, what a human approves); who owns each data source the workflow reads and writes, and where it is processed; the controls and evidence record built into execution; and the roles that review exceptions, hold oversight authority and maintain the system after handover. Each element needs a named owner in the business, not only in IT.
Is layering AI on top ever the right choice? Yes. For peripheral, low-stakes, or highly varied work, adding an assistant or copilot is fast, low-risk, and delivers real convenience. The error is not layering, it is mistaking a layered copilot for transformation on a workflow that actually drives your P&L.
Why do so many AI projects fail to deliver ROI? MIT’s NANDA report attributes the gap to approach rather than technology: most deployments show adoption without transformation, often because systems cannot retain feedback or adapt over time. Gartner separately expects over 40% of agentic AI projects to be canceled by the end of 2027 (CDO Magazine), citing unclear value, cost, and weak risk controls. The full evidence, and what the few that succeed do differently, is in why enterprise AI pilots fail.
How do I choose which workflows to rearchitect? Rearchitect workflows that are core to the P&L, repetitive and high-volume, measurable today, and governable. Layer AI on everything else. Sequence it: layer broadly to build fluency, then rearchitect the two or three workflows where the numbers clearly justify the redesign.
How does governance fit into a rearchitected operating model? Governance is a design input, not an add-on. Policy enforcement, human oversight, and a full audit trail are built into how the workflow executes. For regulated sectors, this is the precondition for deployment: an ungovernable autonomous workflow will not ship regardless of its accuracy.
When do the EU AI Act high-risk rules apply to a rearchitected workflow? Regulation (EU) 2026/1744, the Digital Omnibus on AI, set the application date for high-risk systems under Annex III (use cases such as employment, education and critical infrastructure) to 2 December 2027, and for high-risk AI built into products covered by Annex I to 2 August 2028 (European Commission). From those dates, Article 14 requires the system to be built for effective human oversight and Article 26 requires the deployer to assign that oversight to people with the competence, training and authority to exercise it.
Why are forward-deployed engineers relevant to rearchitecting? Because the hardest parts of rearchitecting, legacy systems, undocumented process logic, and compliance constraints, live inside the company. Forward-deployed engineers embed in the business to build around those blockers, which is why leading AI firms have adopted the model. See our explainer on forward-deployed engineers and the 2026 benchmarks on enterprise AI agent ROI.
Where to start
Start with one workflow, one baseline metric, and a target operating model you could defend to an auditor and to your works council. Nucleo works with European mid-caps on exactly that step: engineers inside your stack, the redesigned workflow in production, the data governed to EU rules, and the system yours when we leave. See how an engagement works, or write to us with the workflow you would rearchitect first.
Sources: MIT NANDA, The GenAI Divide: State of AI in Business 2025; Gartner, Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (June 25, 2025); CDO Magazine, Over 40% of Agentic AI Projects Likely to Be Abandoned by 2027; Gartner, 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026 (Aug 26, 2025); Eurostat, Use of artificial intelligence in enterprises (data extracted December 2025); Eurostat, 20% of EU enterprises use AI technologies (11 December 2025); Regulation (EU) 2026/1744, Digital Omnibus on AI; European Commission, AI Act regulatory framework; EU AI Act, Article 14 and Article 26; German Works Constitution Act (BetrVG), English translation; First Round Review, So You Want to Hire a Forward Deployed Engineer.