There’s a conversation happening in nearly every enterprise IT leadership team right now, and many are interpreting it incorrectly.

Here’s the setup. Three things are arriving at once. Low-code has quietly stopped being a category on its own and become a default feature buried inside almost every serious platform. In this context, “low‑code is dead” is best understood as a terminology shift, not a disappearance of enterprise needs. Gartner forecasts that the low-code development technologies market will reach $58.2 billion by 2029, and 75% of new enterprise applications will be built on low-code by the end of 2026.Vibe coding, in which you describe software in plain language and watch it appear, has gone from a viral demo to something most Fortune 500 companies have adopted in some form.And agentic AI, where software reasons and acts on its own, is on one of the steepest adoption curves enterprise technology has seen since the rise of public cloud. Gartner expects 40% of enterprise applications to integrate task-specific AI agents by the end of 2026, up from less than 5% a year earlier.

The tempting read, especially in the boardroom, is that this is a clean generational swap. Vibe coding absorbs low-code. Agents absorb workflows. Platforms fade away. That read is partly right and dangerously incomplete.

Let’s break down why.

The build is fast. The rest of the lifecycle isn’t.

If you look at an enterprise application across its full lifecycle, there are seven phases: discovery, design, build, test, deploy, operate, and change. Vibe coding compresses the first three. A business analyst can describe an idea and have a working prototype in hours, which accelerates stakeholder feedback by a factor of five to ten compared with document-based review. That’s a real shift and is one that’s bringing significant value in the early phases.

But over five years, the last four phases account for roughly 70% of an application’s total cost. AI-assisted development hands back 30%–50% of effort in the build phase. In operate and change, it returns less than 15%. So, you’re compressing about 30% of the cost while leaving the expensive 70% more or less untouched. That’s a meaningful gain. It is not a transformation. A platform that addresses all seven phases is.

The pattern shows up phase by phase. Design and architecture is where vibe coding starts to show structural limits, because generated applications embed implicit dependencies and assume defaults that don’t match enterprise reality. Test coverage extends to happy paths but rarely to the negative paths and regulatory scenarios that matter in production. Deployment, with environment promotion and rollback, sits largely outside the vibe coding loop. And change is the largest open question in the industry today: re-prompting a system produces a new version, but tracking what changed, why, and whether it’s safe to ship is still an active research problem.

The risk story many leaders still miss

The productivity numbers around vibe coding travel fast. The security numbers travel slower, and these are what should worry leaders the most. Veracode tested 100 leading large language models across 80 coding tasks and found that AI-generated code contains security vulnerabilities 45% of the time, with no real improvement across newer or larger models. A December 2025 study of 15 vibe-coded production apps found 69 vulnerabilities, with every single application missing basic cross-site request forgery (CSRF) protection and security headers. A production scan of more than 1,400 vibe coded applications found that 65% had security issues and 58% contained at least one critical vulnerability. The Cloud Security Alliance found that AI-assisted commits expose secrets at more than twice the rate of human-only commits.

This isn’t a reason to ban the tools. It’s a reason to be precise about where they belong. Vibe-coded output is best treated the way you’d treat a high-fidelity prototype from a vendor: useful for alignment and validation, expected to be re-engineered before production, and never shipped unreviewed into a regulated workload.

And agents bring their own twist: nondeterminism. The same input can produce different outputs across runs. While fine when you’re brainstorming, it’s not fine when the decision touches real money or a regulated customer outcome. Gartner’s own forecast is blunt: more than 40% of agentic AI projects will be cancelled by the end of 2027, driven by escalating cost, unclear business value, and inadequate risk controls. The build velocity is real. The operational assurance is not.

What actually scales: hybrid architecture

Here’s the thing the replacement narrative misses. Every major IT shift of the past 30 years was an absorption, not a replacement. Mainframes didn’t vanish when client-server arrived. Client-server didn’t vanish when web architectures arrived. On-premise didn’t vanish when SaaS arrived. They became one layer in a bigger architecture.

One of the most under-appreciated platform advantages is sustainability of change: enterprise applications change continuously after go-live. Model-driven approaches can reduce maintenance cost by structuring applications into standardized layers that enable modularity, reuse, and maintainability across large deployments. This supports agility without compromising governance or performance.

Prompt-driven AI-powered development might work in design –time where the risk tolerance is higher. As we move into run time, the risk tolerance is lower and the need for predictability and constraints is higher. Agentic AI and vibe coding accelerates reasoning tasks and prototype creation. A deterministic backbone and governance through a platform makes sure it scales safely.

Agentic AI and vibe coding are following the same path, and the analysts have now formalized the winning pattern. It’s a deterministic backbone with agentic reasoning at selected nodes, all sitting under a single governed control plane. Forrester has even named the new market categories: the agent control plane, which inventories and governs heterogeneous agents across vendors, and adaptive process orchestration, the layer that routes work and enforces audit trails. Gartner’s projection is that by 2030, 70% of enterprises will pivot to a consolidated automation platform that orchestrates processes, AI agents, bots, APIs, and human actions, up from 5% today.

Three questions tell you which delivery model fits a given application. Does it execute regulated decisions or carry direct customer impact? Can its decision logic be fully written down in advance? Will it live longer than three years?If you’re answering yes to these questions, you want platform-based orchestration with agents at the leaves, not a vibe-coded app held together by prompts. Determinism wins where the decision space is bounded. Agentic reasoning adds value at the leaves, not the trunk.

The hybrid pattern isn’t a compromise either, and that’s worth saying plainly. It’s the architecture that has consistently won at enterprise scale across every wave of modernization, because it lets each layer do what it does best. Systems of record stay the transaction truth, governed and audit-anchored. Adaptive process orchestration coordinates the work, routing between systems, applying decisioning, owning the audit trail, and routing between humans, bots, and agents.Agents sit at reasoning nodes inside processes, invoked with bounded scope and structured inputs and outputs, under the platform’s identity and governance perimeter. And the control plane provides the vendor-agnostic governance on top. Vibe coding still has a home in that picture. It lives in the design studio, where its speed is an asset and its security weaknesses are contained. The mistake is letting it escape the studio.

From token cost to cost per outcome

Most architecture debates end the moment someone asks what it costs to run. AI can accelerate design and build drastically. At the same time, token costs can rise quickly and unexpectedly.Modern agentic Al workflows solve a task with a team of agents using their own context windows and parallel execution, which can drive up token consumption rapidly. Token cost is becoming increasingly complex to understand fully, and as adoption increases, the topic is becoming more important.

As a contrast, model-driven generation on a platform foundation creates governed rules instead of open-ended code. Integrations are controlled, avoiding open agent loops. Workflows and development use models only where they add real value. Ultimately, fewer unnecessary model calls mean lower and more predictable cost.

Where Capgemini stands

Our position with clients has a name: Customer First in the agentic era. It’s not a slogan, it’s an operating-model claim. The leadership question is no longer “can we build an agent.” Anyone can. It’s “can we delegate decisions safely, at scale, and prove it afterwards.”

That means treating identity, policy, auditability, and human oversight as platform capabilities you build once, not project artifacts you reinvent every time. It means scaling the breadth of use cases before you deepen autonomy. Most organizations are far better served by ten agents at recommend level than one agent running fully autonomously, and Gartner’s project-cancellation figure comes largely from organizations that scaled autonomy before scope. And it means measuring decision quality, interventions, reversals, error severity, and customer impact, rather than the vanity metric of percentage automated.

We see this play out the same way across banking, insurance, telco, and the public sector. The platform owns the lifecycle. Agents own the reasoning nodes. Vibe coding owns the design studio. Our job is to make that operationally real inside a client’s regulatory perimeter. The results back this up: a financial services organization that put AI-driven anomaly detection inside governed incident workflows saw mean time to resolution drop more than 70%. A European insurer that redesigned journeys on a governed platform cut time-to-market from 24 months to six. In every case, the AI sat inside a governed process. That’s the recurring precondition for return.

The bottom line

If the question is “how fast can we build,” vibe coding wins. If the question is “how fast can we change safely, at scale, over five years,” the platform wins.

Most enterprise CIOs are making a five-year decision while being marketed a one-quarter narrative. The productivity is real, and so is the operational assurance gap. But the organizations that pull ahead won’t be the ones that picked a side. They’ll be the ones that layered all three waves inside a governed architecture, so that speed at the build plane never turns into risk at the run plane.

That’s the decision worth getting right.