Enable javascript in your browser for better experience. Need to know to enable it? Go here.
Blogs Banner

Why agentic AI pilots don’t scale

By Daniel Schmalz

 

Can your current technology and data foundation carry the weight of your AI ambitions?

In today’s enterprises, agentic AI rarely fails because of the models. It fails because of the foundation. This blog post explains why agentic AI pilots don’t scale and gives you a sequence of actions to fix it.
 

Anyone who wants scalable, trustworthy AI must release two underestimated brakes: deeply interwoven legacy monoliths and missing data readiness.
 

Thoughtworks delivers the end-to-end, risk-minimized, AI-accelerated path from a single source: from the monolith, via a modern, evolutionary target architecture, to automatically generated, agent-consumable data products and thus to true agentic AI readiness. Not as a cost block, but as business reimagination

The starting point: Everyone wants agentic AI, few actually get there

Nearly every board of directors today states the ambition to become “AI-native”: autonomous agents are meant to take over end-to-end processes, and AI is meant to become an integral part of the company’s products.
 

The ambition is right, but the reality in most large enterprises looks very different.
 

What we currently observe across the board: dozens of proofs of concept (PoC) emerge in parallel across the most diverse departments, ambitious, often technically impressive, but isolated.
 

The vast majority of these PoCs never reach production maturity. No measurable efficiencies with P&L visibility materialize, no AI scales into operations, no structural competitive advantage emerges. The organization learns a lot, but it doesn't change.
 

The research points the same way. McKinsey’s 2026 state of AI survey found that 37% of respondents attribute at least some earnings impact to AI, about the same as last year, and only about six percent qualify as high performers.
 

For agentic AI specifically, Gartner predicts that by 2027, 40% of enterprises will demote or decommission autonomous AI agents because of governance gaps identified only after production incidents.
 

The reasons are multi-layered. At the mindset level, executives cling to established structures and resource fiefdoms instead of embracing new methods and technologies as an opportunity.
 

At the organizational level, AI responsibilities are fragmented: data and AI teams sit in business domains, in the chief data officer organization, in IT and in separate centers of excellence, without a coherent operating model.
 

The technology experts bring methodological competence but little domain context; the business domains bring deep functional expertise but too little delivery capability.
 

Vibe coding produces fast, local PoCs there, but as soon as professional integration into the enterprise ecosystem is required, the silos hit their limits. The result is local optimization instead of enterprise-wide scaling.
 

The most underestimated reason, however, lies deeper, in the technology foundation. 

Why agentic AI pilots fail: Two invisible brakes in the technology foundation


Legacy dependencies: In most enterprises, core business capabilities still depend on organically grown monoliths: mainframe, embedded code, ABAP/SAP landscapes, historical applications. Small changes trigger large change tickets and massive resource commitments, because the business logic has been deeply interwoven over decades.
 

Architectural change is therefore only possible at a snail's pace, a pace fundamentally incompatible with the demands of agentic AI readiness. Agents need accessible, decoupled, API-enabled business capabilities. A monolith delivers the exact opposite.
 

Data readiness: Trustworthy AI requires trustworthy data. Enterprises today want speed: use cases are expected to go live in weeks, not years. Yet high-quality, consumable, well-governed data is the new and widely underestimated bottleneck along the entire value chain.
 

The problem is compounded by legacy data ecosystems such as DataStage or Teradata: batch-heavy instead of real-time, operationally burdensome, poorly suited for the cloud, locked into proprietary tools and vendors, with costs that grow with data volumes without the architecture becoming any more flexible.
 

The historically grown integration and processing logic of these “Extract, Transform, Load” (ETL) stacks represents a massive hurdle for any serious AI adoption.

The necessary shift in perspective to turn modernization into a growth lever


Many modernization initiatives are never started, or are repeatedly postponed, because from a business perspective they initially appear to be "nothing but cost." That perspective falls short.


The classic modernization drivers (cost reduction, talent risk, operational agility, reduced vendor dependency) remain valid. But a new class of motivators is now moving to the forefront, which we summarize under the term business reimagination:

  • Agent-native architectures

  • AI-native workflows

  • The ability to proactively steer rapidly shifting customer expectations and market dynamics, real-time operations

  • Not least, security: as AI systems become ever better at discovering vulnerabilities, the shift to modern, hardenable environments becomes a security imperative in its own right


Modernization is therefore no longer a cost line item, but risk mitigation, a growth lever and the ticket of entry into the agentic AI era.

The solution to ensuring agentic AI pilots scale successfully: One value chain, three stages


Over recent years, Thoughtworks has not only refined how monoliths are modernized into modern, AI-ready environments. It has actively shaped the market in doing so: from mainframe to embedded code and applications, from ABAP/SAP to legacy data ecosystems.
 

Our end-to-end offering follows a clear logic:

Monolith → modern target architecture → modern data architecture & data readiness → scalable agentic AI

Stage one: AI-first legacy modernization for lower risk and faster delivery
 

1. Understand your legacy estate with AI-first reverse engineering


One of the most challenging parts of any modernization is understanding the existing estate: code that no one in the organization fully comprehends anymore and that is barely documented. Thoughtworks analyzes the legacy code automatically and transforms it into a knowledge graph and a detailed specification:

  • architecture designs

  • dependencies

  • APIs

  • mapping of code to functions and business capabilities


This significantly reduces effort, lead time and thus the overall investment, while, as a by-product, creating the documentation that has been missing for years.
 

2. Design a target architecture built to last


Modern, flexible, secure and built to evolve. It helps that Thoughtworks did not merely co-invent the microservices concept (Fowler & Lewis, Sam Newman's Building Microservices), but has refined it with clients over many years: decomposition along business capabilities using domain-driven design, framed by the principles of evolutionary architecture.


3. Replace the monolith gradually with strangler fig


Critical systems are not replaced in a risky big bang, but gradually "grown over": new capabilities emerge alongside the legacy system, progressively take over functionality, and displace the monolith in a controlled way, with operations continuing throughout.

This pattern, coined by Thoughtworks’ Chief Scientist Martin Fowler, together with the patterns of legacy displacement that grew out of it, forms the methodological backbone of every Thoughtworks modernization.


Stage two: Data readiness from bottleneck to accelerator


1. Build data products in hours instead of weeks


Roughly six years ago, Thoughtworks invented data mesh. From many enterprise implementations, one element has crystallized into an industry standard: data products, products that directly integrate value, data quality and consumability for other participants across the enterprise.

The problem: manually building even simple data products takes days, and complex data products take about 12 weeks or more.

With the data product workbench, Thoughtworks has codified and heavily automated its data product delivery capability: simple data products are created in hours, complex ones in a maximum of two weeks.

The approach remains strictly use-case-driven and works retrospectively: from the use case through aggregated data products all the way to source-aligned data products.
 

  • An AI-native data product designer drafts the products based on open data product standards; human-in-the-loop remains guaranteed at all times through attributes, data quality parameters and fully editable specifications.

  • The data products are then developed automatically (spec-driven development) and deployed into the target environment, including automatically generated continuous integration and continuous delivery (CI/CD) pipelines, repositories and the medallion architecture.

  • In operation, the curated data products can be consumed directly by AI agents via model context protocol (MCP) connectors, while being automatically registered, including their metadata, in the client's data catalog, where they can be maintained effectively.


2. Migrate legacy data ecosystems automatically
 

For DataStage, Teradata and comparable landscapes, Thoughtworks has developed the AI-powered legacy data migration accelerator. It performs a deep analysis of the legacy data landscape and automates its migration into modern target platforms, as a standardized, scalable, and repeatable delivery model. The results:
 

  • up to 60 times faster migration

  • translation and transformation of data pipelines within hours

  • up to 90% shorter end-to-end lead times


Manual rework is largely eliminated, scaling across large legacy landscapes becomes efficient, and the foundation is laid for near-real-time data streams feeding AI applications and models.
 

Proven in practice: Unieuro, Italy's leading consumer electronics retailer, faced hundreds of undocumented data pipelines and a migration to Google Cloud Platform that would have taken years by hand.
 

With the AI accelerator, code translation and documentation generation were automated. Migration became up to 60 times faster than the manual approach. A multi-year project shrank to just a few months, and Unieuro now runs on a scalable platform serving as the foundation of its European expansion.

Stage three: Scalable agentic AI on a foundation that carries the load


Only on this basis does agentic AI move from PoC theater to operational reality:
 

  • Decoupled, API-enabled business capabilities that agents can orchestrate

  • Curated, quality-assured data products that agents consume directly via MCP

  • Near-real-time data streams for real-time operations

  • An evolutionary architecture that grows with every new AI generation instead of standing in its way.

The value at a glance
 

  • From PoC to P&L: AI initiatives reach production maturity and generate measurable, scalable efficiencies, instead of fading away in silos.

  • Risk under control: Strangler fig and the patterns of legacy displacement safeguard critical systems throughout the entire transformation. No big bang, with operational risk kept under control.

  • Lower investment: AI-first reverse engineering, the Data Product Workbench and the data migration accelerator shorten lead times (data products in hours instead of days; complex products in 2 instead of about 12 weeks; data migration up to 60 times faster, end-to-end lead time reduced by up to 90%).

  • Knowledge reclaimed: Automated knowledge graphs, specifications and documentation end the dependency on dwindling expert knowledge.

  • Future-proof and secure: An agent-native, evolutionary target architecture with a modern security posture, ready for real-time operations and every AI generation to come.

  • Thought leadership: Strangler fig, microservices, evolutionary architecture, data mesh. The concepts this offering is built on were coined by Thoughtworks itself and translated into enterprise reality over many years.


So the question is no longer whether your enterprise will deploy agentic AI. It’s whether your foundation is ready for it, and if not, how much longer you can afford to wait for the answer.

Disclaimer: The statements and opinions expressed in this article are those of the author(s) and do not necessarily reflect the positions of Thoughtworks.

Keep up to date with our latest insights