The Financial Times recently published an insightful piece titled "AI engineers embedded in business units risk creating new tech silos" (paywall), highlighting a growing trend in large enterprises: embedding AI specialists—often called "forward-deployed engineers" (FDEs)—directly into business units like finance, marketing, or operations. These engineers, frequently supplied by AI vendors such as OpenAI, Anthropic, or cloud hyperscalers, are tasked with rapidly prototyping and deploying AI solutions tailored to departmental needs. The goal? Accelerate innovation and outpace sluggish central IT.
Sound familiar? It should. This is the modern reincarnation of a pattern we battled two decades ago with ETL (Extract, Transform, Load) pipelines and departmental Business Intelligence (BI) teams. Let's break it down, draw the parallels, and explore why this "move fast" mentality, while seductive, risks entrenching data silos, application fragmentation, and—critically—vendor lock-in.
The Ghost of ETL Past: Departmental BI in the 2000s
Rewind to the early 2000s. Enterprises were drowning in data from ERP systems, CRM tools, and legacy databases. Central IT, burdened by governance, security audits, and enterprise-wide roadmaps, often took months (or years) to deliver BI dashboards or analytics reports. Frustrated business units—finance needing real-time P&L forecasts, marketing craving customer segmentation—took matters into their own hands.
They hired or reassigned ETL specialists and BI analysts who reported dotted-line to the department but solid-line to central IT for "oversight." Tools like Informatica for ETL, Cognos or Business Objects for BI, and homegrown SQL scripts proliferated. These teams built shadow data marts: finance pulled transactional data into a dedicated SQL Server instance; marketing ingested web logs into a separate Oracle schema.
The upsides were real:
- Speed: A marketing campaign analytics dashboard could go from idea to production in weeks, not quarters.
- Relevance: Solutions were hyper-tailored to departmental KPIs, bypassing one-size-fits-all enterprise warehouses.
But the downsides were insidious and long-lasting:
- Data Duplication and Inconsistency: The central data warehouse (often a Teradata or Oracle EDW) held the "source of truth," but departments replicated subsets, leading to version skew. Finance's revenue figures might diverge from the enterprise view by 2-5% due to differing transformation logic.
- Siloed Applications: Custom ETL jobs and BI cubes were maintained in isolation, creating technical debt. When a key analyst left, the solution became a black box.
- Support Nightmares: Central IT ended up firefighting these rogue systems during audits, migrations, or scaling events—without full ownership.
- Governance Gaps: Shadow data introduced compliance risks (e.g., GDPR precursors like SOX) and hindered cross-functional insights.
In T10 Consulting engagements from that era, we'd often discover 10-15 redundant data flows for the same customer master data. Consolidation projects cost millions and took 18-24 months, but they unified analytics and reduced TCO by 30-40%.
Forward-Deployed Engineers: ETL 2.0 with AI Superpowers
Fast-forward to today. Generative AI promises to democratize intelligence, but central IT—now juggling cloud migrations, zero-trust security, and MLOps platforms—remains a bottleneck. Enter FDEs: vendor-supplied experts embedded in the business to build RAG pipelines, agentic workflows, or fine-tuned models on departmental data.
The FT article cites examples from banks and retailers where these engineers spin up AI agents in days using vendor APIs (e.g., OpenAI's GPT endpoints or Anthropic's Claude). Finance gets an AI-powered fraud detector; marketing deploys a personalization engine. Just like 2000s BI teams, they report to the business for priorities but rely on central IT for... well, not much, initially.
Parallels to the ETL Era Are Striking:
| Aspect | 2000s Departmental ETL/BI | 2020s Forward-Deployed AI Engineers |
|---|---|---|
| Driver | Central IT too slow; need for agile insights | Central IT/MLOps too slow; need for AI velocity |
| Structure | Dotted-line to dept, solid-line to central IT | Embedded in dept, often vendor-contracted |
| Tools/Tech | Informatica, SQL, departmental data marts | Vendor APIs, LangChain, departmental vector DBs |
| Speed Win | Weeks for dashboards vs. months enterprise-wide | Days for AI prototypes vs. quarters for central |
| Core Risk: Silos | Duplicated data, inconsistent transformations | Fragmented knowledge graphs, inconsistent prompts/models |
| Maintenance Trap | Analyst turnover → orphaned scripts | Engineer rotation/vendor contract end → unsupported agents |
| New Twist: Lock-In | Proprietary BI licenses, but switchable | Deep vendor API entanglement; IP in prompts/models |
The Hidden Costs: Silos, Fragility, and Vendor Lock-In
FDEs amplify the old problems with AI-specific venom:
-
Data and Application Silos 2.0:
- Departments ingest data into vendor-specific stores (e.g., Pinecone for vectors in marketing, Weaviate in finance). Cross-unit queries? Impossible without heroic integration.
- RAG implementations duplicate embeddings of the same documents, leading to hallucination drift. One T10 client found three separate RAG systems querying the same policy docs with conflicting outputs.
-
Fragility and Technical Debt:
- FDEs build quickly but often skip enterprise standards: no centralized observability, weak data lineage, or scalability testing.
- When the engineer rotates out (contracts are typically 6-12 months), the business is left with a "prompt warehouse" that's undocumented and unmaintainable.
-
Vendor Lock-In: The Killer Differentiator:
- Unlike 2000s tools, modern AI relies on proprietary models and APIs. Your finance AI agent? Built on GPT-4o with custom fine-tuning—migrate to Claude? Rewrite prompts, retrain embeddings, renegotiate data egress.
- Vendors incentivize this: free credits for FDEs, but you're hooked on their ecosystem. Switching costs can exceed $1M in rework, per our N8N and RAG migration projects.
- IP risks: Departmental data fed into vendor models may train their foundational systems, leaking competitive edges.
The FT piece warns of "new tech silos," but it's worse—locked silos. In the ETL days, you could at least export CSV and rebuild. With AI, you're often captive to the vendor's evolution roadmap.
A Better Path: Centralized AI Platforms with Federated Execution
At T10 Consulting, we've architected AI transformations for Fortune 500 clients that avoid these pitfalls:
- Enterprise AI Backbone: Central MLOps platform (e.g., MLflow + Kubernetes) with shared vector databases (PGVector or Elasticsearch) and prompt registries. Departments access via APIs, ensuring lineage and governance.
- Federated Deployment: Business units define requirements; central AI engineers build reusable agents/workflow templates (N8N, LangGraph). FDEs become "AI consultants" for customization, not sole owners.
- RAG-as-a-Service: Unified ingestion pipelines prevent duplication. We use tools like Unstructured.io for document processing, stored in a central lakehouse (Delta Lake on Databricks).
- Vendor Agnosticism: Abstract APIs with adapters (e.g., LiteLLM) to swap models without rework. Contracts include data non-retention clauses.
This model delivers 80% of the speed with 100% governance. One client reduced AI silos by 60% in 9 months, cutting vendor spend by 25%.
Conclusion: Learn from History or Repeat It
Embedding FDEs feels innovative, but it's shadow IT in AI clothing—trading short-term velocity for long-term fragility, silos, and lock-in. Twenty years ago, we consolidated ETL sprawl into enterprise data platforms; today, we must do the same for AI.
If your organization is deploying FDEs, pause and audit: Map data flows, assess lock-in risks, and invest in centralized AI infrastructure. T10 Consulting specializes in these architectures—AI agents, N8N orchestrations, scalable RAG—designed for the enterprise, not the department.
Sources: Financial Times article linked above; T10 Consulting case studies (anonymized); industry benchmarks from Gartner and Forrester on shadow IT costs.