Strategic perspectives on enterprise AI, technology architecture, and the decisions that shape how organizations adopt and operate intelligent systems.
The “RAG is dead” argument is getting ahead of the architecture change. RAG has been one of the main ways companies connect AI models to their own information, but early implementations often treated company knowledge like a pile of searchable files. The next phase requires better decisions about sources, permissions, auditability, and where the AI should go for different kinds of answers.
Read on LinkedInMost responsible AI programs start too late. By the time monitoring dashboards flag a bad answer or bias tools detect an uneven outcome, the model has already been shaped by training data, features, retrieval logic, and prompt context. The real controls have to move upstream.
Read on SubstackThe next AI planning constraint isn't compute or cost. It's power, and the politics of who pays for it. State moratoria, federal moratorium proposals, ratepayer pledges, and midterm politics are turning AI capacity into a public resource allocation problem most 2028 roadmaps haven't priced in.
Read on SubstackLife sciences companies that invested in AI infrastructure early are now reporting real operating returns. The next bottleneck isn't model performance, it's governance for agents operating inside regulated workflows.
The responsible AI controls most companies built — output review, prompt testing, human signoff — were designed for bounded systems that summarized and drafted. Once agents start taking action across live workflows, reviewing the output isn't enough. The control surface moves to action.
The AI stack is changing faster than enterprise planning cycles. Build targeted capabilities that deliver value now, then standardize based on what actually works.
AI agents can accelerate data conversion mechanics, but data quality, hidden business logic, and validation gaps still determine whether migrations succeed or fail.
Legal AI failures often start upstream, in the document layer. PDFs preserve visual fidelity but not machine-readable meaning, and that gap breaks downstream analysis.
Agentic AI is already delivering results in production, but the pattern is specific: tightly scoped problems, real system access, clear evaluation criteria, and engineering discipline.
AI in life sciences regulatory is further along than most realize. Both sponsors and regulators are now AI-powered, creating governance questions the industry hasn't answered.
Large SaaS platforms created a class of over-specialists whose entire value was platform-specific. AI-assisted development just turned that dependency into a liability.
The constraint that anchored the buy side of build vs. buy — the need for developers with specific stack expertise — is gone. The default answer has flipped.
Note-taking apps are becoming operational systems. Once a 'second brain' can execute work on schedules and triggers, it inherits the failure modes of any production system.
AI governance that lives in quarterly committee reviews can't keep pace with weekly deployments. Governance must be engineered into code, tests, and runtime controls.
With enterprise AI, the output works or it doesn't. Most organizations haven't built the evaluation environments needed to hold AI to that standard.
Inference costs dropped 99% in a year, but enterprise AI still isn't accelerating. The bottleneck is the cost of wiring AI into production systems.
The average enterprise runs 12 AI agents, half in silos. That's often the right default — connect agents only when there's a measurable cross-team workflow benefit.
Model selection shifted from vendor procurement to portfolio management. Different tasks need different cost, latency, accuracy, and compliance tradeoffs.
Enterprise AI often fails because organizations commit to a model that isn't good enough for the task and compensate with process instead of correcting the decision.
AI's highest-value enterprise work isn't in new systems — it's inside the legacy platforms everything depends on, where complexity has compounded for years.
Model choice still matters, but the harder problem is turning capable models into dependable systems — reliable tool use, structured outputs, and operational control over time.
In life sciences, the real AI advantage comes from owning your models, your data, and the infrastructure that lets them evolve — not from choosing the right vendor.
Tech companies proved that the gap between good models and real impact is operations — versioning, retraining, governance, monitoring, and rollback at production scale.
AI-assisted knowledge work shows 10-30% productivity gains — real but uneven. Organizational friction absorbs much of the speed increase before it reaches outcomes.
Multi-agent swarms reproduce the worst dynamics of organizational bureaucracy at machine speed. Small, structured cognitive architectures — not more agents — are what reach production.
Enterprises don't need giant all-knowing models. A portfolio of small, purpose-built AI specialists delivers faster time to value and cleaner ROI.
The autonomous-agent vision driving today's vendor pitch isn't coming from enterprises — it's coming from vendors. What's actually working inside enterprises is AI-assisted engineering: integrations, legacy cleanup, test generation, accelerated debugging. Deliberate modernization, not hands-off cross-system automation, is where the real progress is.
The real ROI killer in enterprise AI isn't the model — it's the hidden cost of corrections that never show up in dashboards or official reporting.
Most GenAI pilots work in isolation. Scaling to production agents requires integration into legacy systems that were never designed for modern AI.
Most organizations measure AI ROI with metrics that flatten operational reality. Tracking correction rates and improvement trendlines reveals what dashboards miss.
AI coding tools revealed that the way we scale development teams is structurally inefficient. Small, high-context teams with AI outperform large coordinated groups.
Featured in The Data Wire on why enterprises must treat token cost and model design as architectural constraints from day one — and why smaller, specialized models often win.
Foundation models are commoditizing fast. The real competitive edge is integration — how efficiently you wire AI into workflows, manage token costs, and architect for production.
AI projects fail structurally, not culturally. Coordination overhead, disconnected compliance teams, and too many stakeholders stall initiatives before the model ever becomes the issue.
AI-native labs aren't built by buying more AI. They're built on data discipline — governed, contextualized lab data products, an edge-aware foundation, and a layered model stack that fits regulated science.
Arguing for small composable services before 2010, well before microservices had a name. The technical case was only half the story — the other half was making engineers run what they built.