Beyond the hype: building durable value from AI

For the last two years, executives have been told that chasing the application of increasingly intelligent generative AI models would allow for the automation of entire roles; entire departments, even. The current reality of AI is more mundane, but also more useful. Ambitious pilots are colliding with the reality of efficacy, safety, and ethics as evidenced by the lawsuits against Workday, State Farm, Tesla, and various health insurers – just to name a few. Budgets driven by a fear of missing out and the hype of it all are failing to demonstrate real return on investment. Models are bumping into messy data and revealing brittle processes. In all, the hype cycle is cooling right on schedule. But that’s not a failure; it’s an inevitable equilibrium between the demonstration of (admittedly) astounding new technology, and durable value.

Two truths can, and should, coexist:

1. The GenAI comedown is inevitable.
2. The underlying innovation is still extraordinary and will pay off, especially for the organizations that do the unglamorous work now.

For AI adoption, slow is smooth, smooth is fast. The firms that win won’t chase hype; many of the same organizations that not long ago were struggling to climb the analytical maturity curve towards “traditional” AI/ML capabilities have suddenly dived headlong into GenAI. I believe the current lack of P&L impact from GenAI solutions reflects that lack of preparedness. To win, organizations must industrialize the basics, measure outcomes, and wire generative capability into places where it compounds work people already do.

The comedown is healthy. Enterprises aren’t struggling because of the quality or usefulness of the foundation models. They’re struggling because value creation, at scale, requires integration as well as inspiration. When you move from a proof-of-concept to a production workflow, you inherit all the adult problems: data quality and lineage, versioning, access controls, observability, cost management, evaluation harnesses, rollback plans, user trust, and change management. None of those are solved by a more powerful, intelligent model. All of them are solved by investments in infrastructure, ML and AI Ops, and operating discipline.

If your board is asking, “Why hasn’t GenAI moved the P&L?”, the most honest answer is usually not “we picked the wrong foundation model,” or even “we picked the wrong use case.” It’s “we don’t have the scaffolding to repeatedly deploy, observe, and improve AI systems in the flow of work.” That’s fixable, but it isn’t flashy. It is an enabling investment that sets an organization up for future success.

Let me be clear: I’m not fundamentally bearish on generative AI’s potential. I am bearish on the idea that it will deliver value without the rest of the stack. The successful pattern I’m seeing is simple:

• Start with a well-defined decision or task that already creates or protects value (claims triage, outreach drafting, knowledge retrieval for support, QA summarization).
• Put guardrails and measurement around it so you can operate safely, with low risk to user trust, and prove impact.
• Then let GenAI remove friction: summarize, translate, draft, classify, and propose next actions. Keep a human-in-the-loop until safety is verified.

GenAI’s biggest near-term wins aren’t about replacing jobs, but about reducing the time-to-value for the work you already know how to do. Particularly ones document-heavy, high-volume, and formulaic. That’s a very big surface area. All that said, model improvements are inevitable. At some point research will drive a leap towards AGI.

Organizations that mistake the cooling of the hype cycle for a true winter, and view the 95% ROI failure rate with cynicism and fear, will only find themselves playing catch-up later. Invest in readiness now – focusing on these key capabilities:

Data foundations: A unified semantic layer and productized data for your critical entities, governed metadata, and automated quality checks. If your inputs wobble, your models will, too.

MLOps/LLMOps/AgentOps: Reproducible pipelines, versioned datasets, evaluation suites that reflect the task, shadow/canary deploys, safe rollbacks. You also need cost observability and smart routing: sometimes the right answer is a gradient-boosted model plus retrieval, not a 400B-parameter sledgehammer.

Governance and risk: Policy to control mapping for privacy, IP, safety. Human-in-the-loop for material decisions. Model cards and lineage so you can answer “what influenced this prediction?” without a week of archaeology. Red-team testing that includes prompt injection, data exfiltration, and tool misuse. Trust is a difficult asset to recover once it’s broken.

Operating model: Cross-functional, domain-aligned teams that integrate product management, data scientists, data engineers, and application developers. Models can’t deliver value in isolation – org design should encapsulate the ability to deliver value, soup to nuts. Incentives tied to business KPIs rather than “# of models shipped.” Change management is not optional; if you don’t rewire incentives and training, adoption stalls.

Instrumentation for outcomes: Build KPI trees that connect leading indicators (adoption, handle time, error rates) to lagging P&L metrics (revenue lift, cost takeout, risk reduction). If you can’t measure it, you can’t automate it or scale it—and you definitely can’t defend it in next year’s budget.

There’s a quiet truth in the background: classical machine learning is already a workhorse, and it’s built on these exact capabilities; GenAI is a special case of broader ML. Recommendation, routing, forecasting, anomaly detection, optimization are well-proven algorithms that move conversion, cost, and risk today. They also exercise the exact muscles you need for GenAI: clean data, reliable pipelines, outcome-linked measurement, and a habit of observing and iterating in production.

If your organization cannot ship and monitor a churn model end-to-end, you aren’t ready to run an agent that books refunds via tool use and RAG. That’s not gatekeeping; it’s responsible sequencing. Use classical ML to get reps on the fundamentals. Same stack, same governance, but proven business value today. Then, leverage GenAI where it compounds those wins.

This is not an AI winter; it is simply expectations returning to a level that lets you actually run a business. Treat AI like any other critical capability: build strong foundations, deliver small, measurable increments, and add complexity only when it creates value.