Most Enterprises Are Investing In AI In The Wrong Order

Hiren is CTO & co-founder of Simform. With 15+ years in tech engineering, he leads teams in cloud architecture, MLOps and secure agentic AI.

gettyIn many enterprise AI programs that I have seen, the model is not what blocks production. It’s the layer underneath it. In those programs, the data was neither trustworthy nor was the platform reliable. What AI needs to run was simply not there.

S&P Global Market Intelligence surveyed more than 1,000 enterprises across North America and Europe and found that 42% of companies abandoned the majority of their AI initiatives before reaching production. The year before, that number was 17%. On average, 46% of proof-of-concept projects never made it out. That gap between adoption and production is where the foundation problem shows up.

I use three dependent layers when evaluating the readiness of an AI initiative to scale: a governed, trustworthy data foundation, a reliable and observable platform and governed agentic systems on top.

Visible AI capability competes better for capital than invisible infrastructure. Model licenses and agent capabilities go in first because they produce quick results leadership can point to before the quarter ends. The data foundation comes later, if at all, because a working agent is easier to demonstrate than data lineage, access controls or governance maturity.​

In a recent client engagement with a manufacturing company, the team approved an AI initiative with strong talent, well-scoped problem and real investment. Production-scale capital went into AI capability before the data foundation that should have supported it. The foundation investment did not get approved until months into the program. By the time the gap surfaced, nothing the model produced could be trusted because the data lineage was missing.

The decision-makers who set that investment order set the architecture. They placed agent capability before data readiness. Engineering spent the following months working around a foundation that had never been built, absorbing unplanned effort the original schedule had never accounted for.

Correct sequencing reduces visible feature velocity early because some investment goes into reusable dependencies rather than deployable capability, and that is the trade-off. The alternative is capability spend that outpaces its own dependencies.

But someone has to make that case to the people setting the investment order, before the money moves. The conversation must help the board pair every capability investment with the dependency it requires, the evidence that dependency is ready and the cost of skipping it. That is not a technical argument. It is a time-to-return argument: “Here is what we are betting on, what it requires and what the rework will cost in engineering time and delayed returns if the dependency is not there.​”

A readiness gate governs scale: how much production investment, traffic, autonomy or scope a layer receives, and not whether teams can experiment. Before the next layer receives materially more investment, the previous one has to demonstrate it is ready.

The first gate is the data foundation. It is ready when the team can confirm authoritative sources, data lineage, freshness, access controls and ownership for the information the workflow actually depends on. The question it has to answer is simple: Is this output trustworthy?

The second gate is the platform. It is ready when it meets a predetermined operating envelope—expected peak load plus headroom, with latency, availability, observability, cost per transaction and recovery time all within predefined thresholds. The question it has to answer: Does this hold under load?

The third gate is the agentic layer. It is ready when the accountability questions have written answers before the first workflow goes live: what the agent is authorized to do, when a human takes over, whether consequential actions can be traced or rolled back, who owns the outcome if the decision is wrong and what remediation looks like.

Walking into an investment conversation means arriving with a readiness scorecard for each layer and a position: The next layer does not get investment until the previous one can demonstrate it.

That sounds clean on paper. In practice, it means holding the line against boards that want AI capability before the infrastructure exists to support it, and against business units that have already announced timelines.

There is a real cost to sequencing correctly. Quarterly reporting cycles do not reward discipline that pays off over a longer horizon than a single reporting period. Whoever holds the line will face pressure. The argument for sequence has to be made before the investment meeting, with enough specificity that the board can understand what they are actually choosing between.

The enterprises moving ahead on AI are not the ones that spent the most. They are the ones that spent in the right order.​

When an AI initiative stalls, I look at investment sequence before assuming the model is the problem. When I dig into the blocking issue, it is almost always in what was not invested underneath the capability, not the capability itself.

Capital allocation is what determines whether those dependencies exist when engineering needs them.

Before the next AI budget is set, ask one question that almost never gets asked: Which layer are we scaling next, and what evidence says the dependencies underneath it are ready?

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?