AI-Native Enterprises Need Data Foundations, Cost Visibility and Constant Tuning
- Manu Swami
- 370 views
Data readiness, inference cost tracking and continuous tuning form the core of a sustainable enterprise AI strategy, according to Sonata Software's CTO.

Enterprises are racing to adopt AI, and the bills are arriving faster than most budgets were built to handle. Gartner says AI-native companies are growing up to 40% faster than traditional SaaS companies, but that growth is colliding with a cost problem. Manu Swami, CTO of Sonata Software, argues that a workable enterprise playbook rests on three disciplines: a data foundation, intelligent economics and continuous optimization.
A bill that can dwarf the project
One company was billed half a billion dollars in a single month for AI usage. That may be an outlier, but Swami reads it as an early warning. The move from flat subscriptions to usage-based pricing has made AI spending volatile in a way most IT and finance processes weren't designed to track. Model prices per unit are falling, yet the cost of completing a real task often lands above what teams expected. A single workflow can trigger many model calls, and those calls add up quickly at enterprise scale.
The instinctive fix is a hard spending cap. Swami's point is that a cap set without regard to what each task is for limits innovation along with the waste. Here's the thing: the goal isn't to make AI cheaper at any cost. It's to make sure every dollar buys a measurable result.
Fix the data before the model
Swami is blunt about where modernization efforts stall. Intelligent foundations come first: the unglamorous data platforms and responsible AI guardrails. Skip them, and every later discipline inherits the same blind spots.
Gartner estimates that through 2026, 60% of AI projects will be abandoned because the underlying data was never AI-ready. The pattern is familiar. A model looks strong in a demo because someone spent a week quietly cleaning the dataset behind the scenes, then falls apart the moment it reaches production. That's rarely a model problem. It's a foundation problem, usually fragmented systems that were never reconciled.
The practical implication is that data modernization isn't a warm-up before the "real" AI work. It's the infrastructure everything else sits on. Teams that treat data readiness as an ongoing discipline, rather than a one-time cleanup, give the other pillars something solid to stand on.
Govern spend instead of rationing it
Some use cases will push consumption costs up far faster than others, and high usage is entirely justifiable when it produces real value. Jellyfish data cited in the same discussion suggests the most productive engineers are also the heaviest token consumers. High spend and high value often move together.
That changes the leadership job. The task is to govern AI, not ration it. Intelligent economics treats every AI-driven action as a unit of cost and value, weighing what a task costs to run against the business outcome it unlocks.
The obstacle is visibility. According to a recent KPMG report, only 26% of organizations have real-time visibility into the cost of running AI at scale. You can't govern a line item you can't see, and many teams are still finding out what their AI workloads cost after the invoice lands.
Inference economics and model routing
Inference economics refers to the recurring operating cost of every task or query handed to AI. Managing it takes the same rigor teams apply to any other production workload. The payoff can be substantial. Harvey, a legal AI company, recently tested a hybrid setup that outperformed running the frontier model on everything, at 39% of the cost.
The lesson generalizes. The most expensive model isn't always the most accurate for a given task, and routing work intelligently can beat brute force on both quality and price. In practice, teams in this position typically start by sorting their AI tasks by complexity and value, then testing whether lighter models handle the simpler ones acceptably. The savings tend to be largest on high-volume, low-complexity work.
Optimization never really finishes
Cost visibility only tells part of the story. Enterprises also need to know whether AI is moving the metrics that matter, and that question doesn't end when a workflow goes live. It just changes shape.
Continuous optimization means treating AI systems less like projects with a finish line and more like infrastructure that has to be watched, tuned and adjusted as conditions change. That starts with anchoring initiatives to outcomes leadership actually cares about. It also means putting as much energy into adoption as into deployment. The best-governed initiative delivers little if the people it's meant for drift back to old workflows within a few months.
This is where the three pillars connect. Continuous optimization is how an enterprise checks whether its data foundation and cost governance are compounding into real value. Swami's summary for 2026 is direct: business value from AI is the sum of intelligent workflows, cost efficiency and adoption.
What to check before the next budget cycle
For IT leaders, the near-term questions are concrete. Can you see AI cost broken down by workflow or task, in close to real time? Are the datasets behind your pilots actually production-ready, or only cleaned for the demo? Do you have a way to compare cheaper and more expensive models on the same task before committing to one? And are you measuring adoption, not just whether a tool was deployed?
My read is that the organizations pulling ahead won't be the ones with the most advanced models. They'll be the ones that can tie each AI dollar to a result, and that can tell within a quarter whether the investment is paying off. The cost crisis is real, but it's also a useful forcing function for getting the basics right.