Enterprise AI should be judged on finished work

This article is authored by Tarun Dua, founder & managing director, E2E Networks.

Ask an enterprise CIO about AI and the conversation still begins with the same question it began with in 2023: Which model is best? It was a reasonable question when the technology was proving itself. It is the wrong one now. From conversations with our users running AI in production, one thing keeps surfacing: There is no “best” model, only the right model for a given piece of work, inside a given harness of tools and checks, at a cost the workload can bear. The model that tops a coding benchmark can be the wrong choice for reviewing contracts, and the one that wins on quality can lose the moment its price is multiplied by a million calls a day.

Artificial intelligence. (Thinkstock)

No enterprise buys AI to own the smartest model. It buys AI because there is work to be done: contracts to review, software to write, customer issues to resolve, supply chains to optimise. AI is a way of getting work done. Every previous technological revolution changed how work was performed. This is the first that changes who performs the knowledge work itself.

And the intelligence doing that work is losing the one quality that made it special: scarcity. Two years ago, capable models were expensive to train and rented from a handful of providers through proprietary APIs. Today an enterprise can choose from a broad set of open-weight models such as GLM 5.2, NVIDIA’s Nemotron, Google’s Gemma and Qwen, several of which now sit within striking distance of the best proprietary systems. Just as telling, smaller models have become quietly formidable. Among our own customers we see Qwen’s compact variants handling classification, extraction and support triage, jobs that needed a frontier model two years ago, at a fraction of the cost. No single vendor is a gatekeeper to capable intelligence anymore. The price of inference has collapsed as well; by one Stanford estimate it has fallen more than 280-fold in two years for the same class of model. Intelligence is turning into something an enterprise simply consumes, like bandwidth or storage.

Once intelligence stops being scarce, it stops being the advantage. What becomes scarce instead is dependable work. That is the gap enterprises are running into now: The distance between a model that answers well and a system that delivers a business outcome at nine in the morning and again at two the following night, without a person standing over it.

This is why the next phase of enterprise AI will be defined less by better models than by agentic systems. Where a model generates an answer and stops, an agent works through a multi-step task until it reaches an outcome. It is also why reliable AI work is so much harder than it looks. An agent that gets each step right 95% of the time will complete a twenty-step workflow little more than a third of the time. The arithmetic is unforgiving, and it is the reason enterprises will deploy fewer standalone chatbots and more carefully engineered teams of AI workers.

Getting a model to run is no longer hard. Keeping it running is. Thousands of agents hit the same endpoints at once. GPUs are too expensive to sit idle. Demand spikes and collapses through the day. Sensitive data has to stay inside the country. This layer is called the AI factory. The enterprise brings its data, its workflows and its business systems. The infrastructure underneath has one job: make every act of inference fast, secure and affordable at scale. An enterprise running thousands of AI workers is, in practice, running its business on inference.

All of this needs something to hold it together. That is the orchestration layer, the harness around the models. It matches each task to the model suited for it, sometimes a large frontier-class model, more often a smaller one that is good enough at a fraction of the cost. It feeds that model the right internal knowledge and passes work from one agent to the next. As models become interchangeable, the harness becomes the durable advantage, not the model inside it. Take a claims-processing agent. It pulls the policy, verifies identity, runs a fraud check, escalates when the rules demand and pays out. What decides its fate is not which model reasons best. It is whether that sequence runs the same way at midnight as it did at noon. ERP forced this discipline onto the factory floor a generation ago. Agents need the same discipline now.

And none of it survives without trust. The compliance part is well understood. An agent must stay inside policy, and its actions must be auditable. The newer risk is quieter. Run your most valuable workflows through a vendor’s closed stack and that vendor is learning your business, correction by correction. Satya Nadella said as much in July: AI labs are learning enterprises’ know-how from their own usage. Alex Karp of Palantir was blunter. Enterprises, he said, are paying to lose their competitive edge. The workflows a company automates are its moat. It should think hard about who gets to watch them run.

This changes what enterprises should measure. For years the industry has counted GPUs, parameters and benchmark scores. The measure that matters is finished work: contracts reviewed, claims processed, software shipped. Everything else is an input.

It also points to the way out of the trap Nadella and Karp describe. An enterprise that runs open-weight models on infrastructure it controls keeps its moat at home. The weights can be fine-tuned on its own data without that data leaving its custody. The corrections and workflow patterns that would otherwise teach a vendor its business stay inside its walls. This is what sovereign AI actually means. Not isolation, and not protectionism, but control over the compute, the models and the terms behind your most valuable work.

India’s opportunity here is clearer than most, and we have placed our own bet on it. At E2E Networks we run our cloud inside the country, on a GPU fleet that stretches from RTX PRO 6000 nodes suited to compact models up to the B200 systems coming online in Chennai, with an inference platform on top so that enterprises can put models into production without owning the problem of keeping them running. Match the model to the work, match the hardware to the model, and keep the whole loop under your own control.

The organisations that lead the next decade of AI will not be the ones with the smartest model. Intelligence is becoming plentiful. They will be the ones that turn it, reliably and on their own terms, into finished work.

(The views expressed are personal)

This article is authored by Tarun Dua, founder & managing director, E2E Networks.