Trending Topics
Trending
Enterprise AI doesn’t need better demos: It needs better benchmarks
This article is part of our Opinions section, where we invite industry professionals to share their views on the most pressing technology questions of our time.
Enterprise technology leaders are becoming more cautious about AI, and I think that is a good thing. The excitement has not disappeared, but the questions have changed. Instead of asking for a demo, buyers want to know whether a system will work in their business. Anyone who has lived through an enterprise software demo knows the trick. The happy path is always beautiful.
There is a large gap between an AI system that performs well in a controlled environment and one that can operate reliably in production. In a demo, the data is usually curated, the task is known in advance, and the success criteria are clear. In an enterprise, data is distributed, incomplete, and constantly changing, and is governed by permissions that vary by role, team, and business function.
A 2025 MIT study on the state of AI in business found that 95% of enterprise GenAI pilots fail to deliver their intended outcomes – analysing roughly 300 AI projects across 150 companies.
The problem is not that today’s AI systems are incapable. The problem is that we have no consistent way to measure whether an AI agent is ready for the complexity of enterprise work, and some simply aren’t. Most evaluations focus on whether a system can produce an impressive answer. Enterprise buyers need to know whether it can produce the right answer, from the right source, at a predictable cost, with an audit trail the business can trust.
The purpose of a benchmark isn’t to tell you which AI is best. It’s to tell you whether it’s ready.
Why enterprise AI is harder to test
Most AI benchmarks were built to measure models or agents in controlled environments. They ask whether a model can solve a problem or whether an agent can complete a task, such as writing code or booking a flight. Those are useful tests, but enterprise environments are fundamentally different. In many benchmarks, the task has a clear finish line: either the booking was made, or it was not. Enterprise work is rarely that simple.
Take a simple example. A sales leader asks: Which strategic customers have open support tickets linked to high-priority engineering issues?
The answer is rarely straightforward. The system first needs to understand what “strategic” means in that organisation. It then must connect CRM records, support tickets, and engineering issues that may sit in different systems with different schemas before determining what qualifies as “high priority.” The answer must be assembled from structured and unstructured data across multiple systems, and that data is constantly changing.
That is why enterprise AI cannot be tested only on whether it can complete a task in a clean environment. It must also be tested on whether it can produce the same answer as the volume of surrounding data grows. In a production environment, the correct answer does not change because there are more tickets, more documents, or more historical records. The challenge is finding the same answer amid much more noise.
I believe that enterprise AI needs a new generation of benchmarks because the current ones don’t measure what matters in production.
What production readiness means
For an AI agent to be useful in an enterprise, three qualities matter more than a polished demo: precision, efficiency, and safety.
- Precision means returning the correct answer from the correct source and being able to show how the system arrived at it.
- Efficiency matters because token costs scale quickly. A system that searches broadly across every connected source may perform well in a pilot but become prohibitively expensive as enterprise data grows.
- Safety means respecting permissions, maintaining audit trails, and preventing unauthorized access to sensitive information. These are operational requirements, not optional features.
Production readiness depends on all three.
The model is only part of the answer
In enterprise environments, the architecture around the model often matters more than the model itself. How the system retrieves data, maps relationships, applies permissions, reasons over structured information, and verifies its own answers will usually determine whether it can be trusted in production.
Our own benchmarking work confirmed this in a way I did not expect. When we moved from one version of the same model family to a newer release, our benchmark accuracy dropped. The newer model had learned to take shortcuts it should not be taking. It improved on general tasks and regressed on the specific enterprise work we were measuring. That finding reinforced something I now consider a core principle: do not evaluate an AI platform by the model it runs today. Evaluate it by the architecture that will continue to work as the model changes.
A useful benchmark should answer questions buyers cannot answer today:
- Can the agent cope with realistic enterprise workloads?
- Where does performance begin to degrade?
- Does accuracy remain stable as the amount of irrelevant data grows?
- Can the same answer be reproduced consistently?
Most benchmarks tell you who performed best. Enterprise buyers need something different. They need to know where a system starts to break down. Every enterprise platform has limits. A benchmark that only reports the winner tells you very little about whether that system will cope with your own environment.
That last point is important. In a small benchmark, an agent may succeed by loading most of the relevant data into context. In a real enterprise, that’s not sustainable. The correct answer should remain stable even when irrelevant data surrounds it. If accuracy drops as data volume increases, the system is no longer consistently finding the answer. It is compensating by relying on larger context windows rather than retrieving the information it needs.
Enterprise AI needs a common yardstick
In the early days of the database market, every vendor claimed to be faster than the others. Benchmarks were often private, selective, and difficult to compare. Customers had no neutral basis for evaluation.
The Transaction Processing Performance Council (TPC) changed that by introducing common, auditable standards. Vendors still competed, but they competed against the same yardstick. That did more than help buyers compare products. It accelerated the database market because performance claims could be tested instead of simply asserted.
Enterprise AI is reaching the same point. Every vendor says its agent is enterprise-ready, and demos can be designed to look convincing. But buyers need a way to test whether those claims hold up under conditions that resemble their own business.
That does not mean every benchmark will answer every question. Different industries will need different datasets and scenarios. A financial services firm will care about workflows different from those of a healthcare provider or a software company. But the principle should be the same, and the benchmark should test realistic work, data, permissions, and scale.
It should also be reproducible. If a vendor claims strong performance, customers should be able to understand how that result was achieved. What data was used? What queries were asked? What counted as success? Could another system be tested under the same conditions?
Without that kind of discipline, the enterprise AI market will stay too dependent on demonstrations and claims that cannot be independently verified.
What buyers should ask now?
For enterprise buyers, the practical questions are becoming clearer. Can the system find the right answer when the data sits across multiple systems? Can it show where that answer came from? Does accuracy hold as the dataset grows? Does cost remain predictable as usage increases? Can permissions be enforced throughout retrieval, reasoning, and action? Can the whole process be audited afterwards?
Enterprise AI is becoming easier to demonstrate and harder to evaluate. The questions buyers need to ask are not complicated, but they require evidence that most vendor demonstrations lack. The database industry learned three decades ago that claims without a common standard are just noise.
The buyers who demand measurable evidence now, before committing to a platform, will avoid finding out the hard way that a convincing demo and a production-ready system are not the same thing.
Related articles
About The Author
