AI in enterprise software: separating the useful from the hype
We have reviewed dozens of 'AI-powered' enterprise tools. Most add AI features because it is marketable, not because it adds value. Here is how to tell the difference.
The problem with 'AI-powered'
In 2024 and 2025, 'AI-powered' became a mandatory marketing qualifier for enterprise software — applied to everything from accounting platforms to HR systems, regardless of whether the AI feature solves a real problem. The result is a market where genuine AI capability is indistinguishable from LLM wrappers bolted onto existing products to tick a marketing checkbox. Enterprise buyers are left navigating vendor claims with no reliable way to evaluate them.
Three useful categories of enterprise AI
After reviewing dozens of enterprise products with AI claims, we found that genuinely useful AI features fall into three categories: document processing and information extraction (turning unstructured data into structured records — genuinely valuable); anomaly detection and alerting (identifying patterns in operational data that humans would miss — useful in the right context); and decision augmentation (providing recommendations that humans validate, not decisions that execute automatically — appropriate for high-stakes contexts). Everything else is usually a search function renamed as AI or an LLM chat interface on top of a database.
The evaluation test
When a vendor tells you their product is AI-powered, ask them three specific questions. One: what specific decision or task does the AI perform, and what data does it use? Two: how do you validate that the AI output is accurate before it acts on your data? Three: what happens when the AI is wrong — what is the correction workflow? A vendor with real AI capability will have concrete answers. A vendor with a marketing AI feature will become vague.
Where we apply AI in our own systems
In systems we build for clients, we apply AI in specific, well-defined areas: document classification for compliance workflows, where structured output from AI reduces manual data entry by significant margins; anomaly flagging in financial data pipelines, where AI identifies transactions that warrant human review; and natural language search interfaces for complex data sets, where AI makes existing data more accessible without changing how it is stored. We do not use AI as a general-purpose feature or add it to systems where a deterministic rule would do the same job more reliably.
The reliability question
Enterprise systems have one requirement that consumer AI applications do not: they must produce consistent, auditable, defensible outputs. An AI that is correct 95% of the time is acceptable in a consumer recommendation engine. In an enterprise compliance system, that 5% error rate represents real financial and regulatory risk. Any AI feature in an enterprise system must have a clear boundary between AI-suggested and AI-decided — and a human in the loop for everything that matters.
We are pragmatic about AI — enthusiastic about it where it solves real problems, and sceptical of it where it adds complexity without value. If you are evaluating an AI feature for your systems and want a second opinion from engineers who build in this space, we are happy to have that conversation.