Anna Totterdell
Projects Director
Every enterprise software vendor now claims AI capability. It is on every roadmap, in every pitch deck, and on every pricing page. The label has become so ubiquitous that it has lost all meaning.
This is a problem if you are trying to make real decisions about where AI enablement can actually improve your operations. You need to separate the genuine applications - the ones that change measurable outcomes - from the marketing veneer that vendors apply to features that have nothing to do with AI.
Here is how to tell the difference.
What AI actually does in operations
Genuine operational AI does one or more of these things:
Classification. Taking an input - a document, an image, a data record - and categorising it based on patterns learned from training data. An AI that reads incoming emails and routes them to the correct department based on content is doing classification. An AI that categorises expense claims based on receipt images is doing classification.
Extraction. Pulling structured data from unstructured sources. An AI that reads invoices and extracts the supplier name, amount, date, and line items into structured fields is doing extraction. This replaces manual data entry and works well when the source documents follow recognisable patterns.
Prediction. Using historical data to forecast likely outcomes. An AI that predicts which customers are at risk of churning based on usage patterns, support tickets, and payment history is doing prediction. An AI that forecasts demand based on seasonal trends and leading indicators is doing prediction.
Anomaly detection. Identifying data points that do not fit the expected pattern. An AI that flags unusual transactions in your financial data, unexpected spikes in production defects, or abnormal patterns in system logs is doing anomaly detection.
Generation. Creating new content based on prompts and training data. This is where ChatGPT and similar tools sit. Useful for drafting, summarising, and ideation - but rarely the core of an operational automation.
If a vendor's AI feature does not clearly fall into one of these categories, it is probably not AI. It might be a rule-based system, a statistical model, a search algorithm, or a filtering function with a new label.
What vendors call AI but is not
Conditional logic. "If this field equals X, do Y" is not AI. It is a rule. Rules are useful and valuable but they are not intelligent and they do not learn.
Search and filtering. Sorting results by relevance, showing recommended items based on purchase history, or filtering a list based on criteria - these are often labelled AI-powered but they are database queries with a marketing upgrade.
Dashboards with insights. A dashboard that shows your top-performing products or your slowest-paying customers is doing reporting. The insights are preconfigured calculations, not learned patterns. Useful, but not AI.
Chatbots with decision trees. A chatbot that asks scripted questions and routes you based on your answers is a decision tree with a conversational interface. If it cannot handle a question it has not been explicitly programmed to answer, it is not using AI in any meaningful sense.
How to evaluate AI claims
When a vendor tells you their product uses AI, ask three questions.
What data does the AI use? If the answer is vague or the AI does not connect to your operational data, it is a feature built on generic models that may not reflect your business reality.
What does the AI produce? If the output is a recommendation that still requires a human to decide and act, calculate how much time the recommendation actually saves. Often, the human was going to make the same decision anyway - the AI just adds a confirmation step that feels modern but changes nothing.
Can you measure the difference? Before and after. If the AI is switched on, what specific metric improves? If the vendor cannot answer this with a number - not a percentage from a case study, but a prediction for your business - the AI is not operationally relevant to you.
Where AI genuinely helps in mid-market operations
Document processing. Invoices, purchase orders, contracts, delivery notes - AI can extract data from these documents with high accuracy, reducing manual entry by eighty to ninety percent for standardised formats.
Email and ticket routing. Incoming communications categorised and routed to the right team automatically. This works well when the volume is high enough to justify the setup and the categories are distinct.
Data quality monitoring. AI that continuously checks your data for duplicates, inconsistencies, missing fields, and format errors - effectively a live layer on top of your data and systems integration. This is genuinely valuable because data quality degrades over time and manual checks are incomplete.
Demand forecasting. For businesses with seasonal patterns or variable demand, AI-driven forecasting can meaningfully improve inventory management, resource planning, and cash flow prediction.
Exception handling. Flagging the items that need human attention and letting the rest flow through automatically. This is where AI adds the most value inside business automation - not by replacing the human, but by reducing the volume of work that reaches them.
The honest test
If you removed the AI label from a feature and described what it actually does, would you still want it? If yes, buy it - regardless of what it is called. If no, you are buying a label, not a capability.
The best AI in operations is invisible. It processes documents faster, routes work more accurately, and catches errors before they compound. You do not notice it working. You notice when it is not there.
That is the difference between AI that operates and AI that markets.
