Almost everyone who has tried generative AI recognises the same thing: it writes faster, summarises meetings, drafts emails and helps with coding. And yet, when you look at the bottom line, the effect is hard to see. That distance between personal productivity and company profit has a fairly practical explanation: most of the decisions that move a business’s margin are not made with text, but with structured data — the tables of customers, orders, delivery notes, stock and collections that live inside the ERP.
In a recent article, SAP News Center ("The Gap That No LLM Can Close", news.sap.com) raised exactly this problem, citing a McKinsey study: around 80% of AI users say they gain personal productivity, but only 6% of companies state that AI has significantly increased their operating profit
Why LLMs struggle with structured data
A language model is trained to predict text. It is brilliant at explaining, drafting and conversing, and with techniques such as RAG it can query a company’s internal documentation. But a table is not text: it is a system of relationships. The value of a row depends on the column next to it, on the period, on the customer it belongs to and on another five tables that go with it.
The cited source points out that various independent analyses show noticeably lower LLM performance on reasoning tasks over tables than on language tasks
Language and prediction are two different tools
The useful conclusion is not "LLMs are useless". It is that they solve a different problem. Natural language is the interface; prediction over structured data is the engine. The companies getting real returns combine both: the language model to interact, explain and draft; analytical models to anticipate what is going to happen with collections, stock or demand.
The market’s move: models designed for tables
There is a clear trend in 2025-2026: alongside language models, tabular foundation models are appearing, trained specifically to find patterns in structured data, as well as models trained on time series. The promise is interesting for any finance department: instead of developing, training and maintaining a different model for each question (one for demand forecasting, another for default risk, another for prioritising sales opportunities), you start from a general model applicable to several business questions.
SAP, for example, has announced moves along these lines with the acquisition of Prior Labs and its family of tabular models
What this means if your company works with Business Central
Here is the actionable part. If your ERP is Microsoft Dynamics 365 Business Central —or you are migrating from Navision—, you already have the asset these models need: years of operations recorded in a single system, with well-defined relationships between customers, items, orders, warehouses and accounting entries.
Some examples of questions that are answered with structured data and not with text:
- Default risk. Which customers are most likely to pay late, cross-referencing collection history, ordering patterns and delivery behaviour.
- Demand and stock forecasting. Which items need replenishing and in which warehouse, in a distributor with several centres.
- Turnover and obsolescence. Which inventory is sitting idle before the annual stocktake says so.
- Occupancy and rates. In hospitality, anticipating occupancy gaps by room type and channel.
- Fleet maintenance. Detecting vehicles with an anomalous cost per kilometre before the major breakdown.
At Tisa we have worked on these scenarios in sectors such as retail, distribution, construction, the food industry, sports centres and fleet management (you can see a summary at grupotisa.com/casos-de-exito). The pattern repeats itself: the bottleneck is rarely the algorithm; it is the quality and traceability of the data.
Before investing in predictive AI: four checks
- Do you have enough reliable history? A model that predicts defaults needs years of collections recorded with real dates, not "invoice date" used as an approximation for everything.
- Is the data in a single place? If sales lives in a spreadsheet, the warehouse in another application and accounting in the ERP, the first project is not an AI project: it is a consolidation project. This is where a well-implemented ERP pays for itself.
- Is there governance? Who can see what, with what quality and under which regulations. Without this, AI becomes a risk instead of an advantage.
- Is the use case measurable? "Reduce the average collection period by X days" is an objective. "Doing AI" is not.
How we bring this down to earth at Tisa
Our approach is unspectacular and quite effective: first we put the foundation in order —implementing or upgrading Business Central, consolidating sources, dashboards in Power BI—, and then we choose one use case with clear economic impact and take it to production. Our AI and data line includes precisely that: identifying the relevant data sets, defining governance protocols, evaluating tools according to investment capacity and developing a use case tailored to the business, with training for the team.
The first wave of generative AI has made people more productive. The next one, the one that really moves the bottom line, consists of anticipating: seeing the late payment, the stockout or the margin drop before it happens. And that is played out in your ERP’s structured data.
Shall we talk about your case?
If you want to know what your company can predict today with the data it already has, we offer a no-obligation assessment. Call us on (+34) 971 305 885, write to us at info@grupotisa.com or visit grupotisa.com. More than 30 years implementing management solutions is enough to tell a fad from an opportunity.
Reference source: SAP News Center, "The Gap That No LLM Can Close" (news.sap.com). Original article in English; this text is Tisa’s own interpretation, not a translation.
