The End of "Good Enough" AI: Why Business Data Is the Real Battleground for the Next Wave of Artificial Intelligence

For the last few years, the story of artificial intelligence has been a story about models. Bigger training runs. Higher scores on benchmark tests. Demos that could write a poem or plan a vacation in seconds. It was impressive, and it was also, for most companies, strangely useless. A tool that can explain quantum physics in plain English still cannot tell a supply chain manager whether a shipment will arrive before the factory line stops.

That gap is where the next phase of AI is being fought, and it is the gap that SAP's Philipp Herzig puts at the center of the conversation. The thesis is simple and it flips the industry's priorities on their head: the model is no longer the hard part. The business data is. Anyone can access a powerful language model. Almost no one can hand that model a clean, permissioned, real-time picture of how their company actually runs.

General Models Know the World. They Don't Know Your World

A general-purpose AI model is trained on the public internet. That makes it broadly knowledgeable and specifically blind. It has read about invoicing. It has never seen your invoices. It has absorbed thousands of articles about logistics. It has no idea that your warehouse in a specific region uses a different product code than your warehouse two regions over.

This is not a minor technical detail. It is the whole ballgame. In a business setting, an AI answer that is 90 percent right is not 90 percent useful, it is a liability. A financial report that is almost correct is a compliance problem. A demand forecast that is slightly off is a warehouse full of unsold goods. The bar for enterprise AI is not "sounds smart." It is "is right, can prove it, and can be audited."

The companies that solve this will not be the ones with the most clever model. They will be the ones whose AI can see the actual numbers, in the actual systems, under the actual rules about who is allowed to look at what.

Why Business Data Is So Much Harder Than It Looks

People outside large organizations tend to imagine that companies sit on tidy spreadsheets. The reality is closer to an archaeological dig. Data lives in dozens of systems built in different decades. Customer records disagree with billing records. A field called "status" means one thing in one system and something else entirely in another. Some of the most important knowledge in the company was never written down at all, it lives in the head of a planner who has done the job for twenty years.

So when the industry talks about "AI built for business data," it is really talking about three layers of work that have nothing to do with model training:

• Meaning. A shared dictionary that tells the AI what a "customer," an "order," or a "margin" actually means across the whole company, so the same question gets the same answer everywhere.
• Rules. A hard-wired understanding of who can see what. An AI assistant that happily summarizes salary data for anyone who asks is not an assistant; it is a security breach with a friendly voice.
• Freshness. Access to live operational numbers, not a snapshot from last quarter. Business decisions happen in the present tense.

None of this is glamorous. All of it is what separates a demo from a deployment. And it explains why the most valuable AI companies of the next few years may well be the ones sitting on top of decades of business process knowledge rather than the ones with the flashiest chatbot.

The Semantic Layer Is the New Moat

Here is the strategic shift worth internalizing. When everyone has access to roughly comparable frontier models, the model stops being a differentiator and becomes a utility, like electricity or bandwidth. What you build around the model becomes the advantage.

That "around" is a semantic layer: a structured map of the business that the AI must consult before it answers. Instead of guessing at what a number means, the AI looks it up. Instead of inventing a reason for a sales dip, it queries the actual pipeline, the actual returns, the actual regional split, and then explains what it found.

This matters because it changes the failure mode of AI. A model guessing from memory will confidently make things up. A model grounded in a governed data layer has somewhere to check its work. Grounding is not a feature. It is the difference between a tool you trust and a tool you babysit.

From Answering Questions to Taking Action

The second big shift is that AI is moving from answering to doing. The early generation of enterprise AI sat in a chat box and waited to be asked something. The next generation is being wired into workflows, closing out a reconciliation, flagging a delayed shipment, drafting the follow-up email, and routing it to the right person for approval.

This raises the stakes dramatically. A wrong answer costs you thirty seconds of annoyance. A wrong action, a payment released, an order cancelled, a discount granted, costs real money. Which is why the design pattern emerging across serious enterprise AI looks less like a genius assistant and more like a well-run workplace:

• Clear scope. The AI is given a defined job with a defined boundary, not a vague mandate to "help with operations."
• Human checkpoints. Actions above a certain value or below a certain confidence level go to a person.
• Full traceability. Every step is logged, so when something goes wrong you can reconstruct exactly why.
• Graceful failure. When the AI does not know, it says so and escalates, rather than improvising.

This is unglamorous engineering, and it is exactly what separates the companies that get value from AI from those that get a pilot project and a slide deck.

What This Means for the Future of AI

If the business-data thesis is right, several assumptions currently baked into the industry will need to change.

1. Evaluation will move from benchmarks to business outcomes

Nobody in a boardroom cares about a model's score on a reasoning test. They care whether invoice exceptions get resolved faster, whether forecasting error drops, whether the close takes fewer days. Expect the measurement of AI to shift decisively toward operational metrics you can put in a financial report.

2. Smaller, specialized systems will win in the enterprise

A model that is excellent at supply chain reasoning and nothing else may beat a general model at supply chain work, and cost far less to run. The future is likely a mix: large general models for open-ended thinking, smaller tuned systems for high-volume, high-precision tasks inside a specific domain.

3. Data quality becomes the binding constraint

Every company that has tried to deploy AI at scale has discovered the same uncomfortable truth: the project is really a data cleanup project wearing an AI costume. The organizations that treat data governance as a first-class strategic investment will pull ahead. The rest will keep hitting a ceiling they cannot see.

4. Trust becomes a product feature

As AI takes on more consequential work, the ability to show your reasoning, cite your sources, and respect access boundaries becomes as important as raw capability. Regulators, auditors, and customers will all ask the same question: how do you know? Systems designed to answer that question will have a durable advantage.

Practical Steps for Business Leaders

For anyone trying to turn this from theory into a plan, a few moves matter more than the rest:

• Pick a workflow, not a technology. Start with a process that is slow, repetitive, and measurable. Fix that. Then expand.
• Audit your definitions first. Before you buy anything, find out how many different definitions of "customer" exist in your company. The answer will tell you how much work is ahead.
• Design permissions before prompts. Decide who the AI can act for and what it can see. Retrofit security later and you will regret it.
• Measure the boring numbers. Hours saved, error rates, cycle times. These are what justify the next phase of investment.
• Keep a human in the loop for anything irreversible. The cost of a checkpoint is small. The cost of an unsupervised mistake is not.

The Bigger Picture: AI That Fits the Work

There is a cultural shift buried in all of this. The first era of AI asked, "What can this technology do?" The next era asks, "What does the work actually require?" That is a healthier question, and a harder one. It forces companies to confront messy realities, legacy systems, inconsistent data, unclear ownership, nervous employees, instead of skipping straight to the exciting part.

It also has implications beyond the enterprise. If AI's value comes from being deeply embedded in real operations rather than from dazzling general intelligence, then the technology will spread more slowly and more thoroughly than the headlines suggest. Jobs will not vanish in a single dramatic moment. They will be reshaped task by task, with the routine parts automated first and the judgment-heavy parts preserved longer. That is a less scary story than the one usually told, and probably a more accurate one.

The companies that come out ahead will not be those that waited for a perfect model. They will be those that spent the last few years making their data trustworthy, their definitions consistent, and their processes legible, so that when a capable model showed up, it had something real to work with.

In that sense, the future of AI is not really about AI at all. It is about whether an organization actually understands itself well enough to explain its own business to a machine. That is a much older problem. It just finally has an urgent deadline.