A procurement manager in Leeds opens a chat window on a Tuesday morning. He asks a very capable model to draft a reply to a supplier who has missed a delivery date again. The answer comes back in seconds, fluent and courteous. It is also wrong. This supplier sits inside a framework agreement with a service credit clause, and nobody told the model that. So the manager rewrites the email himself and reports the failure to no one. The scene is illustrative. But the pattern is a familiar one, and it shows why AI business context deserves as much attention as model size.
The instinct, when this happens, is to reach for a smarter model. That is understandable, and models do matter. Frontier models already perform strongly on many academic and professional benchmarks. Still, they cannot know that your Leeds framework agreement exists.
In January 2026 the training-data company Mercor published APEX-Agents. The benchmark asked leading models to complete professional tasks in consulting, investment banking and law. Mercor built a simulated workplace around them, with files and tools scattered the way they are in real firms. Yet the best first-attempt score was 24%, from Gemini 3 Flash. GPT-5.2 managed 23%.
What the benchmark actually tested
Read the design, though, and the result looks less like a verdict on intelligence alone. Mercor’s chief executive, Brendan Foody, told TechCrunch that nobody works with “all the context in one place”. In real life, he added, “you’re operating across Slack and Google Drive and all these other tools.”
The models did not face exotic questions. Instead, instructions left gaps, and the files sat apart from the conversations. Mercor said agents often failed to handle that ambiguity, find the right file or hold context across a whole workflow.
So the benchmark tested reasoning under the conditions of real work: fragmented context, ambiguous instructions and information scattered across systems. A model can reason well and still fail if it cannot find the right file.
Why AI business context sits beside model capability
MIT’s Project NANDA came at the same problem from enterprise deployments. Its preliminary 2025 report reviewed more than 300 publicly disclosed AI initiatives. The team interviewed representatives from 52 organisations and surveyed 153 senior leaders. Only 5% of integrated AI pilots, it reported, were extracting millions in value. Most showed no measurable P&L impact. But the authors described their figures as directional, drawn from interviews rather than company reporting.
The report’s primary explanation was what its authors called the learning gap. Most systems, they wrote, “do not retain feedback, adapt to context, or improve over time”. Respondents also listed model output quality among their concerns, so the report does not say models are beside the point.
Read plainly, though, a learning gap looks less like a shortage of raw intelligence than a shortage of memory and integration. A system that retains no feedback can answer today’s question exactly as it answered March’s. That happens because nobody wired the outcome back into it.
British firms describe a related obstacle from another direction. In July 2026 the Office for National Statistics reported that 41% of UK businesses with at least 10 employees had met no barrier to AI adoption in the previous three months. Among firms that did, lack of expertise was one of the most frequently cited constraints. Earlier ONS research also listed difficulty identifying business use cases, cost and a lack of expertise as the most commonly reported factors. None of those three concerns the model itself.
Where AI business context actually lives
Ask a finance director where the company’s rules live and someone will point at a policy folder. The folder is real. It rarely holds all of them.
In many companies the rest sit in a few heads. Picture a credit controller who knows which three accounts always pay late and always pay. Picture a sales director who discounts to a floor that appears on no price list. Picture an operations manager who knows Tuesday’s report lies until the depot closes its shift. None of these people exists. The habits will be familiar.
None of that is proprietary genius. It is simply the accumulated shape of one business. In 2025 Salesforce’s research team tested the leading models of the time on CRMArena-Pro, nineteen expert-validated tasks across sales, service and quoting. Agents managed about 58% success on single-turn tasks. That fell to about 35% once a conversation ran over several turns.
The pattern inside those numbers matters. The best agents handled defined workflow execution well, above 83% on single-turn tasks. But they struggled with multi-turn reasoning and showed near-zero inherent confidentiality awareness. Salesforce, which sells its own agents, summed up the lesson this way: generic agents lack “the structured data, workflows, and safeguards needed to operate in high-stakes, real-world scenarios”.
A model trained on the world at large arrives knowing a great deal. It does not arrive knowing your contract.
The work nobody sold you
There is an uncomfortable implication here for anyone drawing up an AI budget. The model is only part of the bill. Much of the effort goes into writing down how the company works, in a form a machine can act on.
That means naming which system holds the truth when two disagree. It means defining what counts as a resolved query, an at-risk account, a closed month. It also means deciding who approves an action and where the audit trail lands. This work looks like operations rather than technology, so it can slip between teams.
Boards feel the same gap from a different seat. A director who signs off an automated decision needs to know which rule the machine applied and who set it. That question has no answer while the rule lives only in someone’s habits. So the documentation nobody wanted to write can become the governance record somebody will eventually ask for.
Riophany spends most of its time here, connecting CRM, finance and service systems. The aim is that an agent inherits the company’s rules instead of guessing at them. We described the same discipline in a piece on the operational detail hiding in a maintained spreadsheet. It appears again in an examination of what a voice deployment has to survive before it reaches a customer.
The version of this that works
Notice what the MIT report says works. Its winning vendors focus on narrow, high-value use cases and integrate deeply into workflows. And buyers who get value judge tools on business outcomes rather than software benchmarks.
That is a modest ambition, and it often pays. A system that knows one process well can beat one that knows everything vaguely. Business decisions turn on specifics: this customer, this contract, this exception. So strong AI business context is those specifics, held where software can reach them.
None of this argues against capable models. Better reasoning helps, and it helps most where the ground rules are already clear. But model capability is only one layer of enterprise AI performance. Once a model is capable enough, access to the right business context, workflows, permissions, institutional knowledge and feedback becomes part of the intelligence of the finished system.
The procurement manager in Leeds did not need a model that had read more of the internet. Instead, he needed one that had read the framework agreement. Those are different projects, and only one of them comes off the shelf.