The next gains from AI come from the context you feed it, not the model you pick. Most of that context is still sitting in your team's heads.
Most of what a business knows never gets written down. It sits in the head of whoever has done the job longest, and only comes out in the moment they make a call or are asked a question. Which supplier to use when the usual one is running late. Why that quote got discounted. Which insurer will actually write an unusual risk. None of that is in your CRM, its never touched your cloud storage, so none of it is available to an AI tool, which is why most AI tools give answers that are technically fine and practically useless.
We are writing this from an Australian SME’s seat, because the two clearest arguments on this came out of the big end of town and are being read as a big-end-of-town problem. They aren’t. The short answer is that the value you get out of AI now depends far more on the context you can hand it than on which model you pick, and gathering that context is plain, unglamorous work a team of any size can start to do today instead of “soon”.
Why the model matters less than it used to
In July, Microsoft CEO Satya Nadella published an essay he called the Reverse Information Paradox. It builds on economist Kenneth Arrow’s Information Paradox, but shifts the focus from people selling information to organisations using AI. His argument is that you pay for AI twice. In his words, “You essentially pay for intelligence twice, once with money, and again with something even more valuable, the proprietary knowledge you must reveal to make that intelligence useful.” Models learn from your prompts, from the tools your agents use, and above all from the corrections you make when an answer comes back wrong. Each correction, Nadella writes, gets “distilled into institutional know-how”. So his case is that a company should hold on to its own organisational memory, its feedback and decisions, its institutional context, and the outputs of its own AI work. He closes with a line worth keeping, “In consuming intelligence, you are creating intelligence. And what you create should belong to you.”
Nikesh Arora, the CEO of Palo Alto Networks, made a related point on several podcasts in the last couple of weeks. General models trained on public information are fine for general work, but as he puts it, “Mission-critical enterprise applications need precision models trained on high-quality, industry-specific data.” He also named something we run into constantly here, that “preparing enterprise data infrastructure for deployment of security systems is mostly identical to what’s needed to deploy AI systems.” Same groundwork either way, and for many businesses it is still the hurdle they haven’t cleared, or even started to prepare for.
Our own read is that this is why the argument has moved. Pushing a frontier model forward keeps getting harder and more expensive, while the good open models close the gap behind them. When every business can buy roughly the same capability, what separates two of them is what each can tell that capability about itself.
What this looks like in a small business
Think of it in terms of a small insurance brokerage with six brokers. Their most senior broker has 20 years of client conversations, decisions, and insurance wisdom in her head. Give her an odd risk and she knows which insurer will take it, what wording to use, and what the underwriter will ask. New brokers take years to get near that, and when someone leaves, most of it goes out the door with them. The brokerage tried a general AI assistant on their client email and got back polite, generic drafts no experienced broker would send.
So as a team they changed one small habit. Every time a policy got placed, the broker added two lines to the file about why that insurer, why that wording, and what the client’s real constraint was. Same for renewals, and for the reasons a claim got pushed back. It cost about a minute for each job, and it turned work they were already doing into a record of how they make decisions. After three months they had around a thousand short notes explaining real decisions. Pointed at that, the same AI could draft a renewal recommendation that sounded like their firm and set aside the odd ones for the senior broker. Their newest broker reached a usable first draft in about half the time.
The part that surprised them was the corrections. When the senior broker rewrote a draft, they kept the rewrite and the reason for it, and that file became the most valuable thing in the business. It is exactly the asset Nadella argues you should be keeping for yourself. Getting material like that into a form a model can use is what retrieval-augmented generation does, and it is well-worn plumbing rather than anything exotic.
Where this goes wrong
The instruction people take from all this is “record everything”. Every meeting, every email, every phonecall, even down to every conversation in the hallway. In some cases, we would push back on that. Record absolutley everything and you end up with a swamp nobody can search and a privacy problem you never needed. What is worth capturing is often far smaller. It is the reasoning behind decisions, which almost never gets written down and is the one thing a model cannot guess. For a number of clients, we have added a specialised agent that summarises every meeting recording knowing what to look for based on the type of meeting and the context needs of the business.
There is a compliance side to it too. If you are recording conversations and storing client information, the Australian Privacy Principles under the Privacy Act 1988 govern how you collect, hold, and use personal information, and consent to record is not a formality you can wave through. Settle that before you start collecting rather than after.
Then there is what happens when AI stops drafting and starts doing. Arora is blunt about it, “When you deploy it for a precision-use case and give it arms and legs, it doesn’t matter what guardrails the model comes with, you will have to superimpose better guardrails and controls around it.” Good context makes an agent more capable, and a capable agent working from poor context is just wrong faster. Sorting out what you actually know, and how much you trust it, comes first.
And the honest limit. None of this replaces the senior broker. It captures how she decides so the firm keeps the benefit, but she still signs off anything unusual, and the system only learns from reasoning somebody wrote down. Whatever she never explains stays exactly where it is.
What this means for you
You don’t need a data project to begin. Pick the single decision your business makes most often, and start recording the reason behind it rather than only the outcome. Two lines in the file. Do that for a quarter and you will have something no vendor can sell you and no competitor can copy, because it is a record of how your business thinks. That is the point Nadella and Arora circle from different directions, and it holds for a six-person brokerage as firmly as for a listed company.
If you want to see how we approach that groundwork, our data insights and analysis work is about getting business knowledge into a state AI can use, including for insurance teams. Have a look and see whether it lines up with what you are already sitting on.
