Palo Alto's CEO says average AI intelligence will be free and context is the real advantage. What that means for an Australian small business.
Nikesh Arora, the CEO of Palo Alto Networks, a cybersecurity company worth more than $200 billion, said something this week worth sitting with, “In the long term, average intelligence is going to be free, and the average intelligence will get smarter. Exceptional intelligence will still be paid for, but that will be reserved for curing cancer and designing space data centres.” The premise being that for everyday work, it wont much matter which model is being used.
What Nikesh is interested in is context. In the discussion, he told the panel that the model doesn’t know why a customer’s infrastructure is down, what product they run, how it’s configured, or why the same fault appeared the last five times. None of that lives in the model. It lives inside the company and in the minds of his employees, so he now has more people collecting context than he’s ever had, feeding it into vector databases and GraphRAG Knowledge Bases that any model can sit on top of. Microsoft’s Satya Nadella argued something similar in June, writing that the real opportunity “is not in picking the best model but instead in building a learning and data loop on top of models”.
You probably run a much smaller business than either of them, but regardless, the same thing is true for you. If the head of the largest network security company in the world has stopped shopping for models and started collecting context, it shows that the question isn’t which AI to buy. It’s what you feed it.
Why a general answer sounds right
Ask a frontier model a general question and you’ll get a general answer that reads well. That’s the trap. The broader the question, the more a generalised answer seems correct, because there’s nothing in it specific enough to be wrong. Ask the same model something that depends on how your business actually works and it falls apart. We work with a commercial refrigeration business in Brisbane with 12 technicians servicing a few hundred supermarket and cafe sites. A cafe rings at 6am to say the coolroom is warm. The right first move depends on which unit is at that site, when it was last serviced, whether the same fault came up in autumn, and whether the part is on the van or three days out from the supplier. Their first AI approach was to paste the job details into an AI chat interface and they got a tidy list of things to check. Every item was sensible, and none of it was about the customer. Their best technician said he could have written it from the pub. The model wasn’t wrong. It just had nothing of the business in front of it.
The three kinds of context
Arora splits AI context into three parts. There’s the model, the raw intelligence. There’s the context needed to answer the question in front of you. And there’s the context needed to teach the system how your business answers questions at all. Most people (and sadly most AI consultants) only think about the second one. The third is the one that pays off over time. Palo Alto handles around 400,000 customer cases a year, and Arora tells his team that every one is a learning opportunity, not simply a job to close. He wants the reasoning pulled out of the heads of the people who solved it and written down as playbooks, so the next time it happens the system already knows. Nadella calls the same thing a learning loop, where your people’s judgement gets captured into something that improves with each use. You can offload a task, or even a job, he wrote, but you can never offload your learning.
In terms of the refrigeration business, that meant three unglamorous jobs and a data pipeline that keeps it all up to date. The asset register, so the system knows what’s installed where. The job history, so it knows what went wrong last time and who fixed it. The parts list, so it knows what’s already in each of the service vans. None of that is AI work, but all of it is the reason the AI became useful.
With the specifics in front of it, a smaller and cheaper model gave better answers than the frontier model had given with nothing. Good context lowers the level of model you need, which lowers what each job costs to run. That’s the payoff most businesses miss. Arora put it plainly, you don’t need to pay “$6 a million tokens to answer a call saying how can I help you”.
Context doesn’t arrive on its own
Here’s the honest part. Context doesn’t turn up because you bought a subscription. It arrives through infrastructure, integrations, and workflows. You need to create somewhere for the knowledge to live, connections into the systems that already hold it, and a habit or automation that keeps it current as the business changes. That’s AI integration work, and it’s the part nobody advertises. GraphRAG Knowledge Bases or Vector databases are common places to keep it, which is what Arora described his team building.
It’s slower than the pitch suggests, too. Arora thinks a large enterprise needs three to five years to work through every use case. A small business moves faster, but it still isn’t a weekend.
Two more things worth saying plainly. AI doesn’t fix bad data, it repeats it faster. If half your asset register is wrong, you’ve built something that gives confidently wrong answers at speed. The refrigeration service company spent three weeks tidying records before anything was connected.
Connecting systems also carries a cost people find out about late. In the same panel discussion, Jason Lemkin described turning on a Google Drive connector for an AI assistant. It scanned his documents, found a private ideas file he’d never shared, and the coding tool he had linked up then changed his code without telling him. Arora’s read was blunt: “It’s the wild west.” Small businesses are doing it with no idea whether their data trains someone’s model, which credentials get used, or what permissions those agents hold. If your business is covered by the Privacy Act 1988, the Australian Privacy Principles still govern how you collect, use, and disclose customer information, whatever tool is doing the handling.
The refrigeration business landed somewhere sensible. The AI drafts the first call on most morning jobs, and a person reads and checks each one before a van moves, because sending a tech ninety minutes out with the wrong part costs more than the AI saves. That check isn’t a limitation. It’s the design.
What this means for you
The argument coming out of the biggest AI buyers in the world is that the model is becoming the cheap part. What’s scarce is the record of how your business actually works, and you already own that. It’s sitting in inboxes, spreadsheets, and your best people’s heads, where no model can reach it. Nadella offers a good test: you should be able to swap the model underneath and keep everything the system has learned about your business. If changing models means starting again, the value was never really yours.
So pick one job that goes wrong the same way every month, and write down what a good answer to it depends on. That list is your context, and it’s the honest starting point. There’s more on how we approach AI integration on our website.
