When AI Has an Information Budget: The Next Digital Divide in Education
AI agents are learning how to pay for information. Schools need to start asking what happens when their agents cannot.
I watched a video this week that was supposed to make me think about business ideas.
It worked. At first, I was following right along with the intended exercise. Then I started connecting the ideas to education, and somewhere along the way the opportunity started to feel like a warning.
The video, “Making $$$ selling to AI Agents”, is an episode of Greg Isenberg’s Startup Ideas Podcast. It begins with a recent Cloudflare announcement and works outward from there. If AI agents can carry wallets and pay for things on our behalf, there will be businesses built around selling them what they need.
What do agents need? Among other things, they need information.
The episode offers three practical opportunities: refine messy information in a niche into clean data that agents can purchase; help businesses make their websites and information “agent-ready”; or turn an expert’s archive into a tool that agents pay to consult. The larger message is not to predict the entire future of the internet. Just build a useful paid door.
Honestly, that is good startup advice.
It was the doors that I could not stop thinking about.
The moment a web request becomes a purchase
Cloudflare’s announcement sounds technical, but the basic idea is easy to understand. Its proposed Wallets would allow a person or organization to fund an account and give AI agents smaller virtual wallets. Those wallets could have spending limits, approved sellers, and maximum transaction amounts.
When an agent encounters a resource that costs money, it could pay and continue working without sending a human through a checkout page.
This is part of a larger system Cloudflare has been assembling. In 2025, it introduced Pay Per Crawl, allowing a website owner to let an AI crawler in, block it, or charge it. In July 2026, Cloudflare announced a Monetization Gateway intended to let owners charge for almost any digital resource behind Cloudflare, including a webpage, dataset, API, or AI tool. Cloudflare is also experimenting with moving from pay per crawl to pay per use, so compensation can occur when content contributes to a result rather than simply when a crawler collects it.
Some of these products are still in beta, on waitlists, or described as coming soon. The important part is the direction. Cloudflare calls it an “agent-first Internet,” where an agent can request a resource, learn the price, pay, and receive it in one interaction.
That is an elegant solution to a real problem. The people and organizations creating useful information deserve control over it. Journalists, researchers, publishers, museums, professional associations, independent experts, and online communities should not have to provide free raw material to some of the wealthiest companies in the world.
But once an agent can pay for information, it can also decide not to.
That decision is where this becomes an education issue.
Didn’t we go through this with search engines?
My first thought was that we have seen some version of this before.
Search engines crawled the internet too. Website owners generally allowed it because there was an exchange: search engines indexed their content and sent people back to their websites. Publishers could turn those visits into advertising revenue, subscriptions, donations, or sales. Search engines funded their own enormous indexing operations primarily through advertising.
It was an uneasy bargain, particularly for news publishers. Europe and Australia eventually created mechanisms intended to make large platforms negotiate or pay for certain uses of news content. Google now has licensing agreements with publishers while maintaining that links and very short extracts can remain free. The argument over who creates value and who captures it never completely went away.
AI changes the exchange.
An AI system can retrieve information, synthesize it, and give the user a finished response. The user may never visit the original site. Cloudflare has reported enormous differences between the number of pages crawled by AI companies and the referrals they returned to publishers. That helps explain why content owners are looking for compensation at the point of access rather than hoping for a click later.
Search monetized discovery. AI is moving toward monetizing the answer itself.
For schools, that distinction matters.
The divide after the AI divide
Education has spent decades confronting versions of the digital divide. First, we worried about access to computers. Then broadband. More recently, we have begun asking which students and teachers have access to capable AI tools and which do not.
The emerging infrastructure adds another layer:
There may be people without AI, people with AI, and people whose AI can afford to access the information it needs.
We could call this an information budget.
The phrase is useful because it shifts our attention away from the model alone. We tend to judge an AI product by which model it uses, how well it reasons, or how many messages a user receives. In an agentic system, capability may also depend on what the agent is allowed to purchase along the way.
Imagine two students asking the same research question through the same friendly AI interface.
One student’s agent can consult current journalism, a specialist database, and a primary-source archive. The other student’s agent has no external spending authority. It works with freely available pages, older model knowledge, snippets, cached material, or sources already licensed by the vendor.
Both students receive an answer.
That may be the most troubling part. The next information divide could be invisible because everyone still receives an answer.
The student may never see the doors the agent encountered. There may be no message explaining that a relevant source cost twelve cents, that the system declined to pay, and that the response was assembled without it. A polished answer can easily hide a constrained evidence pool.
This will matter most for the information people often value most: information that is current, specialized, local, or difficult to compile. A model may retain older information from its training, but elections change, laws change, scientific findings develop, public-health guidance evolves, and education policy moves. Current questions require current access.
At some point, “better access” stops meaning a marginal improvement in quality. If the necessary current or specialist source is unavailable, it can mean the difference between accessing the information and not accessing it at all.
Wikipedia shows both the promise and the problem
Wikipedia was one of the first examples that came to mind. It is difficult to imagine modern AI without the enormous body of human-curated knowledge Wikipedia has made freely available.
There is a compelling argument that large AI companies should help fund it.
In fact, Wikimedia is already pursuing a thoughtful version of this model. Wikimedia Enterprise provides high-volume commercial users with structured, reliable access to Wikimedia content. Companies including Google, Microsoft, Meta, Amazon, Perplexity, and Mistral have reportedly entered commercial arrangements. Wikipedia remains free for humans, while companies using its knowledge at scale pay for services designed around their needs. Wikimedia also says it offers free Enterprise access to some mission-aligned nonprofits and startups.
That seems far more sustainable than asking individual donors and volunteers to subsidize commercial AI indefinitely.
It also shows the kind of public-interest distinction education may need. Commercial model training, public school research, and a student asking a homework question are not the same use, even when software makes the request in all three cases.
Wikipedia’s mission makes it inclined to preserve free knowledge. Not every owner of valuable information will have that mission. If more archives, databases, publishers, and experts build useful paid doors, who ensures that public education still receives a key?
If the student cannot pay, perhaps an advertiser will
This was the connection that made me stop for a moment.
Search engines found a way to offer an extraordinarily expensive service without charging most users directly: advertising.
Why wouldn’t AI companies follow the same path?
OpenAI has already announced plans to test advertising for adults using its Free and Go tiers in the United States. The company says ads will be labeled, separated from answers, and unable to influence the organic response. Google already places advertising in some of its AI-powered search experiences. Perplexity experimented with sponsored follow-up questions before reportedly retreating from advertising because of concerns about user trust.
These experiments are not evidence that an AI company currently makes students watch an advertisement to unlock a paid source. That specific scenario is still speculative.
It is no longer difficult to imagine, though.
An AI provider facing both inference costs and information-access charges has several choices. It can absorb the expense, raise subscription prices, restrict paid sources to premium tiers, or subsidize free access through advertising and commercial partnerships.
For education, that could produce a deeply uncomfortable set of options. A well-funded district might provide an ad-free agent with access to licensed sources. A student using a free consumer tool at home might receive advertising alongside the answer. A child account with stricter advertising rules might receive a more limited answer because neither the family nor an advertiser is paying for additional access.
The concern becomes even sharper if advertising moves beyond supporting the service as a whole and begins subsidizing particular information pathways. What happens when a bank sponsors deeper access to financial-literacy resources, a food company sponsors expanded answers about nutrition, or an education vendor underwrites access to research in its product category?
The sources might still be legitimate. The advertisement might still be labeled. Yet commercial incentives would be helping determine which questions receive richer treatment and which information doors open easily.
Students without an information budget may not be denied an answer. They may be asked to pay with attention, exposure to commercial influence, or personal data instead.
Questions schools should ask before this becomes normal
I do not think school leaders need to add “agent wallet policy” to tomorrow morning’s agenda. Much of this infrastructure is new, and some of the business models may never develop exactly as described.
I do think the questions belong in today’s conversations about AI procurement and governance.
When evaluating an AI product, schools should eventually be able to ask which external sources it can access, which sources cost extra, and whether the system discloses when price affected retrieval. Districts should know whether their existing subscriptions to journals, databases, news archives, and curriculum libraries can be used by an agent, or whether they will be charged again for machine access.
EdTech developers have another responsibility. If a system declines to retrieve a source because of price, hiding that decision behind a fluent answer is not sufficient transparency. Products may need to show which sources were considered, which were used, which were unavailable, and whether access rules or spending limits changed the result.
Policymakers, publishers, and open-knowledge advocates will need to consider whether schools, libraries, nonprofits, and other public-interest users deserve different licensing structures from commercial AI companies. Otherwise, rules created to make large technology companies compensate creators could unintentionally push the cost downstream to institutions least able to carry it.
And all of us should become more careful about treating an AI response as if it represents everything the system could find. Soon, it may represent only what the system could afford.
Paid doors are not the villain
I want to be careful here because the easy version of this argument would also be the wrong one.
Paid information is not inherently bad. Advertising does not automatically corrupt an answer. AI companies are not secretly coordinating to keep knowledge away from students. Content owners asking to be compensated are responding rationally to an economic model that has often extracted their work without returning enough value.
The video did not propose an educational dystopia. It identified a real entrepreneurial opening and encouraged builders to act before the entire market is settled.
My discomfort comes from what happens when many reasonable decisions accumulate.
One creator builds a paid archive. One publisher charges per use. One AI provider limits external spending on its free tier. One district chooses the lower-cost plan. One advertiser offers to subsidize access.
No single decision creates the divide.
Together, they could create an internet where knowledge remains technically available but is reached through agents with very different budgets, permissions, and commercial obligations.
We should find ways to compensate the people and organizations producing trustworthy information. We should also protect students’ ability to reach that information regardless of the purchasing power attached to their AI.
Those goals are not incompatible, but they will not balance themselves.
The original digital divide was visible. A student had a computer or did not. A household had internet access or did not.
This one may be harder to see. Every student may have an AI. Every AI may produce an answer. Only some of those agents may be able to open the doors behind it.
Before the agent-first internet becomes ordinary, education needs to ask a question that the market will not answer for us:
If current knowledge becomes something an AI must purchase, what will students without an information budget be asked to give in return?
This article was developed with the assistance of ChatGPT 5.6. The argument, perspective, and final editorial decisions are my own.
