The Agentic Information Economy
An internet populated and run by AI agents will need not just reliable data but also a broader information ecology capable of questioning its own inputs and transmitting corrections where they are needed if information changes. But whether this will serve private interests or the public good remains to be seen.
Shuwei Fang
CAMBRIDGE—Information is undergoing a category change.
It has always been something people consume, but it is also becoming something machines act on.
That means errors will stop being mere digital impressions and start becoming actions with far-reaching consequences.
Given the risks, the system will need safeguards of its own.
Better data or putting a human back in the loop will not suffice.
For as long as the consumer internet has existed, information and autonomous action have mostly been separated by a human interlocutor.
A shopper compares, then buys.
A procurement manager considers various bids, then picks.
The web was mostly built as an information system for human decision-making, and everything about it—the dominant business models, the quality-control mechanisms, the remedies when something goes wrong—assumes that there is ultimately someone at the end of the chain to make sense of things.
It is that person who is being removed.
AI agents are already screening job applicants and making autonomous purchases within a predefined budget.
In software, “dark factories” write and ship code with almost no human review.
When economists at Microsoft Research set out a fuller version of this agentic economy, every human consumer was represented by a shopping agent, and every business by a service agent.
Put those together, and you have a chain rather than a tool.
Your agent would ask a supplier’s agent about the price of a good, and the supplier’s agent would check with a logistics agent that is working from a forecast that no person has read.
Each one’s output becomes the next one’s input.
Although this economy does not exist yet, its constituent parts do.
As they join up, information will stop being what the system delivers, because it will have become what the system runs on.
What it delivers will be actions previously done by humans: buying, analyzing, shipping, deciding.
Whatever the agents are doing, this human-free system’s performance will rest on the quality of the information going in.
The old adage about “garbage in, garbage out” applies more than ever.
This framing is dangerously incomplete, however.
It casts the problem as one of cleaner inputs, of better data, when in fact agents acting alone in a vast economy need something larger: an information ecosystem capable of questioning its own inputs, carrying corrections back to where they are needed, and aiming to be consistently right, rather than merely accurate on average.
The State of Play
We have the parts of this system, but none of the self-correcting reflexes.
To understand what is missing, consider what an agent can actually draw on today.
There is the open web, which was written for human readers.
There are enterprise streams—the ledgers, contracts, and case files inside organizations—which were built to be queried only by their owners.
There are government registers and court filings, which are authoritative and public, but rarely machine-readable.
And there is personal context: whatever an agent has been told about the person it is acting for.
Each of these carries different rules of access, and whatever signals of reliability they do carry were made for a human reader and will stay at the source once the information has been absorbed.
Moreover, the richest of these sources are precisely the ones that nobody should be able to harvest freely, because they belong to the firms and the people they describe.
Nor do current sources automatically update once information has passed down from one agent to another.
Suppose that a contamination result comes back on a Tuesday, and a batch of infant formula is recalled.
By the time the recall is published that afternoon, a replenishment agent has already placed a new order, a logistics agent has already routed the pallets, and by Wednesday morning the tins are on the shelves.
Each one acted correctly on whatever information it had.
Nothing did anything wrong.
The problem is that a recall designed to be noticed by people cannot keep pace with the speed and scale of automation.
For now, such failures remain limited, because most tasks are still simple and self-contained, such as a checkout agent buying within a set budget with a human at either side.
But the trouble grows as tasks become more complex and pass through many more hands.
Then there is a second, much older problem.
Human information was written to be interpreted, whereas machine information has to be executed on.
A book blurb written to sell, a corporate filing written to reassure, a politician’s pledge written to be applauded—a human reader discounts each of these and mostly does not act on them.
That discounting is an act of self-correction, one that is constantly distributed across millions of readers in fuzzy, continuous, and free process.
It fails often, and sometimes catastrophically, but it does so at human speed and scale.
And because it cost nothing to build, nobody ever thought of it as infrastructure, and nobody has priced its replacement.
Putting the person back in is beside the point.
The inputs were never sufficient on their own, because they were only half of a system whose other half was a person.
If that half is going, then some capacity for doubt and correction must be designed into whatever replaces it—and someone must have a reason to do the designing.
Nor will better models solve this.
A correction that never arrives is a failure of plumbing, not of cognition.
Where human discounting was the work of millions of readers with different priors and different stakes, agents currently converge on a handful of models: the judgment may be better on average, but it is the same judgment everywhere, so its errors correlate rather than cancel out.
Capability determines how well an agent reasons, not what it has to reason from, and because capability is what licenses autonomy, it increases the exposure to risk rather than reducing it.
The cost of all this is borne at the point of reversal.
Mistakes can still be undone, but by the time one is discovered it may have been acted on many times over.
The error travels at machine speed and scale, but the correction currently does not.
Under these conditions, the harm, when it comes, will not be a misleading headline that duped a few readers, but a mis-executed action.
Nor will the exposure be confined to private firms and individuals.
Early glimpses of automation, long before agents, give us some idea of what to prepare for.
Between 2016 and 2019, Australia’s Robodebt scheme automatically told hundreds of thousands of welfare recipients that they had been overpaid and demanded money many of them did not owe.
A tribunal found against the method in more than 400 cases, but its decisions set no precedent and were not published, so nothing carried them back to the system that continued raising the debts.
The Price of Being Right
Faced with such risks, many will argue that we shouldn’t hand over our lives to agents in the first place.
But the agentic economy is coming regardless, driven by a powerful commercial logic
The real question, then, is how to seize the opportunity it presents.
We should see the machines’ demand for accurate information as a form of leverage, because they will need it even more than we do, given the scale of the mistakes they could make.
In financial data, clinical evidence, and legal research, accuracy has always been worth paying for, because it could be certified and the stakes justified the cost.
Everywhere else, being right and sounding right have fetched the same price.
But if automation is happening across the economy, that is what could change.
Two factors make it possible to price high-quality information.
First, information that was previously only consumed by people as a final good is becoming an intermediate input for machines.
Trusted consumer reviews, price histories, and safety recalls are no longer units of text to be read but signals to be acted on.
And if any of this information changes, the correction signal becomes another valuable input.
All these inputs can be valued against their contribution to the output (reflecting what economists call derived demand), and unlike a human reader, an agent’s output can, in principle, be inspected.
That brings us to the second factor, which is that an agent’s contribution is measurable.
The difference a given input makes to the result is an observable quantity, and so is the difference made by correcting one that was wrong.
In practice, of course, demand alone does not make a market, much less an ecosystem.
For that, three standards would need to be applied to inputs across any multi-agent economy.
First, an input would have to say where it came from, and stay referenceable when that source moves or disappears.
Second, a correction would have to reach everywhere that the input was used, rather than sitting at the source waiting to be noticed.
And third, a source would have to be able to answer a question without surrendering what lies behind it.
Finance has long had mechanisms for information provenance and self-correction, because it is an industry where information is expensive to get wrong.
When a company revises its earnings, the correction travels to most of the systems that priced on the old figure, and a record of what was known on any past date is sold as a product in its own right.
Where these standards appear next will depend on the same calculation.
What anyone will pay for a good input tracks the cost of a bad one, which leaves out the places where errors are cheap, or where the person a decision is about may not be the person directly paying: the benefits claimant, the visa applicant, the patient on a waiting list.
Fortunately, this gap is not as large as it sounds, because the same mechanism could serve both worlds.
Knowledge about an input’s origin does not need to be created twice just because a case worker rather than a trading desk is asking for it.
The same is true of corrections: once an input carries information about its source, telling everyone who used it that something has changed is an incremental step.
Railways built to shift coal ended up carrying passengers, post, and milk.
Extending something that has already been built costs less than building something new.
There is an obvious way to close that gap, and it is already in public hands.
Governments are among the largest buyers of software anywhere, and they are now deploying agents into precisely the processes where the person affected has no purchasing power: benefits, permitting, immigration, health.
A procurement rule requiring that inputs carry their provenance, and that corrections reach the systems that used them, would price the thing into existence for the claimant and the applicant as well as the trading desk.
The state that produced Robodebt has more reason than most to write it.
The cost is not negligible, of course.
The work of keeping information accurate never stops, and what replaces universal human interpreters will not be free.
Anyone can build the pieces we need—be it an individual, a company, a foundation, an open-source project—but sharing them is unlikely to happen by itself.
Building provenance and correction structures inside one sector or company is a manageable problem, because firms will have a reason to do so.
But making these work across a broader ecosystem to deliver societal benefits is a different problem, because, as with all public goods, it belongs to nobody in particular.
If higher-level coordination is going to happen, it will fall to standards bodies, large buyers, and forward-thinking regulators.
Absent that, the market will form privately, centered around a few firms with their own proprietary ways of establishing what to trust.
Scrutiny would run in only one direction: a firm would know and decide what information its own agent relied on, but a person affected by the decision would not.
This scenario would be not just unfair but unstable.
Several dominant private players might give the impression of diversity, but nobody would know whether three sources are really one source wearing three coats.
Moreover, private enclosure would have a high running cost, because if every firm trusts only its own sources, information from outside would be expensive to use.
An agent that can only act on what its owner vetted can only be trusted inside the building, which is why the civic and commercial cases point the same way.
Whatever replaces the traditional reader will be built one way or another.
But whether it works across the whole ecosystem or only inside a few firms is being decided at this very moment, while the standards are still soft.
Shuwei Fang is a fellow at Harvard Kennedy School’s Shorenstein Center on Media, Politics, and Public Policy.
0 comments:
Publicar un comentario