Moving Agency to the Edge
Personal AI, fiduciary duty, and the difference between being authorised and being recognised.
Digital governance has largely been built around two responses to concentrated technological power. The first is regulation: when platforms collect too much data, manipulate choice, discriminate, or exercise excessive control over access to markets and information, public institutions attempt to define limits, impose duties and create routes of redress. The second is corporate responsibility: companies are asked to adopt better standards, improve transparency, design more responsibly and align their systems more closely with the interests of the people who use them.
Both responses are necessary, and neither changes the underlying direction of the relationship. The platform still receives the data, performs the computation, constructs the representation of the user — the score, the classification, the eligibility model, the inferred profile — and sets the terms of the interaction. Regulation constrains what may happen after that asymmetry has been created; corporate responsibility asks the actor benefiting from it to exercise it well.
A different approach begins by reversing part of the equation. What if more computation took place on the side of the individual, before information crossed into somebody else’s system?
A personal computational layer could evaluate requests, interpret terms, verify credentials, filter information and disclose only what a particular interaction requires. Instead of exposing a complete identity to establish one relevant attribute, a person could supply proof of that attribute alone. Instead of accepting a complex set of permissions in order to reach a service, software acting locally could assess them before the transaction occurred. The point is not to remove institutions from digital interactions, but to change what reaches them, and under whose control.
Europe’s developing digital identity architecture already contains part of this logic. Regulation (EU) 2024/1183 requires Member States to make European Digital Identity Wallets available by the end of 2026, built around interoperable credentials and user-controlled disclosure. The Commission’s age-verification work shows the pattern at its clearest: a person can prove they exceed a required threshold without revealing their exact age, their identity or anything else. The service receives the information it is entitled to use without receiving the underlying data from which that information was derived.
This is a meaningful shift, and it answers only one part of the governance problem. It establishes greater control over what leaves the individual’s boundary. It says nothing about whose interests the software operating at that boundary serves.
From edge computing to fiduciary agency
A personal AI can read a contract faster than its user, assess a request for information, compare financial products, identify a problematic clause, or recognise that a service is asking for more than it needs. But technical capability does not create loyalty. An agent running on a person’s device could still be designed around the commercial interests of whoever supplied it: its recommendations favouring affiliated products, its interpretation of consent serving an advertising model, its apparent role as the user’s representative masking incentives established elsewhere in the value chain.
The question therefore moves from where the computation happens to the relationship governing it. Who does the agent work for?
This is not a new question, and it is not being asked only here. The GliaNet Alliance has been developing a closely related answer through the concept of the Net Fiduciary. Under GliaNet’s model the fiduciary is not the AI itself but a company that adopts a fiduciary posture and corresponding commitments toward the people it serves — duties of care, fidelity and loyalty, embedded in governance structures, technical systems and daily operations, with conflicts of interest resolved in the client’s favour. A personal AI is a technical capability. A fiduciary relationship establishes who is answerable for how that capability is used: the first can filter information, the second creates a duty.
This adds a missing institutional layer to the edge-computing proposition. It is not enough for software to operate close to the individual. Someone must be legally and organisationally responsible for ensuring that the software does not use its position against the person it represents.
This raises a question that belongs to fiduciary duty as a category, not to any one attempt at it: when a fiduciary is a conventional, investor-owned company, what happens when the user’s interest and the firm’s commercial interest genuinely diverge? A pledge can state that conflicts resolve in the client’s favour, but a pledge and its performance under pressure are two different things, and no amount of drafting settles that gap in advance. The same tension sits inside every professional fiduciary relationship, including the oldest ones: a lawyer’s duty of loyalty does not depend on the firm having no commercial interest of its own.
There is a second way to build this kind of carrier, and it answers the shareholder question differently rather than regulating around it. A data cooperative is owned and governed by the people whose interests it represents, rather than by outside investors, so there is no separate shareholder interest for a client’s interest to compete against in the first place. A company adopting fiduciary duties keeps the investor-owned form and places a duty on top of it; a cooperative changes who owns the carrier instead. Neither guarantees enforceability on its own — a cooperative’s fiduciary standing still depends on the legal regime it operates under, just as a company’s pledge depends on whether anything obliges it to keep it — but they are two different answers to the same question: whether the entity representing someone’s interests has to share its loyalty with a shareholder at all.
Ownership is not the only axis on which the two models diverge, and the second one is regulatory rather than structural. The EU’s Data Governance Act (Regulation (EU) 2022/868) requires anyone providing data intermediation services to remain neutral: acting only as an intermediary in a transaction, using the data for no other purpose, structurally separated from any other service the same provider offers, with pricing independent of anything else a client might buy from it. Neutral here does not mean disinterested. An intermediary is allowed, even expected, to negotiate hard for the party it represents. What the rule forbids is a separate commercial interest — using the data for something else, extracting value from it beyond the agreed transaction, pricing the service so it depends on some other product sold alongside it.
A member-owned cooperative satisfies that test by construction. There is no cooperative interest apart from the members’ interest for the rule to catch, because the members are the owners. A private company offering the same narrow service would have to build that separation deliberately — a product walled off from anything else it sells, priced independently, with no other use for the data passing through it — rather than get it for free from its own structure. Ownership and neutrality meet here more directly than they first appear to: cooperative ownership is not incidental to satisfying this test, it is one of the more direct ways of satisfying it.
But this only covers the function the Act was written for. It names data cooperatives explicitly, as a form of intermediation service that helps individuals and small businesses negotiate the terms on which their data is shared — negotiating terms, not advising on a contract, not contesting an insurance decision. That narrower function is what neutrality was built to accommodate, and it is why a cooperative has a regulatory home here that a fiduciary company does not.
What neither model has, cooperative or corporate, is a clear path for the broader work this essay has been describing. Advising, comparing, contesting a classification on someone’s behalf is not data intermediation as the Act defines it, and no ownership structure moves that boundary. Cooperative ownership helps with the question the Act actually asks. It does not enlarge the question the Act is willing to ask in the first place.
Any such commitment begins voluntary, because there is no other way for it to begin. A company can adopt these duties toward the people it serves well before there exists a professional body able to certify or monitor them — the kind of self-regulatory structure medicine and law eventually built — because that structure only becomes possible once enough relationships of this sort exist to regulate. But even a fully mature version of that regime would meet a boundary that has nothing to do with how well it is built: A profession can bind its members. It cannot by itself oblige a third party to recognise what those members do on a client’s behalf.
That is not a shortfall in any one design. It is true of fiduciary duty as such — a lawyer’s mandate meets exactly the same boundary the moment it reaches a court, a bank or a public agency that owes the lawyer nothing. The question this raises is not whether the fiduciary can be trusted. It is whether anything outside the fiduciary relationship is obliged to listen to it.
That is where personal AI stops being a matter of loyalty and becomes a matter of standing — a distinction worth making precise, because it is where the architecture actually breaks.
Authorisation is not standing
Picture one relationship with an insurer, followed across three ordinary moments — a composite, not a case, but the kind of sequence any policyholder could have.
An application for a policy is submitted online, and the intake form asks whether the applicant is over eighteen. A credential in the applicant’s wallet establishes the fact, and the personal AI discloses nothing else — no birth date, no ID scan, no address history. It is acting as a filter, and the underlying mechanism already exists: the service gets the one bit it needs and nothing more.
A year later the policy comes up for renewal with a clause buried in the terms — an arbitration requirement that would be costly for the policyholder to invoke and easy for the insurer to. The personal AI reads it in seconds, flags it, and suggests a competing policy without the clause. It is acting as an adviser, and this works too: better information, surfaced faster, with the policyholder’s interests rather than the insurer’s shaping what gets flagged.
Then a claim is filed, and it is denied. The decision rests on a risk model that classifies the household as high-risk based on a proxy — a postcode, a payment pattern — that does not reflect its actual circumstances. The personal AI concludes the classification is wrong and tries to contest it on the claimant’s behalf.
This is where the sequence stops working, and it is worth noticing exactly why. Nothing about the AI’s loyalty has changed. It is exactly as loyal as it was when it withheld the birth date at the first step. What has changed is that the AI is no longer disclosing or comparing — it is attempting to act inside the insurer’s own institution, and loyalty to the policyholder is not the same as having a role that institution recognises.
An actor can be authorised to speak or act on someone else’s behalf. That authorisation does not automatically give the representative standing inside another institution.
A mandate answers one question: may this actor act for me? Standing answers another: must the institution recognise that representation as capable of triggering an institutional response?
Standing is used here in the broader institutional sense: recognised status that allows a represented claim to count within a decision process, not the narrower legal sense of standing to bring a claim before a court.
The difference already exists in analogue and digital systems alike. A lawyer may hold a valid mandate, but the procedures of the court determine how that representation is recognised. A financial intermediary may be authorised by its customer, while regulation and contract determine what a bank or data holder must accept from it. India’s Account Aggregator framework separates these elements deliberately: the customer’s consent, the regulated intermediary carrying it, the structured consent artefact defining its scope, and the obligations of participating financial institutions.
The institutional consequence does not arise from the software alone. It arises because the surrounding framework recognises the mandate within a defined scope — this consent artefact, for this purpose, nothing wider. This is also why the Account Aggregator is a regulated intermediary rather than a marketplace of individual agents: standing of this kind is close to impossible to build one relationship at a time. It came from regulation obliging banks to recognise a defined class of consent artefacts, not from any single customer’s authorisation, however genuine.
Get Gael Van Weyenbergh’s stories in your inbox
Join Medium for free to get updates from this writer.
The claimant’s AI has a mandate. What it lacks, at the point the claim is denied, is a recognised role for that mandate inside the insurer’s own procedures. Personal AI makes this distinction increasingly consequential, because an agent can be entirely loyal to its user and still reach a boundary at which that loyalty has no procedural consequence.
Moving computation toward the person redistributes capability. Moving agency toward the person requires something more: institutions must recognise what that capability is authorised to do.
The interface is not the agent
Approached from the institutional side, the problem looks different again.
Consequential decisions are increasingly assembled across several organisations. A public agency may hold formal authority, a vendor may construct the model, another organisation may supply the data, a regulator may provide recourse, and a procurement team may control whether the system can be changed at all. Each actor can operate within a legitimate and bounded mandate while the consequence produced by their combination falls between those mandates.
The structural problem is not the absence of trust. It is that consequences cross more boundaries than responsibility does.
From this perspective the personal AI should not be treated as the institutional interface. It is better understood as one possible user-side instrument arriving at that interface — a distinction worth keeping, because several functions otherwise collapse into a single technical object.
The individual is the principal whose interests are at stake. The fiduciary is the legal person owing duties to that principal. The personal AI may be the technical agent through which part of the mandate is exercised. The mandate defines what that agent or fiduciary is authorised to do. And the receiving institution determines, under law or governance arrangements, whether the mandate has standing within its own procedures. These remain separable relationships even when future services package several of them together, and keeping them separate is what allows responsibility to stay visible.
The separation is already familiar in other domains. An investment adviser can operate as a commercial company while owing client-facing fiduciary duties; the regime exists precisely because the company’s interests and the client’s are not assumed to be identical. Personal AI combines that fiduciary problem with two others already visible in the insurance sequence: bounded disclosure at the edge, and institutional learning after a consequential decision.
Control over information, fiduciary representation and institutional learning are three different governance problems. Treating personal AI as the answer to all three hides the differences that determine whether agency is actually effective.
From recourse to learning
Say the claimant eventually gets somewhere: the insurer reviews the case, agrees the classification was wrong, and reverses the decision. Recognition worked, and the remedy may be adequate for this one claim.
But the postcode-based proxy that caused the misclassification is still sitting inside the model, waiting for the next household it will misclassify the same way. If the case closes without that finding travelling anywhere, the institution has corrected one decision and changed nothing else, and the same error will recur for someone else next month.
Recognition is what let the claim reach the institution at all. What happens after that is a separate question: does what the case revealed have anywhere to go?
That is what a governed return channel addresses — an institutional pathway through which knowledge arising among the people affected can reach the relevant decision structure, receive a response, and contribute to future adaptation. Where AI mediates a service, such a channel has to connect two levels of correction at once: the individual case, and the system producing repeated cases.
A remedy completes a case. Learning changes the system that will produce the next one.
A personal AI can therefore improve individual agency without improving institutional responsiveness at all. The two become complementary only when representation can cross the boundary, reach authorised judgement in time to matter, and — where the same consequence recurs — travel far enough upstream to alter its source.
When the fiduciary cannot fix the problem
The difficulty deepens when the consequence does not belong to a single institution.
An insurer may make a decision while relying on a model supplied by a vendor, which in turn relies on external data. The insurer may remain legally accountable while lacking direct control over some of the system’s representational choices. A regulator may be able to require remediation without possessing the operational capacity to redesign the model.
A fiduciary acting for the individual carries that individual’s interests into the arrangement. It cannot acquire the vendor’s mandate, or the regulator’s, or force the three of them to act as one.
Two obvious responses fail, and for the same reason. Giving the individual’s fiduciary authority over the vendor’s model would not remove the imbalance — it would just hand it to a different institution, one the vendor and the regulator would then have to answer to instead. Merging the insurer, the vendor and the regulator into a single body would remove the seam, but also the reasons those functions were separated in the first place: the vendor’s technical expertise, the regulator’s independence from the companies it oversees, the insurer’s accountability for the decision it actually made. Collapsing the boundary destroys what the boundary was for.
What is needed instead is a logic for keeping these independently governed functions connected without requiring any one of them to run the others. The framework calls this Convergence:
Convergence is the coordination logic through which independently governed functions become responsive to one another without requiring a single centre.
The aim is not to eliminate institutional boundaries. It is to govern what crosses them.
From control at the edge to agency across the system
The original case for computing at the edge remains compelling. Moving computation toward the user can reduce unnecessary disclosure, make digital interactions more intelligible, and place a greater share of decision-making capacity on the side of the person rather than the platform. Europe’s emerging identity infrastructure makes some of this technically and institutionally plausible, and GliaNet adds an equally important proposition: a personal AI that genuinely acts for someone requires an accountable relationship behind it, not merely a convenient interface.
The combination also reveals what comes next. Technical control over the edge does not itself create loyalty; fiduciary loyalty does not itself create standing; standing does not itself guarantee that judgement can be reached in time; individual recourse does not itself produce institutional learning; and no bilateral relationship can govern a consequence produced across several independent institutions.
The architecture therefore extends well beyond the agent:
edge computing → fiduciary agency → recognised standing → reachable judgement → institutional learning → convergence
These are not stages of a single technology. They are distinct institutional capacities that need to remain connected, and each can be present while the next is missing.
The deeper promise of personal AI may lie less in creating a more capable assistant than in changing the position from which people enter digital systems. Today the individual typically arrives as a subject to be authenticated, classified, profiled and acted upon. A different architecture would let a person arrive with computational capacity, selective credentials and representation of their own.
Whether that produces meaningful agency depends entirely on what exists on the other side. The objective is not to make the relationship seamless. It is to ensure that the seams between a person, their representative and the institutions acting upon them do not become the places where responsibility disappears.
→ The framework: Trust as Infrastructure
Written with AI assistance. Ideas and arguments by the author.
