Working research & field experiments · 2025–26 · two
studies
I started this research by looking at what changes when AI agents
begin acting directly on behalf of individuals. My initial question was
mostly one of execution: could an agent authenticate, pay and actually
complete an action for someone? But the field experiments gradually
pushed me towards something that may be more consequential and that
happens much earlier, when an agent researches a market and starts
eliminating candidates before any company is contacted. The broader
question became what happens to customer experience when part of the
customer journey, decision context and relationship memory sit in
software the enterprise does not control.
Customer experience has historically been organised around the
enterprise operating the customer journey. Personal agents introduce a
different possibility: an increasingly important part of that journey
may be operated instead from the customer side.
The timing is not incidental. When I began this work, I had to
construct the behaviour myself through experiments. Personal agents such
as Instinct and, more recently, Meta’s Muse make the premise less
hypothetical, as persistent software on the customer side moves from
constrained experiments towards consumer distribution. I do not know
which products will endure, but the assumption that a customer reaches a
company through an interface the company controls no longer holds by
default.
My working thesis is that personal agents are beginning to
operate the customer side of the relationship: they can
carry intent, context and memory across companies, while the enterprise
remains authoritative for commitments and execution.
The experiments
I ran two sets of experiments. In the first, I gave four AI agents a
real consumer decision each — telecom, electricity, motor insurance and
travel — with live web access and one instruction: research the market,
compare, recommend, and stop before contacting any company.
Across the four runs, 40 brands entered consideration and 31 were
eliminated. Twenty of those 31 were dropped without the
agent ever loading a readable page belonging to them.
Every source used by the four agents, separated by whether it belonged to a brand under evaluation. Filled circles materially influenced the decision; outline circles were consulted but did not; diamonds represent information drawn from model memory rather than from a webpage. Three of the six memory sources materially influenced the decision. These results describe the four experiments only.
The second study tested the memory clause directly. Across unrelated
consumer decisions, session continuity alone carried almost nothing
forward, while an agent with persistent notes carried materially more of
the customer’s preferences and changed later recommendations
accordingly. That matched what I had already observed in ordinary use
with Instinct, and the same design direction is now visible in Muse:
memory can shape what an agent surfaces, filters and recommends.
Sixteen agents were given the same final decision — a primary checking account — after four earlier, unrelated decisions. The brief stated no constraints at all. The four agents that had kept their own notes across the earlier decisions all recommended a local institution with branches and a phone line; the twelve without notes all recommended a national online-only bank. Nothing in the brief mentioned branches, phone support or location. Two of the twelve named a local institution as a secondary option.
What this showed me is that agents can carry context across companies
when they have somewhere to put it. Presence is not
persistence. The carrying is an artifact somebody writes.
Silent elimination
Customers relying on third-party information is obviously not new;
they always have. What changes is that software can now assemble that
information, apply the customer’s constraints and reach the decision
itself, at a point where the company has no visibility of any of it.
One of the clearest examples came from the telecom run. AT&T was
evaluated using pricing found on a competitor’s site. The agent
concluded that it was not competitive enough to pursue, never
successfully loaded an AT&T page, and omitted the company from the
final recommendation. From AT&T’s perspective, there was no
observable customer event.
The reverse also occurred. Texas Farm Bureau was put forward as an
insurance option on the strength of third-party information, again
without the agent visiting the company’s own property. Being selected
can therefore be almost as invisible as being rejected.
Silent elimination seems the useful term for this: a
company disappears from consideration before receiving any observable
interaction. And it was not an artifact of the first four journeys.
Across the second study’s runs, 239 of 277 brand eliminations — 86 per
cent — happened without the agent ever loading a readable page belonging
to the brand it dropped.
Measurement before contact
Most customer-experience measurement becomes useful only once the
customer reaches the enterprise. Contacts, sessions, transfers,
abandonment and resolution can all be observed from that point
onward.
Personal agents introduce a layer before it, where the enterprise may
not know that a need existed, that it entered consideration, what the
agent believed about it, or which requirement caused it to
disappear.
The experiments surfaced a related problem. The agents recorded 49
facts about companies that could not be confirmed on those companies’
own sites, and on six occasions they relied on information already held
in model memory, three of which affected the outcome. The agent was
therefore assembling a picture of the market that no company in it could
see, correct or even know existed.
That made me think differently about a familiar quality question. We
normally ask: what did our representative tell the customer? There is
now a possible counterpart: what did the customer’s
representative believe, and tell them, about us?
Where the enterprise remains authoritative
The agents went a long way using public information, but eventually
they reached facts they could not establish reliably: account state,
actual eligibility, address-level availability, a customer-specific
quote or live service conditions.
This is where the boundary becomes useful. The personal agent can
carry context, compare and decide, while the enterprise remains the
source of truth for what is actually available, what it is willing to
commit to, and whether that commitment was fulfilled.
I pushed that boundary into software. The current implementation
keeps five states separate — customer requirement, enterprise
proposition, accepted commitment, execution and evidence of outcome —
and tests whether they can remain connected even when the surrounding
journey is orchestrated elsewhere. It also covers reviewed service
conditions, exact acceptance, bounded disclosure, case replay and
reconciliation against later evidence.
The point is not to build another service agent. It is to test a
different continuity model. If the enterprise no longer owns the whole
journey, it may not need the whole conversation. What it needs is the
authoritative chain of what was needed, what was offered, what was
accepted, what happened and what can be established afterwards.
What this changes for CX
Taken together, the studies suggest a different division of labour.
The personal agent can increasingly operate the customer's side of
discovery, comparison, memory and decision-making, while the enterprise
remains authoritative for its side of the relationship.
That moves the CX problem. The task is no longer only to design and
optimise a journey the enterprise controls. It is also to make the
enterprise's part of a journey orchestrated elsewhere legible,
bindable, resumable and verifiable.
The customer-side agent does not need to become part of the
enterprise stack, and the enterprise does not need to own the customer's
context. What needs to survive between them is the relationship state
that matters.
In that model, continuity is no longer primarily continuity
of conversation. It is continuity of commitment.
Amine El Afrit is a Casablanca-based finance executive working across strategic finance, capital allocation, business transformation and operating-model design. His independent research explores AI-driven business-model change, markets and decision systems.