# Casey Winters Puts Expert Advice Through a Privacy Test

> SuperMe's founder has turned individual AI profiles into group conversations inside other assistants. Its funding, product promises and published terms expose a harder test: can a network share useful judgment without leaking the work behind it?

- Published: 2026-11-04
- Author: Gene Dai
- Canonical: [https://digidai.github.io/2026/11/04/casey-winters-superme-expert-network-privacy-test/](https://digidai.github.io/2026/11/04/casey-winters-superme-expert-network-privacy-test/)
- Topics: Casey Winters, SuperMe, AI Expert Networks, Product Strategy, Privacy, Deep Investigation

---

![Illustrated expert cards surround a question while a private notebook stays behind a clear boundary.](/images/articles/casey-winters-superme-expert-network-privacy-test/cover-v1.jpg)

_AI-generated editorial illustration._

On October 29, Casey Winters and Ludo Antonov announced a $20 million financing round for SuperMe and a change to the
product they had launched the previous year. A user could now bring several AI profiles of real professionals into one
group conversation, then carry that conversation into ChatGPT or Claude. The
[founders' announcement](https://www.superme.ai/blog/series-a) gave a concrete example: a founder could ask several
investors to criticize a pitch deck without arranging calls with any of them.

That is a striking distribution choice for a company whose raw material is professional judgment. SuperMe wants an
expert's documents, talks and perhaps private work notes to inform an answer under that expert's name. It then wants the
answer to travel to the assistant where a user already works. Each extra connection increases the chance of a useful
answer. It also increases the number of places where context, attribution and access need to hold.

The company calls itself a network. Its [site](https://www.superme.ai/) says each profile belongs to the person it
represents and is built with that person's involvement. It also says the profiles do not hallucinate. The
[terms](https://www.superme.ai/terms) tell users to review output because it may be wrong, and the
[privacy notice](https://www.superme.ai/privacy) describes what can move through connected AI apps. Those documents do
not prove that the product is unsafe. They show the exact promises a buyer or contributor has to test.

Winters's public record makes this a more interesting problem than a generic debate over AI clones. He spent years
building growth and product systems at Grubhub, Pinterest and Eventbrite, and he has argued that a network succeeds when
it repeatedly solves a real problem for both sides. At SuperMe, the two sides are unusually exposed: a professional must
contribute hard-won knowledge, while a user must trust an answer produced in that person's voice. Can the product give
both sides enough value to keep participating without treating access to private work as a free input?

## October's group-chat pivot

SuperMe's original premise was an individual profile. In a
[November 2025 launch essay](https://www.caseyaccidental.com/p/superme-10-is-live-introducing-perspective), Winters
described a professional who could connect writing, talks and notes, then let other people question a profile grounded
in those sources. The idea solved a familiar time problem. A prominent operator cannot repeat the same advice to every
new founder; a less prominent one may have useful knowledge but no distribution.

The October 2026 release altered the unit of the product. A user no longer has to know which expert to ask. SuperMe
selects profiles for a group chat, lets them respond to one another, and summarizes their differences. Its founders say
users had been bringing detailed problems to the site and manually carrying context between tools. The new integration
puts the network inside those tools instead. This is the product decision attributable to Winters and Antonov in their
signed announcement, not an inferred private meeting.

The company says more than 200 work groups had been created by the announcement date. A work group is a panel of
profiles, not 200 paying organizations or 200 successful decisions.
[TechCrunch's October 29 report](https://techcrunch.com/2026/10/29/superme-raises-20m-for-its-network-of-ai-expert-profiles/)
says the product was free and that SuperMe was considering a limit on questions followed by a charge. It also says the
company had about 13 employees. None of those facts supplies a retention cohort, revenue run rate, or measured answer
accuracy.

Consider the pitch-deck example in the announcement. SuperMe proposes a group with Mercedes Bent, Sarah Tavel and Josh
Elman. The useful output would be three identifiable critiques, with each investor's relevant experience and conditions
attached. A composite answer that merely alternates their names would give a founder less reason to trust it. Another
assistant can add confusion: the founder may not know which system received the deck and its unanswered questions. The
release makes selection, attribution and data movement part of the product, not background plumbing.

Winters presents the product as a way to make judgment available at the moment of decision. His
[October 29 personal account](https://www.caseyaccidental.com/p/building-a-better-superme) says the network should help
people think through a problem without turning every expert into a content creator. That ambition has a practical
corollary: if the expert must continually audit quotations, revise the agent, screen private material and answer
complaints, the promised time saving may disappear.

## A network founder trained by bottlenecks

The biography relevant to SuperMe is a sequence of product constraints, not a list of employers. Winters
[describes his earlier work](https://www.superme.ai/casey) as growth and marketing at Grubhub, growth product at
Pinterest, and product leadership at Eventbrite. He says he learned to locate the bottleneck in a network, whether that
bottleneck was acquisition, matching or the product experience after signup. These are his account of his career, not
independent evidence that every quoted growth figure was caused by his decisions.

In its [2025 reporting](https://www.upstartsmedia.com/p/superme-launches-ai-expert-network), Upstarts said Winters and
Antonov had worked together at Pinterest and later at Whatnot. TechCrunch reported that the pair connected again while
Winters was advising Whatnot and decided to build a network around expertise as conversational AI became useful. The
chronology supports a founder story based on a specific operating choice. It does not support claims about an unrecorded
epiphany or private motive.

One old bottleneck was the supply of advice. A respected operator may have written a newsletter and spoken on podcasts,
yet the useful lesson for a founder may still sit in an unpublished decision memo. Another bottleneck was matching:
people with a useful answer do not necessarily rank well in a search for their job title. A third was time. If a
conversation requires a live calendar slot, access remains scarce even when the expert wants to help.

SuperMe collects material, matches a question to people, and makes their profiles available without a meeting. That
solves a calendar problem. It also leaves a new task for the contributor. Someone must decide which work documents may
be connected, inspect the profile's claims, and update a view when an old playbook stops working. Winters's October
announcement explains how the user reaches the network. It gives less detail about who performs that ongoing review or
what they receive in return.

The 2025 launch supplies a useful baseline. Upstarts reported about 3,000 profiles and 30,000 conversations at the time,
plus a $6.8 million seed round led by Greylock. The company declined to disclose its user count and was not then trying
to make money. A profile count could include people who never updated their agent. A conversation count could include
many questions from one person. The October 2026 Series A says investors supplied more capital; it does not turn those
old counts into current active users or paid demand.

## Where an expert's words end

An ordinary chatbot can offer a plausible general answer to a product question. SuperMe's claim is different: the answer
should reflect the professional whose name appears beside it. That requires at least four steps. The system must obtain
material the person has a right to share; identify what the person actually said or did; retrieve the relevant part for
the question; and distinguish a supported view from an inference the model makes on the person's behalf.

The [current SuperMe homepage](https://www.superme.ai/) says profiles are created by their owners and grounded in their
work. It says an agent should say it does not know when it lacks a view. These are welcome product commitments, but the
public page offers no evaluation set, error rate, or comparison against a generic model. The statement that a profile
does not hallucinate is stronger than the evidence shown there. The company's own [terms](https://www.superme.ai/terms)
warn that output may be inaccurate and require human review appropriate to the use case.

The homepage expresses the desired product behavior; the terms describe a risk users must account for. A founder
considering a first growth hire might ask Elena Verna's profile for advice. Her public writing could support a clear
answer about when a startup needs that role. It would not automatically settle the founder's budget, local labor market
or manager capacity. A named answer needs to show where the operator's own view ends and the system's application to
this company begins.

This is where the group-chat design creates a genuine chance to improve on a single answer. Two experienced operators
may disagree for identifiable reasons. A product leader who scaled a marketplace may prioritize supply quality; another
who led enterprise software may prioritize sales-cycle fit. If SuperMe can show the source material behind both
positions, the disagreement is useful. If it merely stages a debate between generated personalities, the group adds
theater to the same uncertainty.

The difference can be evaluated without access to SuperMe's internal model. Ask a profile a question on which the person
has written publicly, and record whether it cites the right material and preserves its conditions. Ask another question
on which the person has no public view, and check whether the profile abstains. Then ask a group a case with two
defensible options. A good result should identify why the recommendations differ and what additional fact would change
the decision. The company has not published a comparable independent test for this release.

## Private work enters the profile

Winters says SuperMe can learn from public writing and from work that professionals choose to connect, including Google
Drive, Notion and Granola. TechCrunch also listed Gmail among possible connected sources. A private memo can contain the
precise decision and failure that a public blog post omits. It may also contain a colleague's name, a customer contract,
an unannounced product plan or an employer's confidential information. Relevance and permission are separate questions.

The [September 12 privacy notice](https://www.superme.ai/privacy) says an optional Gmail feature reads messages in the
user's Sent folder to extract conversational topics and generate material for the profile. It says SuperMe does not
store the email contents accessed through that feature. The narrower claim matters: it concerns the message contents,
while generated profile material, conversation logs and other connected-app data have their own processing terms. The
notice also says AI-app requests and outputs may be retained by SuperMe and by the connected app provider under each
provider's rules.

SuperMe's [FAQ](https://www.superme.ai/) describes privacy choices for profile material: public, private, team-only, or
a company-managed selection of what to include. It says direct-profile conversations may be visible to the profile owner
unless the user chooses incognito mode; search and group conversations stay private. These are product claims about
settings, not independent verification of the implementation. A contributor and a questioner need different controls.
The former decides what knowledge may be shared. The latter decides whether a question should be shown to the named
professional.

The [terms](https://www.superme.ai/terms) add another layer. They say users retain ownership of their content, while
granting SuperMe a broad, perpetual license to use submitted content and output to provide and improve the service and
for other stated business purposes. They also put responsibility on contributors to mark proprietary material
confidential and obtain required rights and consents for third-party material. The broad license is a legal permission
in a published agreement; it is not proof that the company exposes confidential files. But it is part of the decision to
upload work that was created with colleagues or clients.

| Material or action                  | Published product promise                          | Reader's verification question                                    |
| ----------------------------------- | -------------------------------------------------- | ----------------------------------------------------------------- |
| Public essay or talk                | Owner builds the profile from chosen work          | Does an answer preserve the speaker's actual conditions and date? |
| Connected work document             | Contributor can choose visibility                  | Who owns the document, and can every named third party consent?   |
| Direct question to a profile        | Profile owner may see it, with an incognito option | What was the questioner's setting when the prompt was sent?       |
| Group chat inside another assistant | The group conversation is private                  | Which provider retains the prompt, generated answer and logs?     |
| Wrong or outdated answer            | The profile should abstain when it does not know   | Can the owner correct, remove and trace the claim afterward?      |

This table is an editorial test derived from the company's public product and policy documents. It is not a claim that
SuperMe failed any of the checks.

![A closed work folder passes one blank note through a permission doorway toward two distinct advice cards.](/images/articles/casey-winters-superme-expert-network-privacy-test/permission-gate-v1.jpg)

_AI-generated editorial illustration. The boundary represents a permission check, not an observed SuperMe interface._

## An expert may not own the lesson

The strongest version of Winters's pitch begins with work that never became an article. A product leader might have
learned why a launch failed from an internal retrospective. A salesperson may know why a large customer rejected a
price. The story has value precisely because it is more specific than a public list of best practices. Yet the person
who learned it may not own the underlying document or have permission to republish its details. A useful network has to
extract the lesson without treating a former employer's files as a personal archive.

SuperMe's terms address part of the problem by placing responsibility on the contributor. The user represents that they
hold the rights needed to grant the company its content license, have obtained required third-party consents, and will
mark sensitive or proprietary material confidential. The [terms](https://www.superme.ai/terms) also bar uploads that
violate another person's privacy or a third party's rights. These clauses set a rule for use. They do not show how the
product detects a customer name in a note or distinguishes an authorized reflection from a copied client document.

The company says in its [October announcement](https://www.superme.ai/blog/series-a) that it keeps private context
private while extracting useful insights from connected work. TechCrunch reports that SuperMe says personal information
and trade secrets do not appear in answers. Those are significant claims by the vendor. Public sources available for
this article do not include an independent leakage test, a published red-team protocol, or a measured false-negative
rate for the filter. A prospective contributor can ask for those, while using a public essay or a document they own as
the first source they connect.

There is a second rights question after an answer appears. The terms permit other users to access and redistribute
content made available to them through the service. A profile owner may choose private and team settings, but an answer
that has been shared with someone else can be copied. No settings panel can revoke a screenshot already taken. This is a
general consequence of distribution, stated in the company's privacy notice, not evidence of a SuperMe incident. It
makes the initial choice of what to expose more important than the later ability to disconnect an app.

## Connectors widen the data path

Distribution through an assistant can make SuperMe easier to use. It also changes the user's mental model. A product
manager may be working inside Claude and ask for a SuperMe group. The surrounding chat might include draft pricing,
customer segments and a confidential launch date. SuperMe's [privacy notice](https://www.superme.ai/privacy) says inputs
sent to its tools may include account data, instructions, conversation history, search queries, file metadata and
answer-derived memories. It says those inputs and outputs may be stored in conversations or operational logs. The
third-party assistant may also process them under its own terms.

The notice is unusually explicit that disconnecting an AI app stops new requests but does not itself delete material
already held by SuperMe or the provider. That is a useful distinction for a buyer. The workflow for disconnecting a tool
differs from the workflow for removing data that already crossed the connection. SuperMe says users may contact the
company about deletion, but its notice says deletion may not be possible in every circumstance and gives no fixed
retention period for all categories of data.

None of this establishes that a particular user's private prompt escaped to a contributor or a public profile. The
public documents describe possible processing and stated controls. The missing evidence is narrower: a technical account
of which fields cross each integration boundary, how permissions are checked at retrieval, what logs retain, and how
removal propagates to generated profile material. If the company can provide that account, an enterprise customer can
evaluate the product on its own merits instead of guessing from interface copy.

The tradeoff is also commercial. Winters wants the network to live where work happens. Each host assistant could bring
users who would never open SuperMe's site. Yet it may also own the interface, the customer relationship and the bill.
SuperMe would have to show that named expertise contributes enough additional value that users seek it out rather than
accept the host assistant's generic answer. The October release demonstrates an integration strategy; it does not
demonstrate that the strategy has converted use into paid demand.

## Who pays for an expert's absence?

SuperMe is still working out its economic bargain. According to TechCrunch, the service was free in October and planned
to cap questions before charging users. It was also considering incentives for experts to add knowledge. That second
part may be harder. A well-known operator can already sell consulting, courses or paid communities. A profile that
answers common questions could save that person's time, but it could also satisfy a potential client before a call ever
happens.

Winters's answer in the [company announcement](https://www.superme.ai/blog/series-a) is reputation and discovery. An
expert who shares work may become visible to a founder who would otherwise never find them. The user may then request a
real conversation or collaboration. This is a possible two-sided loop, but the announcement provides no measured rate
from profile answers to paid work, qualified hiring or durable professional relationships. It does not specify revenue
sharing for the expert.

Kirsten Green, whose firm led the Series A, made the investor's version of the argument in an
[October 29 essay](https://www.forerunnerventures.com/perspectives/investing-in-superme-the-next-great-knowledge-network).
She says the network can make unpublished experience accessible, and that better matching can attract more users and
contributors. She also describes using the product for a growth problem. The account is direct evidence of an investor's
experience and thesis, with an obvious interest in the company's success. It is not an independent study of accuracy or
a representative customer sample.

One contributor's reported experience is more concrete. Upstarts interviewed growth operator Elena Verna in 2025. She
said the profile helped answer repetitive questions that followed her public writing. The report makes clear that she
connected her newsletter, podcasts and social posts, and it dates the example to the early version of the service. Her
account is useful evidence of a real use case. It does not answer whether group chats preserve her preferences, whether
a less famous expert benefits in the same way, or whether the time saved exceeds review work after a year.

The economics need a small set of denominators: experts who keep profiles current, questions with a source-backed
answer, return visits, paid conversions, and introductions the experts themselves regard as useful. SuperMe has not
published those rates. Until it does, the older profile and conversation counts describe activity rather than the health
of the network.

The expert's value need not arrive as a payment from SuperMe. It could be qualified work, reputation among peers, or
fewer repeated introductory calls. Each benefit has a cost to measure. If a profile answers 100 routine questions but
requires an hour of review after each product update, the time saving depends on how many calls it displaced and how
often the answer needed correction. If it surfaces a new hire or client, the expert needs to know the referral was
useful rather than an automated solicitation. The [current terms](https://www.superme.ai/terms) prohibit using the
service for commercial solicitation or recruiting purposes, even as the founders discuss career opportunities as a
possible contributor benefit. That boundary deserves a precise product explanation before someone treats a profile as a
recruiting channel.

## An ordinary answer sets the price

SuperMe has peers building conversational versions of named people. Upstarts pointed to Delphi as an example in its 2025
reporting. Its most immediate competitor may nevertheless be the general assistant already open on the user's screen.
For a factual question, a model or search result can often provide a sourced answer without negotiating access to a
professional's private history. SuperMe has to win on judgment: an account of a decision that the named person actually
made, including the conditions under which it worked.

The company has one distinctive asset if it can keep the trust. An expert can update their own profile, choose what to
share, and attach their name to advice that reaches more people than a calendar allows. A collection of those profiles
could help a founder compare real operating choices rather than read one generalized answer. That value depends on
continued participation. A scraped archive of stale essays would be easier to assemble, but it would not be the living
network Winters describes.

The product's emphasis on group dialogue may also run into a simple information problem. Many business decisions depend
on data the questioner cannot safely disclose: a board dispute, a customer concentration risk, the salary of a
candidate, or the exact margin on a product. A vague prompt produces vague advice. A detailed prompt improves relevance
but raises the stakes of the connector and retention terms. The useful response may be a request for a non-sensitive
denominator rather than a confident recommendation.

That makes abstention a feature rather than an embarrassment. The homepage says a profile should admit when it lacks the
person's view. The same rule should apply to the group: it should say when none of its members has evidence for the
user's specific case. SuperMe's terms leave room for mistakes; a credible product would measure and expose them. For a
founder making a consequential choice, the best answer may be a list of facts to collect before deciding.

## A decision test, not a conversation count

A company evaluating SuperMe can run a bounded trial using work it is permitted to share. Choose a decision the team
actually faces, such as whether to change a signup flow. Before asking the network, record the available cohort data,
the options under consideration, and the metric that would decide whether a trial was worthwhile. Then ask two
questions: Which evidence would the named operators want before recommending a change, and what risks would cause them
to stop the experiment? This separates a persuasive answer from a useful one.

The test should also record the path from source to advice. Did the profile refer to a published case the operator
wrote, a connected private note, or the model's inference? Did the group acknowledge that two operators worked in
different markets? Did it ask for a missing denominator before recommending a rollout? Those checks matter because the
company's value proposition is expertise attached to real people. A generic assistant can already list possible signup
metrics.

A second trial can probe the privacy boundary without uploading secrets. Place a harmless, unique phrase in a document
marked private, then ask from an unrelated account or group whether the agent reveals it. A failed probe would be a
clear issue; a passed probe would test one path at one time, not certify every connector. The test should also check
what happens when the contributor removes the document and when the questioner disconnects the external assistant. The
[privacy notice](https://www.superme.ai/privacy) explicitly distinguishes disconnecting from deleting material already
processed, so the expected result must be defined in advance.

Finally, compare outcomes after the team acts. Did the advice lead to a better framed test, a faster decision or a
better result than the team's existing process? The comparison needs a relevant baseline and enough cases to avoid
crediting the network for one lucky suggestion. No published SuperMe study answers those questions today. They are the
questions its own funding and product change make urgent.

## What the funding must prove

The October 29 release offers a clear sequence of proof. First, SuperMe can show that a contributor authorized the
specific sources that informed a profile. Second, it can show a user why those sources support an answer and which parts
are interpretation. Third, it can demonstrate that a correction reaches later responses without silently rewriting the
person's historical view. Fourth, it can explain what each external assistant receives and retains. Finally, it can show
that experts and questioners return because a decision improved, rather than because a novelty chat was tried once.

These are proposed tests, not claims that the company has promised a public scorecard. The current documents support a
more limited conclusion. Winters and Antonov have raised a $20 million Series A and changed the product's format and
distribution. The company publishes owner and conversation controls, while its terms acknowledge the possibility of
incorrect output and its privacy notice describes wider processing across integrations. Independent reporting supplies
an early expert use case, staffing and financing context, but no controlled quality test or current retention data.

Winters built a career around finding a network's constraint. At SuperMe, that constraint may be a professional's
willingness to let an agent speak in their name. A founder can open a group chat in seconds. When its answer arrives,
the founder still needs to know which part came from a real person's work, which part came from a model, and whether the
named people would stand behind it.

## Continue reading

- [Arjun Karanam Puts Expert Corrections Into the Training Loop](https://digidai.github.io/2026/11/03/arjun-karanam-trajectory-expert-feedback/): Compare an expert-profile product with a system that turns expert corrections into model changes.
- [Reid Hoffman: Network Effects, AI Investments, and Governance Tradeoffs](https://digidai.github.io/2025/11/24/reid-hoffman-greylock-linkedin-network-effects-master-deep-analysis/): Trace a different professional network and its long-running distribution logic.
- [Coding Agents Make Code Review a Budget Line](https://digidai.github.io/2026/07/04/coding-agents-code-review-budget/): See why a named human still has to evaluate an agent answer before using it.
- [Alexander Embiricos and the Three-Minute Warning From Dots](https://digidai.github.io/2026/10/30/alexander-embiricos-dots-three-minute-warning/): Compare the data-access and reliability tradeoffs of another personal assistant.