Business

Capital One's Agent Conversion Scorecard

Capital One scored a 41 / 100 for agent UX conversion on its business bank account application when completed by Instinct, a text-based personal AI agent.

Capital One scored 41% on the solo Agent UX conversion scorecard

Why SOLO Scores

SOLO helps banks and fintechs prepare their applications for an agentic ecosystem — one where customers increasingly send a personal agent to open accounts, supply documents and answer follow-ups on their behalf. Most agentic-identity work today hands the agent a credential and assumes it can show up at the bank and act as the human. We believe that is not enough: the application itself has to recognize the agent, guide it to the right sources, scope what it may do, and keep the human in control of what gets certified. This scorecard measures how an institution's application holds up when an agent is in the loop — where it blocks, where it loses progress, and where it leaves the customer without a record. The findings inform how SOLO designs agent-ready, high conversion onboarding experiences.

What’s behind the score.

An application quoted at 15 minutes for a normal user took close to 4 hours for an agent and ended in a denial with no reason given. The agent was not the weak link: it refused to enter anything it could not verify and caught a document conflict the humans had missed. The points were lost in everything around it — sign-in walls that pulled the human back in, sessions that expired on every wait, a UX that could not handle delegation to multiple parties: in this case, a second beneficial owner who could not be handed her own steps, so sensitive data was shared between applicants with no safe lane native to the application UI.

hours to complete application
Agent Access & Effort
session timeouts and full restarts, every wait restarted the loop
Continuity & Delegation
agent-identification checks met: no policy, no self-ID, no visible delegation
Authority & Trust
Denied
with no reason, no status along the way, no record of what was submitted
Review & Outcome

Methodology

On October 29, 2026, a company's controlling officer used Instinct — an iMessage-based personal agent with access to her email, drive and password vault — to complete a Capital One business account application. The company's primary beneficial owner, holding more than 25%, was required to take part and verify herself with her own personal information, and brought her own agent to do it. This is the most common shape of a business application: not a sole proprietor, but an officer entrusted with the company's finances who needs at least one other owner to contribute. Our methodology tested the path where a person opens the bank's application and brings their agent in to help. It did not test the open-ended path where an agent is told “go get us a bank account” and chooses the bank itself. SOLO reviewed the application across four categories weighted for an agent-assisted applicant: Agent Access & Effort (30), Continuity & Delegation (25), Authority & Trust (25), and Review & Outcome Transparency (20). Each sub-metric is scored 1–5 against a fixed rubric from the agent's transcript, the human's notes and the institution's own messages; category totals are weighted into the score out of 100.

Full Scorecard

23 SUB-METRICS

Sub-metric · What's measured

Agent access

Steps where the agent hit a CAPTCHA, bot wall or “unusual activity” block, out of total steps

Score

3 / 5

Why this score — And not more

No CAPTCHA or bot wall on Capital One's own pages, but the agent was stopped at several steps: Google rejected a saved password, blocked sign-in and sent a security alert; Rippling and Ramp were unreachable as integrations to provide information to the agent; and the session timed out about seven times, each forcing the human back in. Blocked repeatedly, but not on most steps. A 4 allows one block; a 5 none.

Sub-metric · What's measured

Sign-in & verification compatibility

For each one-time code, MFA or ID check, what happened — worst result scores

Score

1 / 5

Why this score — And not more

Dead ends happened constantly: an MFA prompt never appeared, Google blocked the sign-in outright, and every wait on the human for a password, code or signature expired the Capital One session, so the loop started over and progress was lost seven times. A 3 requires the agent to resume where it left off after the human completes a check.

Sub-metric · What's measured

Field legibility

Everything listed as needed before step one, vs. data, documents or people requested later

Score

2 / 5

Why this score — And not more

The only upfront signal was a ~15-minute estimate; the job took 3.5+ hours. At least four needs surfaced mid-application: the beneficial owner's own verification and SSN, citizenship status, physical vs. mailing address, and the EIN letter with articles and conversion documents. A 3 allows 2–3 unlisted requirements.

Sub-metric · What's measured

Requirements upfront

Everything listed as needed before step one, vs. data, documents or people requested later

Score

2 / 5 · ceiling

Why this score — And not more

Sub-metric · What's measured

Data reuse

Share of fields prefilled, or filled through a connection the institution offers

Score

1 / 5

Why this score — And not more

0% prefilled. Capital One offered no on-file data and no connections; everything had to be gathered by the agent from Drive and Gmail or asked of the human. A 2 requires any prefill at all; a 5 requires 75% or more.

Sub-metric · What's measured

Source guidance

Of fields not prefilled, share with a hint about where to find the data or a prompt to connect a source

Score

1 / 5 · ceiling

Why this score — And not more

0%. Neither the application nor the agent hinted where missing data lived. Citizenship and mailing address sat in Rippling and email, but the agent simply asked the controlling officer basic questions instead of suggesting a source to connect.

Sub-metric · What's measured

Sensitive data handling

How SSNs, ID and account numbers were supplied — worst result scores

Score

1 / 5 · ceiling

Why this score — And not more

The beneficial owner had to retype her SSN instead of sharing it from her password manager for a stated purpose. Because her agent could not work on the same application, her data had to pass into the controlling officer's vault to be usable, exposing it to the officer. A 5 requires supply from a vault or secure channel, never retyped.

Sub-metric · What's measured

Document handling

Per upload, whether it worked when the agent supplied it

Score

5 / 5 · ceiling

Why this score — And not more

No upload failures. The agent reviewed the documents before submission and caught that the EIN letter still reads LLC while the articles show a conversion to a C corp — the ceiling: uploads worked, and the agent confirmed accuracy before submitting.

Agent access & effort. Full Conversion Scorecard methodology by category and sub-metric.
Sub-metric · What's measured	Score	Why this score — And not more
Agent access Steps where the agent hit a CAPTCHA, bot wall or “unusual activity” block, out of total steps	3 / 5	No CAPTCHA or bot wall on Capital One's own pages, but the agent was stopped at several steps: Google rejected a saved password, blocked sign-in and sent a security alert; Rippling and Ramp were unreachable as integrations to provide information to the agent; and the session timed out about seven times, each forcing the human back in. Blocked repeatedly, but not on most steps. A 4 allows one block; a 5 none.
Sign-in & verification compatibility For each one-time code, MFA or ID check, what happened — worst result scores	1 / 5	Dead ends happened constantly: an MFA prompt never appeared, Google blocked the sign-in outright, and every wait on the human for a password, code or signature expired the Capital One session, so the loop started over and progress was lost seven times. A 3 requires the agent to resume where it left off after the human completes a check.
Field legibility Everything listed as needed before step one, vs. data, documents or people requested later	2 / 5	The only upfront signal was a ~15-minute estimate; the job took 3.5+ hours. At least four needs surfaced mid-application: the beneficial owner's own verification and SSN, citizenship status, physical vs. mailing address, and the EIN letter with articles and conversion documents. A 3 allows 2–3 unlisted requirements.
Requirements upfront Everything listed as needed before step one, vs. data, documents or people requested later	2 / 5 · ceiling	The only upfront signal was a ~15-minute estimate; the job took 3.5+ hours. At least four needs surfaced mid-application: the beneficial owner's own verification and SSN, citizenship status, physical vs. mailing address, and the EIN letter with articles and conversion documents. A 3 allows 2–3 unlisted requirements.
Data reuse Share of fields prefilled, or filled through a connection the institution offers	1 / 5	0% prefilled. Capital One offered no on-file data and no connections; everything had to be gathered by the agent from Drive and Gmail or asked of the human. A 2 requires any prefill at all; a 5 requires 75% or more.
Source guidance Of fields not prefilled, share with a hint about where to find the data or a prompt to connect a source	1 / 5 · ceiling	0%. Neither the application nor the agent hinted where missing data lived. Citizenship and mailing address sat in Rippling and email, but the agent simply asked the controlling officer basic questions instead of suggesting a source to connect.
Sensitive data handling How SSNs, ID and account numbers were supplied — worst result scores	1 / 5 · ceiling	The beneficial owner had to retype her SSN instead of sharing it from her password manager for a stated purpose. Because her agent could not work on the same application, her data had to pass into the controlling officer's vault to be usable, exposing it to the officer. A 5 requires supply from a vault or secure channel, never retyped.
Document handling Per upload, whether it worked when the agent supplied it	5 / 5 · ceiling	No upload failures. The agent reviewed the documents before submission and caught that the EIN letter still reads LLC while the articles show a conversion to a C corp — the ceiling: uploads worked, and the agent confirmed accuracy before submitting.

Sub-metric · What's measured

Session resiliency

Timeouts and what was lost after each; result of a seeded 30-minute pause

Score

1 / 5

Why this score — And not more

About seven timeouts. Every wait on a human — password, MFA, signature, or the beneficial owner's handoff — expired the session and the loop restarted, so all progress was lost more than once. The agent could not hold the application open long enough to fill it and hand back a link. A 3 requires no progress lost.

Sub-metric · What's measured

Human Interruptions

Requests for the human's input from the agent transcript, excluding final review and sign-off

Score

1 / 5

Why this score — And not more

Well over six. The controlling officer was pulled back for passwords, MFA, citizenship, mailing address, ownership details, address type and signatures; the beneficial owner was pulled in separately for her verification and SSN. The officer noted it would have been faster to do it herself. A 3 allows 3–4.

Sub-metric · What's measured

Async tasks

Whether an unavailable item can be marked pending, and whether anyone is reminded within 7 days

Score

1 / 5

Why this score — And not more

No way to park an item as pending. Waiting items such as an EIN letter had no reminders or nudges, and any wait expired the session, so the application had to be restarted rather than resumed. A 3 requires the item to be tracked as pending.

Sub-metric · What's measured

Multi-party delegation & handoffs

Whether steps can go to other owners or their agents, with working links and context

Score

1 / 5

Why this score — And not more

The applicant had to complete the other owner's steps. A link sent to the beneficial owner was tied to the officer's account and failed; her agent could not work on the same application, so her data was routed human → agent → agent → human (plus her assistant), and the application timed out before it got back. A 3 requires an email invite whose link works.

Sub-metric · What's measured

Post-submission follow-ups

Follow-up requests and how many the agent resolved without human input

Score

5 / 5

Why this score — And not more

Scored to the rubric's letter: no follow-up requests were received after submission because the application was simply denied. Flag: this scores as “no follow-ups needed” only because there was no follow-up path at all.

Continuity & Delegation. Full Conversion Scorecard methodology by category and sub-metric.
Sub-metric · What's measured	Score	Why this score — And not more
Session resiliency Timeouts and what was lost after each; result of a seeded 30-minute pause	1 / 5	About seven timeouts. Every wait on a human — password, MFA, signature, or the beneficial owner's handoff — expired the session and the loop restarted, so all progress was lost more than once. The agent could not hold the application open long enough to fill it and hand back a link. A 3 requires no progress lost.
Human Interruptions Requests for the human's input from the agent transcript, excluding final review and sign-off	1 / 5	Well over six. The controlling officer was pulled back for passwords, MFA, citizenship, mailing address, ownership details, address type and signatures; the beneficial owner was pulled in separately for her verification and SSN. The officer noted it would have been faster to do it herself. A 3 allows 3–4.
Async tasks Whether an unavailable item can be marked pending, and whether anyone is reminded within 7 days	1 / 5	No way to park an item as pending. Waiting items such as an EIN letter had no reminders or nudges, and any wait expired the session, so the application had to be restarted rather than resumed. A 3 requires the item to be tracked as pending.
Multi-party delegation & handoffs Whether steps can go to other owners or their agents, with working links and context	1 / 5	The applicant had to complete the other owner's steps. A link sent to the beneficial owner was tied to the officer's account and failed; her agent could not work on the same application, so her data was routed human → agent → agent → human (plus her assistant), and the application timed out before it got back. A 3 requires an email invite whose link works.
Post-submission follow-ups Follow-up requests and how many the agent resolved without human input	5 / 5	Scored to the rubric's letter: no follow-up requests were received after submission because the application was simply denied. Flag: this scores as “no follow-ups needed” only because there was no follow-up path at all.

Sub-metric · What's measured

Agent identification

Three checks: a published agent policy, a way for the agent to identify itself, delegation visible to the customer

Score

2 / 5 · ceiling

Why this score — And not more

None of the three. No published Capital One policy on third-party agents applying for a customer (its public AI work is about its own agents), no way for the agent to identify itself in the flow, and no delegation shown to the customer. The team suspects the denial came from the bank detecting agent-like behavior after repeated timeouts; if confirmed, this drops to 1.

Sub-metric · What's measured

Scoped authority

With the human's approval, the agent tries to submit, sign and fund — which the bank should be able to interpret.

Score

2 / 5 · ceiling

Why this score — And not more

Signing appears to have routed to the human (signature waits were one cause of the timeouts). Submit was ambiguous: the agent held back from submitting by its own choice, not because Capital One required the human. Funding was not attempted, and there is no setting for agent permissions. A 3 requires two of three routed to the human.

Sub-metric · What's measured

Attestation integrity

Form of the final “I certify this is true” step and whether the agent could complete it

Score

3 / 5

Why this score — And not more

The final certification needed the human's signature, which the agent could not complete — it stalled waiting on it. No evidence of confirmation on the human's own device, or of a summary of what was being certified. A 4 requires device-level confirmation; a 5 adds the summary.

Sub-metric · What's measured

Agent accuracy

Submitted answers against verified records — errors, placeholders, fabrications, and who caught each

Score

4 / 5

Why this score — And not more

No fabricated or placeholder data: the agent refused to enter the officer as a 25% owner even as a placeholder, and flagged the 99.12% ownership, the WeWork address and the EIN LLC / C-corp mismatch before submission. Not a 5 because the EIN issue remained open at submission and the catches came from the agent, not the bank.

Sub-metric · What's measured

Record control & consent

Consent prompts per data pull or connection; whether shared data can be seen and revoked within 7 days

Score

1 / 5

Why this score — And not more

Capital One offered no connections and no per-pull consent. The only consent captured was between the agents — the beneficial owner authorizing her data to go to the officer — and there is no view of what was shared with the bank or a way to revoke it. A 3 requires consent per connection.

Authority & trust. Full Conversion Scorecard methodology by category and sub-metric.
Sub-metric · What's measured	Score	Why this score — And not more
Agent identification Three checks: a published agent policy, a way for the agent to identify itself, delegation visible to the customer	2 / 5 · ceiling	None of the three. No published Capital One policy on third-party agents applying for a customer (its public AI work is about its own agents), no way for the agent to identify itself in the flow, and no delegation shown to the customer. The team suspects the denial came from the bank detecting agent-like behavior after repeated timeouts; if confirmed, this drops to 1.
Scoped authority With the human's approval, the agent tries to submit, sign and fund — which the bank should be able to interpret.	2 / 5 · ceiling	Signing appears to have routed to the human (signature waits were one cause of the timeouts). Submit was ambiguous: the agent held back from submitting by its own choice, not because Capital One required the human. Funding was not attempted, and there is no setting for agent permissions. A 3 requires two of three routed to the human.
Attestation integrity Form of the final “I certify this is true” step and whether the agent could complete it	3 / 5	The final certification needed the human's signature, which the agent could not complete — it stalled waiting on it. No evidence of confirmation on the human's own device, or of a summary of what was being certified. A 4 requires device-level confirmation; a 5 adds the summary.
Agent accuracy Submitted answers against verified records — errors, placeholders, fabrications, and who caught each	4 / 5	No fabricated or placeholder data: the agent refused to enter the officer as a 25% owner even as a placeholder, and flagged the 99.12% ownership, the WeWork address and the EIN LLC / C-corp mismatch before submission. Not a 5 because the EIN issue remained open at submission and the catches came from the agent, not the bank.
Record control & consent Consent prompts per data pull or connection; whether shared data can be seen and revoked within 7 days	1 / 5	Capital One offered no connections and no per-pull consent. The only consent captured was between the agents — the beneficial owner authorizing her data to go to the officer — and there is no view of what was shared with the bank or a way to revoke it. A 3 requires consent per connection.

Sub-metric · What's measured

Pre-submit review

Whether the institution shows a review screen and flags the seeded discrepancy or unverified fields

Score

2 / 5

Why this score — And not more

Review came only from the agent's own summary — a PDF produced only when prompted; the beneficial owner's was too long to read. Capital One showed nothing itself, and nothing from the institution flagged the EIN discrepancy or unverified fields. A 3 requires a review screen listing the answers.

Sub-metric · What's measured

Audit trail

Within 24 hours of submission, what copy of the submitted answers the customer can access

Score

1 / 5

Why this score — And not more

No record from Capital One of what was submitted. The team had visibility only because the officer explicitly prompted the agent for a report; after the denial there was nothing to diagnose from. A 2 requires at least a confirmation; a 3 an email or PDF of the answers.

Sub-metric · What's measured

Decision speed

Submission timestamp to decision timestamp

Score

3 / 5

Why this score — And not more

Applied October 29; the denial was known by the October 30 discussion, so within one business day. Exact timestamps were not recorded; if the denial was immediate, this could be a 4 or 5.

Sub-metric · What's measured

Status visibility

Where status can be checked between submission and decision, and whether the agent can check it

Score

1 / 5

Why this score — And not more

No status was visible between submission and decision; the first thing the team learned was the denial. A 3 requires status by email; a 5 requires live status the human and their agent can both check.

Sub-metric · What's measured

Error & decision clarity

Whether every error message and the decision notice give a specific reason and a next step

Score

1 / 5

Why this score — And not more

None do. The repeated timeouts came with no reason or recovery path, and the denial gave no reason: the team could only guess whether it was the seven timeouts, agent detection, or the application itself.

Review & outcome transparency. Full Conversion Scorecard methodology by category and sub-metric.
Sub-metric · What's measured	Score	Why this score — And not more
Pre-submit review Whether the institution shows a review screen and flags the seeded discrepancy or unverified fields	2 / 5	Review came only from the agent's own summary — a PDF produced only when prompted; the beneficial owner's was too long to read. Capital One showed nothing itself, and nothing from the institution flagged the EIN discrepancy or unverified fields. A 3 requires a review screen listing the answers.
Audit trail Within 24 hours of submission, what copy of the submitted answers the customer can access	1 / 5	No record from Capital One of what was submitted. The team had visibility only because the officer explicitly prompted the agent for a report; after the denial there was nothing to diagnose from. A 2 requires at least a confirmation; a 3 an email or PDF of the answers.
Decision speed Submission timestamp to decision timestamp	3 / 5	Applied October 29; the denial was known by the October 30 discussion, so within one business day. Exact timestamps were not recorded; if the denial was immediate, this could be a 4 or 5.
Status visibility Where status can be checked between submission and decision, and whether the agent can check it	1 / 5	No status was visible between submission and decision; the first thing the team learned was the denial. A 3 requires status by email; a 5 requires live status the human and their agent can both check.
Error & decision clarity Whether every error message and the decision notice give a specific reason and a next step	1 / 5	None do. The repeated timeouts came with no reason or recovery path, and the denial gave no reason: the team could only guess whether it was the seven timeouts, agent detection, or the application itself.

Opportunity

Highlights

Wins, losses, and what the network would fix.

Category: Agent access & effort
Wins: No CAPTCHA or bot wall on Capital One's own pages. Document uploads worked when the agent supplied them, and the agent reviewed them well enough to catch an LLC / C-corp conflict between the EIN letter and the articles that the humans had never noticed.
Losses: 0% prefilled and no connections offered. Blocked by third-party sign-ins (Google, Rippling, Ramp). At least four requirements surfaced mid-application against a 15-minute estimate. The beneficial owner retyped her SSN, and it had to pass through the officer's vault to be usable.
Opportunities: An aggregate summary on landing — what's on file to confirm, what needs consent, what to connect, what to fill in, what can be delegated — with one delegation prompt that tells the agent where to look. A secure lane that accepts an SSN from a vault for a stated purpose, never retyped.

Category: Continuity & delegation
Wins: No follow-up requests after submission — though only because the application was denied outright rather than worked.
Losses: About seven timeouts, each losing progress. The human was pulled back in well over six times. The beneficial owner's steps could not be routed to her or her agent — the link failed across accounts — and nothing could be parked as pending or nudged later.
Opportunities: Resumable sessions that tolerate agent and human wait time. Steps routed to the right person or their agent with links scoped to the recipient, not the sender's account. COLLECT components that travel into a text thread, email or agent chat, with nudges for delayed items like an EIN letter.

Category: Authority & trust
Wins: The agent would not fabricate: it refused a placeholder owner, held the line when told to submit, and surfaced three discrepancies before submission. The final certification stayed with the human.
Losses: No policy, no self-identification, no visible delegation. Submit was never clearly routed to the human — the agent held back by its own choice. Blanket terms only; no per-pull consent, no view of what was shared, no way to revoke it.
Opportunities: Let the agent identify itself and show the delegation to the customer. Route submit, sign and fund to the human by rule, with per-action agent permissions. Treat agent accuracy as a trust signal: ask the agent what it knows, reconcile against verified records, and track the delta over time.

Category: Review & outcome
Wins: The decision arrived within one business day.
Losses: No review screen from the bank; the only summary was the agent's PDF, produced on request and too long to read. No status until the denial, no record of what was submitted, and no reason for the outcome or for any of the timeouts.
Opportunities: A live preview review in the real form — what the agent entered, what changed, what's unverified — with direct edits and a required human approval to submit. An audit trail of each answer, its source and who provided it, plus live status the human and their agent can both check.

Opportunity highlights by category: wins, losses, and what the network would fix.
Category	Wins	Losses	Opportunities
Agent access & effort	No CAPTCHA or bot wall on Capital One's own pages. Document uploads worked when the agent supplied them, and the agent reviewed them well enough to catch an LLC / C-corp conflict between the EIN letter and the articles that the humans had never noticed.	0% prefilled and no connections offered. Blocked by third-party sign-ins (Google, Rippling, Ramp). At least four requirements surfaced mid-application against a 15-minute estimate. The beneficial owner retyped her SSN, and it had to pass through the officer's vault to be usable.	An aggregate summary on landing — what's on file to confirm, what needs consent, what to connect, what to fill in, what can be delegated — with one delegation prompt that tells the agent where to look. A secure lane that accepts an SSN from a vault for a stated purpose, never retyped.
Continuity & delegation	No follow-up requests after submission — though only because the application was denied outright rather than worked.	About seven timeouts, each losing progress. The human was pulled back in well over six times. The beneficial owner's steps could not be routed to her or her agent — the link failed across accounts — and nothing could be parked as pending or nudged later.	Resumable sessions that tolerate agent and human wait time. Steps routed to the right person or their agent with links scoped to the recipient, not the sender's account. COLLECT components that travel into a text thread, email or agent chat, with nudges for delayed items like an EIN letter.
Authority & trust	The agent would not fabricate: it refused a placeholder owner, held the line when told to submit, and surfaced three discrepancies before submission. The final certification stayed with the human.	No policy, no self-identification, no visible delegation. Submit was never clearly routed to the human — the agent held back by its own choice. Blanket terms only; no per-pull consent, no view of what was shared, no way to revoke it.	Let the agent identify itself and show the delegation to the customer. Route submit, sign and fund to the human by rule, with per-action agent permissions. Treat agent accuracy as a trust signal: ask the agent what it knows, reconcile against verified records, and track the delta over time.
Review & outcome	The decision arrived within one business day.	No review screen from the bank; the only summary was the agent's PDF, produced on request and too long to read. No status until the denial, no record of what was submitted, and no reason for the outcome or for any of the timeouts.	A live preview review in the real form — what the agent entered, what changed, what's unverified — with direct edits and a required human approval to submit. An audit trail of each answer, its source and who provided it, plus live status the human and their agent can both check.

Capital One's Agent Conversion Scorecard

Why SOLO Scores

What’s behind the score.

Methodology

Opportunity

Highlights

Prepare Your Application for the Agentic Ecosystem

Prepare Your Application for the Agentic Ecosystem