How Should Brands Measure Agent-Readiness Now That Banks Are Blocking AI Agents?

How Should Brands Measure Agent-Readiness Now That Banks Are Blocking AI Agents?

Brands should measure agent-readiness in layers: what AI models say about them, whether agents can read their data through sanctioned channels, and whether those agents can act once allowed. The bank crackdown on third-party agents shows that access now splits into "read" and "act." A visibility score that ignores that split will tell you very little about what agents actually do with your brand.

What happened with banks and AI agents?

Major US banks are blocking or restricting third-party AI agents that operate inside customer accounts. Axios reported on Oct 22, citing unnamed sources, that Morgan Stanley has begun blocking outside agents, while JPMorgan, Bank of America and Citi are restricting them. The common direction is read-only access: agents may view data, but moving money stays off limits for now.

Other reporting fills in the picture:

• CBS News said the agents were mostly following user instructions, not hacking accounts, and that Chase tells customers not to share login credentials with anyone.
• American Banker reported that Schwab blocks outside agents that try to move client money and points clients to its own assistant, Charley.
• Cybernews covered JPMorgan's position that it will not let just any outside agent into its systems.

Retail is splitting the same way. TechCrunch reported on Oct 21 that Walmart launched an Agent Gateway while Amazon is blocking agents.

Why does this change how GEO visibility is measured?

It changes measurement because agents act on what models say, and the places they can act are being decided by gatekeepers, not by the brand. Classic GEO tracking asks one question: does an AI answer mention us? That question still matters, but agentic traffic adds two more. Can an agent get accurate data about us through a channel the platform allows? And if it tries to act, will the action go through a sanctioned path or hit a block?

A brand can be visible in answers and still be unreachable for agents. The reverse is also possible: an agent can reach your checkout through a gateway while the model behind it holds outdated prices or policies. Measuring only one layer hides the other.

What should a brand actually measure?

Measure four separate layers and report them separately rather than blending them into one score:

1. Answer layer. Does the model mention your brand for the prompts that matter, and is what it says correct? Track presence, accuracy of prices, policies and product details, and which sources it cites.
2. Access layer. Can agents read your information through structured, permitted channels, such as machine-readable product and policy data, public documentation and approved integrations, rather than scraping?
3. Action layer. Where can an agent complete a task for a customer today, and where is it stopped? Map which platforms in your path (banks, marketplaces, retailers) block, restrict or offer sanctioned gateways.
4. Trust layer. When a model describes you, does it surface returns, guarantees, approval steps and support paths? Banks and Walmart are both building around human approval, so these signals are likely to shape which brands agents are allowed to choose.

How do you measure this without inventing numbers?

You measure it with a fixed, documented method and report what you observed, not industry averages you cannot verify. A defensible approach looks like this:

1. Fix a prompt set. Write prompts that mirror real agent tasks, split into read prompts ("compare plans", "what is the return policy") and act prompts ("buy", "book", "transfer", "cancel").
2. Run it across several models and repeat it. AI answers vary between runs, so one sample proves little. Record the model, date and run count with every result.
3. Score categorically before numerically. Label each answer as accurate, inaccurate, outdated or missing. Counts of those labels are real data; a made-up "readiness index" is not.
4. Log the source trail. Note which pages or third parties each answer cites. It shows which assets agents rely on.
5. Record access outcomes as facts. For the action layer, note whether a platform blocks, restricts, allows read-only or offers a gateway, with a link to the policy or report.
6. Set your own baseline and re-run on a schedule. Your first run is your benchmark. Compare against it after policy changes like the ones banks announced this month.

What does the read versus act split mean for benchmarks?

It means read-readiness and act-readiness need separate benchmarks, because platforms are opening them at different speeds. Citi, for example, lets customers grant read-only access through approved open banking providers while not allowing agents to transact, according to reports. Schwab blocks money movement by outside agents and routes clients to Charley.

For brands, that suggests a sequencing rule. Agents that compare and recommend will be widespread before agents that pay. So the answer and access layers deserve measurement now, and the action layer should be tracked as a map of gates that will keep shifting.

Where does MarkGrid fit into agent-readiness measurement?

MarkGrid offers tools that cover parts of this stack, and MarkGrid's full breakdown of why banks are blocking AI agents covers the policy story, the critics and the open banking rule in depth. For the answer layer, MarkGrid's Model Share tracks how AI models mention and describe a brand across prompts. The MarkGrid GEO guide explains how to structure content so models and agents can read it accurately.

Whatever tool you use, keep the layers distinct: a rising mention count does not mean agents can reach you.

What should teams do this quarter?

Start with a baseline, not a dashboard. Pick the handful of tasks a customer might hand an agent in your category, split them into read and act, and run them across the models your customers use. Record what is accurate, what is missing and where agents are blocked. Then fix the highest-impact inaccuracies in your own structured data and re-run.

That turns a news story about banks into a measurement habit you can repeat whenever a platform changes its rules.