The New Front Door Is an AI Agent. Is Your Company Even There?

Update, October 2026: we followed this piece with Are You Willing to Become the Backend?, on what AI agents mean for strategy.

The people you serve have started sending an AI assistant ahead of them. That leaves every financial firm with two questions to answer:

1. What does the assistant say about you? When someone asks Claude or Grok about your products, your fees or how to file a claim, is the answer something you shaped, or whatever the assistant happened to find?
2. Can their agent act for them on your systems? Agents such as OpenAI's dots, Meta's Muse and xAI's Grok Bot already open browsers, fill in forms and complete tasks for their users. When one of them arrives at your site, is the path designed for it, or does it work by accident, if it works at all?

Try this before your next leadership meeting. Ask Claude the question your customers ask most often about your firm, phrased the way a customer would phrase it. Then ask Grok which firms it would recommend in your category, in your market, and see whether your name comes up at all.

Whatever comes back is what many of them now hear before they hear from you. Policyholders, investors, small business owners, credit union members, brokers, advisors and HR buyers are putting to Claude, Grok, ChatGPT, Gemini and Copilot the questions they used to type into a search bar, and many take the answer at face value.

The front door has moved

For the past twenty years, digital customer experience has meant the website, the mobile app and the customer portal. Financial firms spent heavily to get those right. They mapped customer journeys, redesigned onboarding, simplified claims and fought for a place on the first page of search results. That investment still pays off, but it rests on an assumption that is starting to break: that customers will come to your door themselves.

More and more, the journey begins somewhere you don't control. A customer types a question into an assistant, and the assistant does the research for them. It reads your public pages, the PDFs you forgot were still online, review sites, forum threads and competitors' comparison pages, then condenses all of it into a single answer. Many customers act on that answer without ever opening your website. The visitor that matters most can't see your branding, your layout or the care that went into your design. It sees only the text it can reach, and it decides what to repeat.

Some customers have already moved past asking. They hand the assistant the task itself, such as getting a quote, comparing savings rates, filling in an application or starting a claim, and the assistant comes to your site to do it.

This is the early stage. The agents named above launched within the past few months, and we expect them to become far more capable over the next six months. As they improve, more people will trust them with real tasks. In financial services, that will likely start with research, quotes and status checks, and reach account changes and money movement as trust builds. Over time, a growing share of the visitors arriving at your site will be agents acting for someone.

Software teams already have a name for this: agent experience, or AX. The term began as a way to describe how easily AI agents can work with developer tools and platforms. It carries more weight in financial services, because the tasks involve claims, fees, eligibility and money, and a mistake costs the customer something real.

Customer experience asks how a person feels as they move through your service. Agent experience asks whether an AI acting for that person can find what it needs, understand it correctly and finish the job. It is customer experience in a new form, and like every other part of customer experience, it is either designed with intent or left to whatever happens by default. The customer is still the person you serve, even when an assistant does the visiting.

Question one: what does the assistant say about you?

People ask assistants two kinds of questions, and each carries a different risk. The examples below are illustrative.

The first kind is about the category, such as which brokers allow options trading in an IRA or which credit unions someone can join. Here the risk is that you never appear. A brokerage that has offered options trading in IRAs for years, but explains it only in a PDF and a logged-in help center, gets left out while three competitors are named. A credit union that opened membership to its whole county is still described by the employer-only rule on an old page, and a likely member opens an account somewhere else.

The second kind is about you specifically, such as how to file a claim or what a wire transfer costs. Here you usually appear, and the risk is that the answer is wrong. An assistant can hand a plan member the fax number from a claims guide the insurer retired years ago, or quote a small business owner a wire fee from an outdated schedule. The claim stalls, the statement shows a higher fee than expected, and in both cases the customer blames the firm.

Search engines at least showed people a page of links, so a customer could see where each result came from and judge it for themselves. An assistant gives one confident answer, usually with little sign of where it came from. Left to default, that answer is assembled from whatever the assistant can reach: an old PDF, a third-party review site, a forum thread, or a competitor's comparison page that describes you in the least flattering terms.

In none of these situations did anyone at the firm write the wrong answer, yet in every one the source was the firm's own content. Some of these failures quietly cost a customer and leave no trace. Others become complaints, and in financial services, regulators track complaints closely.

Question two: can their agent act for them on your systems?

Logins, two-factor codes and CAPTCHAs are not the hard part. Most agents simply hand those steps back to the customer and carry on. The friction that hurts is the kind neither the agent nor the customer can see, because it comes from how your systems grew over the years.

• Old paths that still work. Most financial firms have layers of history online: a landing page from a campaign that ended years ago, a legacy application form from before the portal, a retired quote tool that was unlinked but never switched off. A person browsing your site would never find them. An agent can, because it reads search indexes, sitemaps and old links, and it has no way of knowing which path is current. It completes the old form, the submission lands in a system or queue nobody watches anymore, and the agent reports success.
• Steps that only your staff know. Completing a task like a 401(k) rollover or a claim depends on knowing the order of things: which form comes first, which documents to attach, what happens after you submit. That knowledge usually lives in internal procedures and call scripts, and a customer who gets stuck simply calls and gets walked through it. An agent has nothing to follow. It pieces the sequence together from scattered pages, guesses at the gaps, and either stalls or does the steps in the wrong order.
• Confirmations built for human eyes. Many flows end with a page that a person reads as "done," but that never states plainly that the request went through or gives a reference number. An agent that can't tell whether it succeeded may try again, and a second submission can mean a duplicate application, a duplicate claim or a duplicate payment.
• Fraud controls that can't tell friend from foe. An agent moves faster than a person, retries when something fails and often connects from cloud servers. To your fraud controls, that looks like an attack. The likely result is a locked account, and the real customer ends up on the phone with your service center to prove who they are.
• Handoffs that lose the work. An agent fills in most of an application, then reaches a step only the customer can complete, such as a signature. If the customer's session doesn't carry over what the agent entered, they start again from the beginning, or they give up.

Illustrative scenario. A small business owner asks an agent to request a call from a local bank about a line of credit. The agent searches and lands on a landing page from a promotion the bank ran two years ago. The page is no longer linked from the bank's site, but it is still live, and its form still accepts submissions. Leads from that form go to a queue nobody has worked since the campaign ended. The agent reports that the request is in, the owner waits for a call that never comes, and two weeks later signs with another lender.

Some agents will make it through your site anyway, simply because the paths they find happen to work. That is agent experience by default. It works until the next redesign moves a button or renames a field, and in the meantime nobody inside the firm knows which agents are using the site, which paths they take or how often they fail.

Regulation raises the stakes on every one of these choices. When an agent opens an account, changes an address or moves money for a customer, the firm still needs a clear record of what was done and on whose behalf. A deliberate path for agents makes that easier, because it leaves a cleaner record than an agent improvising its way through old forms. Letting every agent in invites fraud, because the same tools that help a real customer also help someone pretending to be one. Blocking every agent pushes customers toward competitors who serve them.

Blocking can still be the right design. In September, Amazon blocked Meta's Muse from shopping on its site, saying it had never authorized the access and that Muse did not identify itself as an AI agent. Whatever you think of that call, it was a decision, made on purpose by a company that understood what was at stake. Most financial firms haven't made that decision in either direction. Their systems let some agents through, stop others and quietly fail the rest, and nobody chose which. Whether to serve agents, block them or allow them only certain tasks is now a question every firm has to answer deliberately.

A self-check you can run this week

You don't need a new tool or an outside vendor to see where you stand. An afternoon and a spreadsheet are enough for a first pass.

Part one: what the assistants say

1. Write down ten questions your customers, and the brokers or advisors who serve them, actually ask. Your service center and sales teams can tell you which ones come up most often. Include category questions ("Which carriers offer term life without a medical exam?") alongside questions about you ("How do I roll my 401(k) over to you?" or "What is your overdraft fee?").
2. Put each question to the assistants your customers are most likely to use: Claude, Grok, ChatGPT, Gemini and Copilot. Phrase each one in a customer's words, including the shorthand people actually use.
3. For each answer, record whether you appear, whether the facts are right, which source the assistant seems to rely on, and which competitors appear in your place.
4. Save the answers and repeat the exercise every quarter. Results shift as the assistants update and as your own content changes, so a single snapshot goes stale quickly.

Part two: what an agent can do

1. Pick the tasks customers most want to complete with you, such as getting a quote, updating an address or starting an application.
2. Ask an assistant with an agent or browsing mode to complete each task on your public site, the same way a customer would ask it to.
3. Note where it stalls and where it has to guess. Check whether it found an old form or page instead of the current one, and whether it could tell for certain that the task was complete.

Every gap you find marks a place where your firm is running on default. Together, those gaps become your list of what to fix first.

Moving from default to design

What assistants say about you

• Stale content. Old fee schedules, retired product pages and outdated claims guides tend to stay online for years after anyone stops maintaining them, and assistants find them as easily as your current pages. Retire them, redirect them, or mark them clearly as replaced so there is no doubt which version is current.
• Facts that disagree. When your fees, rates, eligibility rules or contact numbers differ from one page to the next, an assistant has no way of knowing which is right, and it may pick the wrong one. Decide on one source of truth for each of these facts and make every page match it.
• Pages AI can't read. Many AI crawlers read the raw text of a page and don't run the scripts that build most modern websites. Key facts also tend to sit in places they can't reach, such as PDFs, logged-in help centers and rate calculators. Put the facts customers ask about most in plain text on a public page.
• Who you let in. Check your robots.txt file and your content delivery network settings, because some providers now block AI crawlers by default. Blocking can be a reasonable choice for a regulated firm, as long as someone made it deliberately and understands what it costs in visibility.

How agents act on your systems

• Choose the tasks. Decide which tasks a customer's agent may complete with you, such as getting a quote or checking a claim status, and which should stay with a person because of the risk or judgment involved.
• Publish the steps. For the tasks customers need most often, write a plain public guide: what is needed, in what order, which form or page to use, and what confirms the task is done. It helps people as much as agents, and it takes calls off your service center. Some firms also publish a short machine-readable map of their site that points agents to the right pages. The llms.txt file is one emerging convention for this, though not every assistant reads it yet.
• Retire old paths. Find the old campaign pages, legacy forms and endpoints that still accept submissions, including the ones search engines and sitemaps still list. Switch them off or redirect them to the current process.
• Make completion unmistakable. Every completed task should return a plain confirmation and a reference number, and a repeated submission should be recognized as a repeat instead of creating a duplicate.
• Know who is acting. Your fraud and bot controls should be able to tell a customer's agent from an attacker without locking the real customer out, and an agent acting for a verified customer should get only the permissions its task requires.
• Readable forms and data. Clear labels, predictable steps and plain error messages help agents and people alike, and schema.org markup tells an agent which number is a fee, a rate or a deadline. For high-value tasks like quotes, a published interface lets an agent skip the screen entirely, and emerging standards such as the Model Context Protocol make that easier to build.
• Handoffs that keep the work. When an agent reaches a step only the customer can complete, carry everything it entered into the customer's session so nobody has to start over.

Who owns the front door?

In most firms, marketing owns the website and the customer experience team owns the journey. Compliance owns the accuracy of what the firm says, and IT owns the portal and the systems behind it. The AI front door touches all four and belongs to none of them, which is exactly how it ends up running on default.

Name one owner anyway. The work will still span those teams, but one accountable executive should run the quarterly check, have the authority to get sources and paths fixed across them, and report to leadership on what the assistants are saying and what agents can and can't do.

Every firm already has an agent experience, whether or not anyone planned it. It is easy to miss, because the conversations happen where you can't see them. That is the reason to start looking now, and to decide what the experience should be before the default decides for you.

Frequently asked questions

What is agent experience (AX)?

Agent experience is how easily an AI assistant or agent can find, understand and correctly act on your firm's information and services. The term comes from software development. In financial services, it is customer experience in a new form, delivered through the customer's assistant.

What is the difference between agent experience by default and by design?

By default, assistants and agents work with whatever they happen to find on your site and elsewhere, and nobody checks the result. By design, you decide what they should find, which tasks they can complete, and how you monitor and correct what they say and do.

How can I check what AI assistants say about my company?

Ask the main assistants the questions your customers ask, both about your category and about your firm, and log the answers. Then ask an assistant with an agent mode to complete a few common tasks on your site. Repeat both every quarter.

What makes a website ready for AI agents?

Key facts in plain text on public pages, published steps for the tasks customers need most, old forms and pages retired, clear confirmations that prevent duplicates, fraud controls that recognize legitimate agents, and handoffs that keep the agent's work when the customer takes over.

Can we control what AI assistants say about us?

Not directly. You can control the sources they read by keeping your public content current, consistent and readable, and by making deliberate choices about which AI crawlers can reach your site.

Is this a marketing issue or a governance issue?

Both. Visibility is a marketing question. Accuracy and agent access, especially around fees, coverage, eligibility and money movement, are governance questions, and in a regulated firm they need a named owner.