Blog Design

Your website has a second audience: software acting for customers

AI can take some of the work out of using the web. But if we need an assistant to find out what we’re agreeing to, we should ask what the website is doing in the first place.

Suchetana Bauri13 min read

A web page with a form and a Confirm button, a person on one side and an assistant on the other; the assistant’s dotted path stops at a review screen before the button

You want to know what it costs to move a Saturday booking. On your phone, you tap ‘Manage booking’ and find a sign-in screen. The policy is elsewhere. You go back, find it and read that a fee ‘may apply’. All you wanted was to know whether moving the booking would cost more than keeping it. By the time you put the phone down, you feel faintly unreasonable for having asked.

I use AI for precisely this sort of thing. I want it to find the paragraph I missed, explain what it means and, if I ask, get me as far as the point where I can make the change. I also use it to learn: sometimes the most useful answer is an explanation of which question I should have asked in the first place. What I don’t want is for it to read ‘may apply’ and come back with a confident price. Or, worse, to change the booking and tell me afterwards.

This is the case for designing websites for a second audience: software acting for customers. Not because the machine deserves a better website than the person does, but because what it can find, infer and do now shapes what the person gets to know and decide.

The website was already asking us to do the work

The booking company may have perfectly ordinary reasons for putting the fee behind a sign-in. The amount might depend on the fare, the date or how late you leave it. What the public page could tell you is where the actual price will appear and whether you can see it before committing. “Sign in to see your change fee. Nothing will be changed until you confirm.” That describes a sequence you can trust. ‘Manage booking’ gives you somewhere to click.

A site can be built to make a sale easy without anyone taking equal responsibility for what happens after it. The team that owns checkout can measure a completed purchase. The customer trying to establish what a change will cost may register as a visit to a help page, a chat session or nothing at all. The company sees activity in three places. The customer experiences one unfinished task.

The UK’s Competition and Markets Authority uses the term ‘sludge’ for excessive or unjustified friction, such as making a subscription easy to buy and awkward to cancel.1 That is not the same as saying every clumsy journey is deliberate. Perhaps the policy came out of one team, the account area out of another, and the booking software was bought years ago. Nobody needs to have planned the obstruction for the business to feel less urgency about fixing it than the person trying to leave.

After a few rounds of signing in and re-entering a reference number, giving up can feel like a personal decision rather than an outcome the interface helped produce. It would be interesting to know how many dashboards count that as a customer choosing to keep their booking.

An assistant can spare us some of that work. That is a real gain, and I would rather not romanticise twenty minutes spent looking for a cancellation clause. But if the information is contradictory or the outcome is hard to confirm, delegation moves the work around. We used to check the policy. Now we check what the assistant says the policy says.

Being found is not the same as getting something done

There are three separate tests. Can the software find the right page? Can it understand what the page actually says? Can it take a permitted step and establish what happened? A site can pass the first and fail the next two. The booking policy may be indexed, the assistant may quote ‘a fee may apply’ correctly, and neither may get you any closer to knowing what Saturday will cost.

Some people may never reach the website at all. In Pew Research Center’s study of 900 US adults, traditional Google results received a click on 8% of visits with an AI summary and 15% of visits without one, though Pew reconstructed results after the searches and could not establish that summaries caused the difference.2 When someone gets enough of an answer to stop looking, did they get the part that mattered?

Retail referral data describes a different part of the journey. Adobe Digital Insights reported that visits from AI referrals to US retail sites grew 71% year on year in September 2026, counting people clicking links from AI sources, not software completing purchases on their behalf.3 It doesn’t tell a retailer whether an assistant can notice a price change or stop before payment.

Researchers have tried to measure that other thing. In Online-Mind2Web’s test of six agents on 300 tasks across 136 live websites, the best of the six completed 61.3% of tasks by human judgement.4 That isn’t a forecast for today’s assistants. The lesson I’d keep is to check the outcome: if an assistant says “I’ve sent it”, look for the message.

As with measuring AI adoption at work, activity is not the same as a better outcome. A visitor who finds the fee and chooses to keep their booking has completed the task they came to do. A sales metric alone cannot distinguish that informed decision from someone giving up before they saw the price.

One website, one account of the facts

I came across a small example on my own site. When I checked on 20 November 2026, my homepage marked the Google UX Design Professional Certificate as ‘In progress’, while the machine-facing summary listed it among certificates without that qualifier. I had in fact completed it, so the summary was right and the homepage was out of date. I’ve since made them match. It showed how easily an account of the same person can change when two words go missing, or stay behind.

The fix is to check the status and use the same wording in both places. A customer comparing two products, two training programmes or two insurance policies needs the same courtesy. Put the terms that would change their decision in visible text, near the decision. Link evidence beside the claim. If there’s no fixed price, explain how it is quoted. ‘It depends’ is useful when you say what it depends on.

The qualification is often the part that feels easiest to cut. ‘In progress’ is two words; ‘only on weekdays’ is three. Neither does much for the rhythm of a product description, and both can change the answer entirely. I wrote about this in In an AI-saturated market, proof is part of the brand. Evidence is more useful when the conditions that make a claim true travel with it. That matters especially when the customer sees one sentence from an assistant rather than the whole page.

Google Search Central says that important information should be available as text and that structured data should match what visitors can see. Google’s guidance also says you don’t need llms.txt for its generative AI features.5 If you maintain an extra summary for another service, ask who will use it and who will keep it true.

Nor does designing for software mean turning every website into a list of specifications. A film school ought to feel different from a building society. Photography, type and pacing help tell visitors what kind of place they have arrived at. But they should not be the only cues that tell you where to go, or the only place a material condition appears. A richly designed page can still have one plain sentence that survives being read aloud, copied into a message or quoted away from the photograph. That is an editorial decision as much as a technical one.

Make the stopping point part of the design

A button that says ‘Continue’ might take you to a review screen or commit you to something. If you have to click to find out, that is a poor place to put a decision. I made a related argument in The interface owes you an explanation: the more a system can do, the clearer it should be about what it is about to do. Tell the person what happens next. Give form fields visible labels, say what needs correcting when something fails and make confirmation recognisable.

These are accessibility concerns first. The Web Content Accessibility Guidelines cover keyboard use, labels, errors and status messages.6 An agent filling a form without complaint is not evidence that a disabled visitor can use it. Equally, a person who finds a workaround has not proved the design is clear. It is worth listening for ‘I assumed’, ‘I went back’ or ‘I just tried it’. Those words tell you where the visitor supplied an explanation the website didn’t.

Then decide where delegation ends. Let the assistant find a slot, compare terms or prepare an enquiry. Before a booking, payment or message is committed, show the person what will happen and let them approve it. If verification is required, hand control back without making the person start again unnecessarily. And if nobody can establish whether a submission went through, say so. A second attempt may be the first time the customer sees an error, or the second time the business receives the order.

Not all friction is a defect. The sign-in may protect the booking; the form forgetting your chosen date adds work without protecting anyone. A good handover keeps the check, preserves the work where it can do so safely and makes clear what information will be shared.

There are tools worth trying. Chrome’s experimental Lighthouse checks can flag missing control names and unstable layouts; they cannot tell you whether your customer got the right booking.7 WebMCP could let compatible assistants call a named site function rather than guess which control to click, but it remains a proposal in a Chrome origin trial, not a World Wide Web Consortium standard.8 I’d try it on a read-only task such as checking availability before giving it any means to commit a payment.

What changing a railway ticket asks of you

On IRCTC’s beta train-search page, the Indian Railway Catering and Tourism Corporation puts the station and date fields near the top, with shortcuts for checking a reservation or tracking a train beside them. Below, a family walks towards a station next to a phone the size of a building. The banner says your journey is ‘now just a tap away’. It is an advertisement for the app, of course, but it captures how a booking service likes to picture the journey: everyone is still setting out.

In October 2025, the railway minister told NDTV that passengers would be able to change the date of a confirmed e-ticket online from January 2026 without a cancellation fee, subject to availability and any higher fare. As of 20 November, IRCTC’s public FAQ still described cancelling the original ticket and booking another; I couldn’t verify whether the announced option was live for signed-in passengers.9

Either way, ‘move my journey to Sunday’ does not mean ‘give up my Saturday seat before I know whether there’s one on Sunday’. Before anyone commits, the service should say whether the new seat is confirmed, what extra fare is due and when the original seat is released. An assistant can help compare the options. It should not have to infer the price of getting it wrong.

The interface itself has been getting attention. The Ministry of Railways says the redesigned beta launched in July 2026 shows availability across travel classes and reduces the number of booking steps.10 Those are worthwhile changes. In November, after passengers reported trouble accessing the service, IRCTC said the portal had remained operational but some logins were delayed under heavy festive demand.11 Both can be true: a clearer interface can make a real difference, and the service can still be unavailable to a person at the moment they need it. ‘Operational’ is a useful word until you are the one waiting to get in.

There’s a useful contrast in the GOV.UK page for changing a driving test. Before you sign in, it tells you when you may have to pay again, which days count towards the notice period and that you must change a car test yourself.12 That doesn’t make the rule more generous. It makes the rule easier to understand, including where an assistant’s job ends.

A human-and-agent readiness audit you can run

Pick one journey customers already care about. For my site, I’d use the route from the Design services page to a relevant example of work, then to an enquiry ready for review. For a rail booking, use the alternatives, the cancellation terms and the point before either ticket is touched.

TestBriefWhat counts as success
FindFind the service for a team whose customers struggle to complete a task on its website.The person or assistant reaches the relevant page and correctly describes the offer. For search discovery, don’t give the assistant the URL.
VerifyFind a relevant example. Tell me what was delivered and what the evidence does not establish.The account matches the case study and points back to it. A reported result doesn’t become a promise.
Prepare and stopDraft an enquiry, find where it should go and stop before sending it.The draft and destination are correct, the next step is clear and nothing is sent. Stopping is a success, not an incomplete task.

Write the expected answer or end state before testing. Give the same brief to people and to the relevant software clients. Include customers with relevant access needs where possible, and record keyboard and screen-reader checks separately. Try a few fresh agent sessions and note the client, date and result of each. ‘Two correct answers out of three, one with an invented qualification’ tells you more than an aggregate readiness score. It is a starting point for diagnosis, not a dependable estimate of how often the error will happen.

Keep live work read-only unless you have authorisation for more. Use staging and dummy data to test submissions, failed fields or uncertain outcomes, and check the resulting record rather than relying on a reassuring message. In the findings log, keep the human, search and browser-agent results separate. For each failure, write down what you expected, what happened, the evidence, who owns the fix and when you’ll retest. Fix a wrong term, a blocked route or an unauthorised action before worrying about any optional integration.

Run one case in which the correct outcome is not to proceed. Tell the assistant there is a firm spending limit and let the staging journey present a higher total at review. Check whether it notices, shows the new amount and stops. A booking is easy to demonstrate when everything goes to plan. If you want to learn what the design does for people, watch what happens when the plan changes. That is why I’d rather start with a broken workflow than with a list of things an agent could be made to do.

When you look at the results, distinguish a missing fact from a missing decision. If the change fee is decided but poorly explained, the content needs attention. If nobody has settled when a customer should see it, that decision belongs to the service owner. If the assistant ignores a clear restriction that a visitor understood, record a failure of that client rather than rewriting the whole site. This is close to a distinction I’ve made about communicating AI change at work: sometimes people need clearer instructions; sometimes they are still waiting for someone to decide what the instruction is.

The useful result is a short list of changes you can verify. Perhaps the public policy needs one sentence saying when the price appears. Make the change and run the same brief again. If people still get stuck, the audit has at least spared you the satisfaction of fixing the wrong thing.

What we shouldn’t delegate

A public website is one place where two people can look at the same sentence. A friend can send you the policy; a support worker can point to the clause; a customer can show a business what it said yesterday. A private answer from an assistant may be more convenient and more relevant to your circumstances. If it can’t point back to something you can inspect, though, it is harder to tell whether a disagreement is about the terms, the way they were explained or the answer generated for you. I want the convenience of the private conversation without losing the usefulness of the public page.

And I want to be able to do without it. That is partly about people who don’t want another service involved in a transaction. It is also about the pleasure of using something that explains itself well enough to let you make up your own mind. An assistant that finds the actual change fee and shows its workings can help me learn how the system works. An assistant that has to be summoned for every ordinary decision teaches me something too, just not something especially flattering about the system.

That is the cultural cost worth watching. Not that a machine sometimes reads the web for us, but that understanding the terms of an ordinary transaction starts to feel like a specialist skill you have to bring with you. A useful assistant can take some of the work out of finding an answer. It shouldn’t become the price of getting one.

Notes

  1. Competition and Markets Authority, Evidence review of online choice architecture and consumer and competition harm, 2022. See the section on ‘sludge’ for the distinction between justifiable and excessive friction.
  2. Athena Chapekis and Anna Lieb, Google users are less likely to click on links when an AI summary appears in the results, Pew Research Center, 22 July 2025. The study used March browsing records and gathered search results in April.
  3. Adobe Digital Insights, AI traffic drives 83% higher revenue per visit, November 2026. The article uses Adobe’s September US retail referral figures, not its holiday forecast.
  4. An illusion of progress? Assessing the current state of web agents, arXiv. The Online-Mind2Web results cited here are human judgements of agent performance, not a human completion rate.
  5. Google Search Central, AI features and your website and Google’s guide to optimising for generative AI features on Google Search. The guidance applies to Google Search, not every AI service.
  6. World Wide Web Consortium, Web Content Accessibility Guidelines (WCAG) 2.2. Relevant criteria cover keyboard use, labels and instructions, error identification, and status messages.
  7. Chrome for Developers, Lighthouse agentic browsing scoring. The experimental category reports diagnostic checks rather than a weighted score or a measure of task completion.
  8. Chrome for Developers, WebMCP; Web Machine Learning Community Group, WebMCP draft. The browser trial and draft status are as checked on 20 November 2026.
  9. Akhilesh Sharma, In a first, Railways to allow changing travel dates for booked tickets, NDTV, 7 October 2025. The minister set out the planned start and conditions. IRCTC’s beta train-search page and e-ticket FAQ were checked on 20 November 2026. See question five of the FAQ. No signed-in booking was inspected.
  10. Ministry of Railways, Four decades of passenger reservation in Indian Railways, Press Information Bureau, 17 November 2026. The release distinguishes the July website beta from planned integration with the next-generation reservation system.
  11. Anish Mondal, Was Indian Railways’ ticket booking website down? IRCTC clarifies, the Indian Express, 10 November 2026. The article distinguishes reported access difficulties from IRCTC’s statement that the portal remained operational.
  12. GOV.UK, Change your driving test appointment. The example is about how the public page explains notice, charges and who can act, not an assessment of the signed-in service.