Chatbots for Booking Questions

Two rules settle most of this before any question of software. Say that it is a bot. And never make it the only way to reach a person.

Both are written down somewhere, one in a state statute and one in a federal regulator's assessment of what goes wrong, and between them they describe a tool that is either genuinely useful to a guide or actively destructive, with very little in between.

The rule that says disclose

California legislated on this in 2018, and the chapter is short enough to read in a minute. The operative section makes it unlawful to use a bot to communicate or interact with another person in California online, with intent to mislead that person about its artificial identity, for the purpose of knowingly deceiving them about the content of the communication in order to incentivise a purchase or sale of goods or services in a commercial transaction, or to influence a vote.

Then the sentence that turns a prohibition into a design instruction: a person using a bot is not liable under the section if the person discloses that it is a bot.

The disclosure itself has a standard attached. It must be clear, conspicuous, and reasonably designed to inform the people the bot communicates with that it is a bot. Not buried in terms, not in grey text under the input box, and not phrased so that somebody could reasonably miss it.

The chapter sits in the part of the code dealing with representations to the public, was added in 2018 and became operative in July 2019, which means this has been settled law for years while most operators have never heard of it. Everything else a guide does to get found is at the getting-booked hub.

Galveston Bay, TX

Why that is easier than it sounds

The safe harbour is unusually generous. It does not require accuracy, quality, or any particular capability. It requires honesty about what the thing is.

So the compliance work is one line of text at the top of the widget saying this is an automated assistant, and a person is reachable at a number you then give. That is the whole exercise, and it costs nothing.

What the statute is aimed at is the opposite case, where the artificial identity is the point, and somebody is led to believe they are negotiating with a person while being steered toward a transaction. A guiding business has no reason to be anywhere near that, and the disclosure is also just better business, since a client who discovers mid-conversation that they have been talking to software feels tricked in a way that no booking survives.

The failure a federal regulator wrote up

The consumer finance regulator published a report on chatbots in June 2023, and although banking is a long way from guiding, the diagnosis transfers exactly.

Its framing is that working with customers to resolve a problem or answer a question is an essential function and the basis of relationship banking, and that customers rightfully expect timely, straightforward answers regardless of the processes or technologies used. Regardless of the technology is the operative phrase: the customer's expectation does not adjust downward because you installed something.

It then names the commercial pressure honestly. As sectors integrate these solutions into customer service, there will likely be strong financial incentives to substitute away from support offered in person, over the phone, and through live chat.

And the finding worth pinning above the desk: deficient chatbots that prevent access to live, human support can lead to law violations, diminished service, and other harms.

Prevent access is the phrase to test against

Run your own setup against those three words. A widget that answers a question and offers a phone number does not prevent access. A widget that captures a name and address before revealing anything, or that has no exit to a human at all, does.

The distinction matters more in guiding than in banking, because the thing being bought is a person's judgement for a day. Nobody books a guided trip because the website was efficient. They book because they believed the person on the other end knew the water and would look after them, and a system designed to keep that person away from them is working against the sale it was installed to produce.

The practical rule is that the escape hatch should be visible in the first message, not offered after three failed attempts.

The regulation being written now

State legislatures are moving, and the shape of what is coming is visible in the codes already amended. Utah has an Artificial Intelligence Policy Act sitting in its commerce and trade title, which creates an Office of Artificial Intelligence Policy with a director appointed to run it.

Its duties are worth reading because they describe a state building a rulebook rather than issuing one. The office runs a learning laboratory programme, consults businesses and other stakeholders about potential regulatory proposals, and makes rules establishing required participant disclosures to consumers, reporting requirements, and criteria for extending participation. It reports annually before 30 November to a legislative committee on the learning agenda, the outcomes, any regulatory mitigation agreements, and recommended legislation.

The learning laboratory itself exists to analyse the risks, benefits, impacts and policy implications of these technologies to inform the state framework, to encourage responsible deployment, to evaluate the effectiveness and viability of current or proposed regulation with the companies themselves, and to produce findings and recommendations.

Read across the two states and the direction is clear enough to plan around. Disclosure is the settled obligation; everything else is being worked out in public, and an operator whose bot already says what it is has nothing to redo when the rest arrives.

What it can genuinely do for a guide

The useful case is narrow and real: the same six questions, asked at hours when nobody is going to answer.

Most enquiries a guiding business receives arrive outside working hours, and most of them are not really enquiries. They are checks. How long is the trip, what is included, do I need a licence, where do we meet, what happens if the weather turns, how much is it. A system that answers those at nine in the evening removes the reason somebody drifts to the next operator on the list, which is the same enquiry-speed problem examined in the direct messages piece.

It is also a decent capture point. Somebody who asks a question at midnight and leaves a number is a booking you would otherwise have lost, provided the handover is honest about the fact that a person will reply in the morning.

And it enforces a discipline worth having anyway. Writing the answers a bot will give forces you to decide what the answers actually are, which is the same work as writing the pages, covered in the water and city pages piece. If the answers are good enough, the page may remove the need for the widget entirely.

The questions it must never answer

Anything about conditions. A system stating what the fishing will be like has made a promise on your behalf that you would never make yourself, and clients hold guides to it.

Anything about licensing, permits or rules. Those vary by state, by water, by species and by season, they change, and a confidently wrong answer sends somebody onto the water without what they need. That is a question for the authority that issues the licence, and the honest response from a widget is to say so and hand over.

Anything about price outside the standard list. Group trips, unusual durations, travel and multi-day arrangements are negotiations, and a machine quoting into one loses money on the good version and the trip on the bad one.

And anything that has gone wrong. A complaint routed to software is a complaint that becomes a review, which is a far more expensive problem than the one you were trying to automate, as the handling in the negative reviews piece makes clear.

Writing the six answers

This is the actual work, and it is worth doing on paper before any software is chosen, because the answers outlive whichever tool is fashionable this year.

Trip length and what is included wants a number and a list, not a paragraph. How many hours, what time you start, whether tackle, bait, licences, ice and lunch are in the price, and what the client brings. Guides consistently under-specify this and then spend the season answering it one message at a time.

Price wants the standard rates and an honest statement that anything unusual is a conversation. Publishing a range with a reason for the range is better than publishing nothing, since a widget that will not talk about money simply relocates the question rather than answering it.

The meeting point wants an address and a sentence about parking, which is the single most frequently asked operational question in the trade and the easiest to answer permanently.

Seasonality wants months and species, stated as what actually runs rather than what is theoretically possible, and it should be the one answer you revisit every year.

The weather position wants your real policy in your own words, including who decides and what happens to a deposit, and the version that works is the one already written down for the cancellation conversation in the cancellation scripts piece.

And the licence question wants a pointer rather than an answer, naming the authority and linking to it, since that is the one where being helpful and being wrong are the same action. The state-by-state position is collected at the licence collection.

Joey Flats, FL

Reading the transcripts

The output nobody uses is the most valuable thing the tool produces, and it takes ten minutes a week.

Every question the system could not answer is a hole in your site. Every question asked in words you would not have used is a phrase real customers say, which is worth more than any keyword tool because it came from somebody who was about to book, and which feeds directly into the work described in the local search piece. And every conversation that ended without a name and number is a specific failure you can look at.

Patterns emerge faster than people expect. A dozen enquiries asking whether children are welcome means the site does not say. A run of questions about a species you barely target means your pages are attracting the wrong reader, which is a positioning problem rather than a chat problem.

Keep a running note of what the transcripts tell you and fix the site rather than the bot. The measure of a good widget in a guiding business is that it gets asked less over time, because the answers moved to where everybody can see them.

What experienced guides do differently

They write the answers themselves rather than letting a system generate them, because the six answers are the operation's actual policy and putting them in writing is useful independently of the widget.

They put the disclosure in the first line and the phone number beside it, treating the number as the product rather than the fallback, in the same spirit as the pre-trip messaging piece, which satisfies the statute and removes the frustration in the same stroke.

They read the transcripts weekly. Every question a bot could not answer is a question missing from the site, and a season of transcripts is the cheapest content research a guiding business will ever get.

And they turn it off in the off-season if it is not earning its place, since a widget answering nobody is a subscription with a spelling mistake in it.

Common mistakes

No disclosure, which is the one with a statute attached and the cheapest thing in the world to fix.

Making it the only route. If the phone number is not visible until somebody has answered three questions, the system is a gate rather than a service.

Letting it answer questions about conditions, regulations or unusual pricing, which are the three places a wrong answer costs real money.

Installing one before the pages answer the same questions, which puts a conversational interface in front of an information problem. And leaving it unread, so the record of what people actually ask sits in a dashboard nobody opens.

What surprises people

That a state has legislated on this specifically, that it did so in 2018, and that the fix is a sentence.

That the safe harbour asks for disclosure rather than accuracy. The law's concern is the artificial identity, not the quality of the answers.

That a federal regulator has written down the failure mode in plain words, and that preventing access to human support is described as capable of leading to law violations rather than merely to annoyance.

And that most enquiries arrive when nobody is answering, which is the entire commercial case for the tool and is usually discovered only after somebody looks at the timestamps.

When not to have one

When you answer the phone reliably. A guide who picks up has already beaten every widget on the internet, and adding one subtracts.

When the volume is low. A handful of enquiries a week is a personal conversation, and automating it removes the only advantage a small operation has over a booking platform.

When the site does not yet answer the six questions in plain text, because that is the cheaper fix and it also helps the people who never open a chat window.

And when nobody will read the transcripts, since an unread system quietly drifts away from what the business actually does, which is the same reason automated replies need a human review loop as described in the automated replies piece.

The version worth running

One line of disclosure. Six answers you wrote. A visible phone number from the first message. A capture that takes a name, a number and some dates, and says plainly that a person will reply. A weekly read of what people asked. And a decision each season about whether it is still earning its place.

That version is compliant, useful and cheap, and it takes an afternoon. Everything more elaborate than it is an attempt to make software do the part of this job that the client is paying a person for, and it fails in proportion to how well it is built, because the better it is at holding the conversation the longer it takes somebody to reach you.

The same restraint applies to the wider question of where automation belongs in a one-boat operation, which is worth deciding once rather than tool by tool, and is taken up in the piece on generated content.