There is a chat bubble in the corner of our site, and behind it an AI assistant. It spends its days answering how much, how long, and whether maintenance is included.
Here is what it actually does, what it cannot do, and what went wrong along the way — because marketing for this category tends to sit a long way from reality.
What it genuinely covers
The main value is not replacing a person. It is covering the hours when nobody is there.
Chinese-owned businesses in New Zealand share a pattern: customers browse in the evening or at weekends, because they work during the day. Nobody answers messages then. The old outcome was that a question went unanswered and by the next morning the person had moved on.
Now they get an answer immediately. Prices, timelines, what is included, who owns the code — those answers are fixed, and AI gives them as accurately as we would.
The second value is lowering the cost of asking. Plenty of people will not send an email or add a contact just to ask a price, because they expect to be pursued afterwards. A chat box carries none of that. Those people used to leave without trace; now at least one exchange happens.
What it cannot cover
Anything requiring judgement. “What would my project cost” genuinely depends on page count, feature complexity, whether content exists, and whether anything has to integrate. AI can give a range and not a quote. It is instructed to say that a specific figure needs a conversation rather than to invent one.
Anything constituting a commitment. Delivery dates, whether a particular feature is possible, whether a payment arrangement is acceptable. Once AI says any of these, the customer treats it as a promise, so those go to a person.
Emotion. An already unhappy customer wants a machine least of all. An assistant appearing at that moment makes it worse.
It quoted the wrong prices — to everyone
Worth writing down.
We raised prices: single-page from 799 to 1,499, multi-page from 1,499 to 2,499. The website was updated, the quote template was updated, and nobody thought about the assistant’s system prompt.
So for a period, every visitor who asked about price received the old figures — around forty percent under actual. It answered confidently, because as far as it was concerned those were the facts.
Nothing raised an error. Every other part of the site was correct. One chat box was quietly giving out wrong information, and it surfaced only because someone happened to ask it in passing.
The lesson is specific: an assistant’s knowledge is a separate copy of your content and does not update when your site does. Any change to prices, policies or scope has to be on a checklist. “Change a price” is now defined here as a four-part action: site copy, structured data, llms.txt, and the assistant’s prompt. Miss one and the above happens again.
Two other traps
It replies in markdown. The chat panel displays plain text and renders no formatting, but the model habitually produces bold, headings and bullets, so users see raw asterisks and hashes.
We wrote “do not use markdown” into the prompt. No effect. We wrote it more forcefully. Still no effect. Constraining output format by instruction is inherently unreliable — a model complies most of the time and then does not.
The fix was stripping the markup in code before the reply is returned. Deterministic, and independent of whether the model remembers. Anything code can guarantee should not depend on a model remembering.
Page caching fought with a security token. Technical but representative: the page carried a time-limited token used to call the API. Once we cached pages, that token expired inside the cached copy, and the assistant stopped working entirely for anyone served from cache.
The symptom is that new visitors are fine and visitors served a cached page get nothing when they click. This class of problem is nearly untestable by hand, because your own test usually happens to trigger a fresh page.
Deciding whether it is worth it
Three questions:
Do your most common questions have fixed answers? If most are price, location, hours, or whether you do X, an assistant will hit a high proportion of them. If most require looking at specifics, it can do less.
How quickly do you reply now? If you already answer within ten minutes, the marginal value is small. If replies often take a day, it fills a real gap.
How costly is a wrong answer? A mistaken price can be explained. In medical, legal or immigration contexts, an incorrect statement is a genuine problem. Those industries should either skip it or constrain it tightly — process questions only, anything case-specific to a person.
On “replacing your front desk”
We do not describe it that way. What it replaces is not a receptionist but conversations that would never have happened: someone looking at 10pm, someone unwilling to hand over contact details, someone who only wants a price before deciding whether to engage.
Those people were not going to contact you and were not going to become customers. Some of them now stay. That increment is real, and it is a different thing from saving a salary.
If someone tells you an AI assistant lets you employ one fewer person, they are selling something.
Ours is open on this site — ask it about price, about timelines, ask it something awkward and watch where it becomes evasive. It is one of the features we can build, but trying it beats reading about it.