How long AI chat takes to set up

The install is about forty minutes. The content is the project, and the drafting is now largely automatable. The review is not.

Two clocks run on a chat project, and vendors quote the wrong one. The first is the install: getting the assistant onto your site, connected to your catalog, reachable on the number customers already message you on. That has taken minutes for years and it still does. The second is the content: everything the assistant is supposed to know about your shipping, your returns, your stock, and the things you would rather it never said on its own. That second clock is the project, and it is why every demo you have watched ran fast. Somebody wrote the content first, off camera. What has actually changed is that the drafting can now be done by a machine, either a crawler reading your public site or an assistant like Claude, ChatGPT or Cursor connected over MCP writing the knowledge cards directly. That moves the whole bottleneck to review, which is the one step nobody can do for you, and it is why the honest answer to how long this takes is a count of minutes followed by a count of days, with the two numbers measuring completely different things.

The install clock, itemized

Here is what the install actually consists of, with the minutes it takes when the person doing it has the passwords already open in front of them.

Call it forty minutes end to end. The number is honest and it is also close to useless as a planning figure, because it assumes one thing that is rarely true.

What turns forty minutes into three days

It assumes you have access. The embed snippet needs whoever can edit the theme or the tag manager. The Shopify connection needs someone with permission to install an app on the store, which in a lot of shops is one named person and not you. The WhatsApp number is usually held by whoever set it up: an agency, a partner, sometimes someone who left. The test link needs nobody, which is why it is the first thing to do. The work is minutes. The waiting is days, and the waiting is what people are describing when they say setup took a week.

So do the access check before you do anything else. Write the four items down and put a name next to each one. If you cannot put a name next to the WhatsApp number, start there, because it has the longest tail.

Two ways to get the content drafted

The first route is the crawler. It reads your public site and produces a prefill: a first set of knowledge cards drawn from your shipping page, your returns policy, your FAQ, your sizing notes. It runs alongside the catalog import, so at the end of it you have both the products and a first pass at the policies. This route is good at everything you have already written down somewhere public. It is broad rather than deep, it costs you nothing but the time to start it, and it is the fastest way to find out how much of your own site is out of date. Half an hour of your attention, most of it waiting.

The second route is MCP. You connect Claude, ChatGPT or Cursor and that assistant writes the knowledge cards with you, in conversation, filing them as it goes. This route is good at the opposite thing: what is not on your site. The exception you make for repeat customers. The courier you stopped using in March. The answer you give on the phone when someone asks whether the fabric is machine washable. You talk, it writes, and it puts the result somewhere you can edit. The cost here is your own attention rather than waiting: a session that gets a small shop most of the way takes 45 to 90 minutes, and you have to be present for all of it, because the only source for that material is you.

Most shops end up using both, and should. Crawl for the floor, then spend a session over MCP filling what the crawl could not know. Neither route produces a finished assistant. Both produce a draft. It is worth knowing that the assistant answers only from your imported catalog and your own cards, not from general web knowledge, so a gap shows up as a hedge or a handoff rather than an invention. That is the right failure mode, but it is still a failure you want to find before a customer does.

Review is the step that cannot be automated

A crawler is faithful, and that is exactly the problem. If your policy page says returns are 14 days and an old landing page in the footer says 30, you now have both, written down with equal confidence. If a policy is worded badly on your website, it will be worded badly in the assistant, and now it will be worded badly straight into the face of somebody who was about to buy, repeatedly, without anybody on your side hearing it happen.

A drafting assistant has the opposite failure. It is fluent, and fluent things fill gaps. Ask it to write a card about your warranty when your About page only implies you stand behind the product, and you will get a warranty policy that reads well and that you never agreed to. It is not lying. It is doing what you asked with the material you gave it. The tell is that the invented cards are usually the best written ones in the set, because nothing in your source material was slowing them down.

So read every drafted card once. For a shop with ordinary policies that is about an hour. Check numbers and thresholds hardest: the free shipping minimum, the returns window, the restocking fee, the lead time, the countries you actually ship to. Those are what customers ask about and those are what a wrong answer costs you money on. That hour is the single thing that decides whether week two is calm or spent apologizing.

One kind of finding deserves a decision rather than a quick fix. When the draft holds two versions of the same policy, because your returns page and a landing page in the footer disagree, you have not found a content bug. You have found a question your shop has never actually settled, and picking whichever one is nearest so you can finish the setup is how the wrong version ends up being quoted to customers for a year. Stop there, decide which is true, change it on the website too, and only then fix the card. Ten minutes, and it is the only part of this whole process that improves anything outside the chat window.

A realistic first week

Assume roughly a working day of your attention, spread thin across five days rather than blocked out in one. Spreading it is not a scheduling preference. It is what lets real questions arrive between the fixes.

Rehearse with a test link before anything is public

Before the widget goes live on the site, generate a shareable test link and send it to somebody who does not know the answers. A colleague from another part of the business, your bookkeeper, a friend who has never bought from you. You are the worst possible tester of your own assistant, because you ask questions in the vocabulary of the cards you just finished reading. It will answer all of them and you will have learned nothing.

What you want is somebody who types "do you deliver to my country" and "is the blue one in stock" and "can I send it back if I already opened it". Give them twenty minutes and tell them the goal is to break it. Twenty minutes of that is worth a day of your own testing.

Two rules make the rehearsal worth doing. Sit with them and say nothing, because the moment you explain what it meant you have contaminated the test. And read the transcript afterward rather than relying on what they tell you, since people are polite about software and the transcript is not. The link needs no permission from anybody, which is why it is the one setup step you can do on the first afternoon regardless of who holds the passwords.

Where this is the wrong plan

If you sell under about thirty products and you get a handful of questions a week, do not do this. A clear shipping page and a monitored email address will serve those customers better, and the arithmetic is not close: the first week above costs you a working day, and then a standing weekly slot for as long as you run it. Below a certain volume that never pays back.

Do not start this in the ten days before your biggest sale either. Peak traffic is precisely when the review step gets skipped, and a rushed draft answering at volume is worse than no assistant at all. The install and the review want a quiet fortnight.

And if the question you are drowning in is where a particular order has got to, no amount of setup time changes the outcome. An assistant answering from your catalog and your written policies cannot look up an individual shipment. That is a readiness question rather than a scheduling one, and it wants settling before you block out the week, not halfway through it.

The maintenance that never ends

Prices change, couriers change, policies change, and the assistant is only ever as current as what it was told. A supplier moves a lead time from three days to ten and nothing in the system notices, because nothing in the system was watching. That is not a defect of any particular product. It is what it means to have written your knowledge down.

A standing weekly slot with the transcripts is the difference between an assistant that improves and one that quietly goes out of date. How long that slot needs to be, and what is worth counting while you are sitting in it, is a measurement question with its own answers. What belongs to setup is what you are reading for, which is narrower than quality: the question that produced a hedge or a handoff twice in the same week, because two of those is a missing card. The per-message traces will tell you which cards were used to build a given answer, which turns "why did it say that" from a mystery into a two-minute check.

A fair test of a vendor's "fast"

Ask them to set it up on your own site, in front of you, with nothing prepared. No slides, no sandbox store, no content written the night before. Everything else is a rehearsal, and rehearsals are fast by construction.

Watch one moment in particular: the first question their drafted content does not cover. What the assistant does there is the only thing you are really buying. If it invents a plausible answer, you have seen your future support inbox. If it says it does not know and offers a person, that is the behavior you want, and now you know what the rest of the week is for.

What finished looks like

Finished is not a perfect assistant. It is one where the questions it cannot answer are the questions you would not have wanted answered automatically anyway: the wholesale inquiry, the complaint, the customer asking for an exception to a policy. Those should reach a person, and it is fine that they do.

So the test at the end of the first week is not "did it get everything right". It is whether the list of things it hands off is a list you chose. If it is, you are done setting up, and what remains is the weekly slot.

Common questions

How long does it really take to set up an AI chatbot?

The install is about forty minutes: a one-line embed snippet on the site, a Shopify OAuth connection, the import that pulls in your product catalog, pointing your existing WhatsApp number at it, and setting business hours and where handoffs get emailed. The content is the real work, and a realistic figure is one working day spread across the first week. After that it becomes a standing weekly slot with the transcripts, and how long that slot needs to be is a measurement question rather than a setup one.

Can the setup be automated?

The drafting can. A crawler reads your public site and produces a first set of knowledge cards, and you can connect Claude, ChatGPT or Cursor over MCP and have that assistant write the cards with you in conversation. The review cannot be automated. A crawler faithfully reproduces a policy your site states wrongly or states twice, and a drafting assistant will confidently write down what your About page merely implies. Budget about an hour to read the drafts, checking numbers and thresholds hardest.

What slows an AI chat setup down the most?

Waiting for access, not the work itself. The embed snippet needs whoever can edit the theme or tag manager, the Shopify connection needs someone with permission to install an app on the store, and the WhatsApp number is often controlled by an agency or a former employee. Each step is minutes once you have the credentials. Chasing them is what turns a forty-minute job into three days, so name the owner of each item before you start.

How soon can real customers use it?

Day three is a reasonable target. Day one is install, catalog connection and asking it ten questions yourself. Day two is fixing the five worst answers, which will be shipping, returns and stock. Day three you let real traffic in and read every transcript. Before any of that, send a shareable test link to someone who does not know the answers and ask them to try to break it, because you test in the vocabulary of the cards you just wrote and they will not.

Do I need a developer to install it?

Usually not. The website widget is a single embed snippet pasted into the theme or a tag manager, and the Shopify connection is an OAuth approval with no code at all. You need a developer only if your theme is locked down or changes go through a release process. In that case the developer's time is about five minutes and the calendar time is however long their queue is.