Who this is for
You run a store or a service website where the same questions arrive every day. Where is my order. Do you ship to Ireland. Does this part fit a 2019 model. What is the return window on sale items. The answers exist already, on a policy page, in a product description, in a help article someone wrote two years ago. Visitors do not find them, so they email, or they leave.
An assistant trained on your own content gives those answers in the visitor's words, at 2 a.m., with a link to the page it took the answer from. It is not a general chatbot with your logo on it. It knows only what you have published and what your store tells it, and that limit is the point.
What it answers from
We load the assistant with the content you already have and keep it in sync as that content changes:
- Pages: about, shipping, returns, warranty, FAQ, contact, store locations and opening hours.
- Products: titles, descriptions, variants, specifications, prices, availability, and any size, fit or compatibility tables you publish.
- Policies: the exact wording of your refund, privacy and terms pages, so the assistant quotes you rather than paraphrasing you.
- Help articles: knowledge-base content from your helpdesk, a shared documents folder, or PDFs such as manuals and spec sheets.
- Nothing you exclude: draft pages, internal notes and discontinued products are left out, and you can see the list of what is in.
Each answer shows where it came from. If a visitor asks about delivery to Canada and the answer is on your shipping page, the assistant quotes that page and links to it. If your content does not contain the answer, the assistant says so, which is more useful than a confident guess.
When it does not know, a person takes over
Every assistant we build has a hand-off. The visitor can ask for a person at any time, and the assistant offers one itself when it cannot find an answer, when the visitor sounds frustrated, or when the topic is one you have marked as human only: complaints, refunds above an amount you set, anything legal. The conversation so far travels with the visitor, so nobody repeats themselves.
Where the hand-off lands is your choice: a live chat seat in your helpdesk, an email to a shared inbox with the transcript attached, or a ticket in Gorgias, Zendesk, Freshdesk or HubSpot with the right tags already set. Outside your working hours the assistant takes the visitor's email address and says when a reply will come.
Order and stock lookups
Connected to your store, the assistant does more than quote pages:
- Order status by order number plus email, with the carrier's tracking link.
- Stock level and expected restock date for a product or variant, where your store holds that data.
- Pickup availability by location.
- Whether an order is still inside your published return window, based on the order date.
The connection is read only unless you decide otherwise. The assistant never changes an order, issues a refund or edits a customer record on its own. If you want it to start a return, it creates a request for a person to approve. Each lookup asks for enough detail to confirm the visitor is the customer, and the transcript records what was looked up and when.
Where it installs
- Shopify and Shopify Plus: a theme app embed or script tag, with the Admin API for lookups. It can sit alongside Shopify Inbox if you keep that.
- WordPress and WooCommerce: a small plugin we hand over to you, with the WooCommerce REST API for orders and stock.
- Magento 2: a module, or a plain script if you would rather not touch the codebase, with the REST API for lookups.
- BigCommerce: a script through Script Manager, with the Storefront and Management APIs.
- Custom applications (Laravel, Node, .NET): a script and a server-side service that reads from your database or API, on your own hosting.
The assistant itself runs as a small service on your own cloud account or your existing server. If you change platform later, it moves with you.
How we test it: the question list
Before scope is signed, we agree a list of 40 to 100 real questions, drawn from your support inbox and from what visitors type into your site search. Each has an expected answer and a source. The assistant must answer every one correctly before it goes live, and "I don't have that, here is a person" counts as correct for questions your content does not cover.
You can add to the list at any point. After launch we read the real conversations with you weekly for 30 days, add the questions it got wrong, fix the content gap or the instruction, and re-run the list. The list stays yours. It is the acceptance test for every change we or anyone else makes later, which is how you keep an assistant honest after the people who built it have moved on.
What it costs to run
The build price is fixed. Running costs are billed by the providers to your own accounts:
- Model fees: for most sites, $40 to $200 a month. A store with around 1,000 conversations a month at a few turns each sits toward the lower end on a mid-sized model; a site answering long technical questions from manuals sits higher. The fit check gives you a number for your volume.
- Hosting for the small service: usually under $20 a month on your cloud account.
- Care plan (optional): $250 to $1,000 a month for monitoring, content re-sync, model updates and a monthly log review by a named engineer. Cancel any month.
There are no seat fees and no per-conversation markup from us. If your volume doubles, you see it on your provider's bill, not on ours.
What it will never do
- Promise a refund, a discount, a delivery date or a price you have not published.
- Answer from general knowledge outside your content. Asked for a recipe or for stock tips, it declines and returns to your topic.
- Invent a product, a policy or a phone number. If a fact is not in the sources, it says so.
- Change an order or a customer record on its own.
- Send your customer data to any account that is not yours.
- Pretend to be a person. It says it is an automated assistant when asked, and a person is always one message away.
A worked example
The numbers below are illustrative. They show the shape of a typical job, not a result we are claiming for a client.
Shopify store, 2,400 products, two people on support. About 720 support emails a month, of which roughly 430 are order status, shipping, returns or "will this fit" questions whose answers are already on the site.
- 720 emails a month reaching a person, average first reply 9 hours, Monday backlog of 60 to 80
- Return window and shipping-zone questions answered differently by each person
+ Assistant live in week 3; 86-question test list passed; order and tracking lookups through the Admin API
+ Month 2: about 430 conversations closed by the assistant, about 290 reaching a person with the transcript attached
+ Running cost in month 2: $78 model fees, $12 hosting, both on the store's own accounts
What it costs
| Work | Includes | Fixed price |
|---|---|---|
| AI fit checkThe first step | One-page plan: what it would answer from, what it connects to, running cost, fixed build price | Free |
| Assistant on your own content | Install on your platform, content loading and sync, source-linked answers, hand-off to a person, 40 to 100 question test list, logs you can read, 30 days of tuning | from $3,500 |
| Assistant with store connectionShopify, WooCommerce, Magento 2, BigCommerce or custom | Everything above plus order status, tracking and stock lookups through your store's API, with identity checks | from $4,500 |
| Extra content sources or helpdesk connectionEach | A second knowledge base, a PDF library, or a ticket connection to Gorgias, Zendesk, Freshdesk or HubSpot | $600 to $2,000 |
| Care planOptional, cancel any month | Monitoring, content re-sync, model and connector updates, monthly log review, small changes, a named engineer | $250 to $1,000 a month |
Half the fee at start, half at launch, by wire transfer or PayPal. Model fees and hosting are billed by the providers to your own accounts, usually $40 to $200 a month in total. We do not take jobs under $2,500.
How it runs
Free AI fit check (2 working days)
You send your site address and the questions you answer most. We read your content, check what your platform's API exposes, and send a one-page plan with a fixed price and a monthly running cost. If the platform's own chat tool already covers you, the plan says so.
Scope and question list (1 week)
Written scope, the 40 to 100 question test list with expected answers, and the hand-off rules. Model and cloud accounts created in your name. Half the fee paid.
Build on staging (1 to 3 weeks)
Content loaded, store connection built read only, assistant installed on a staging copy you can talk to from the first week. Daily written log. Every question on the list is checked and you can add more.
Launch (1 day)
Live on your site. Second half of the fee. You receive the code, the runbook and access to every account.
Tuning (30 days)
We read the live conversations with you each week, fix gaps in content and instructions, and re-run the list. Then the care plan, if you want someone on call, or nothing further.
Start with the fit check
Get a free AI fit check: two working days, one page, a fixed price. Or email info@riantsoft.com with your site address and the three questions you are tired of answering.
Questions people ask
Which AI model will my assistant use, and can I change it later?
We pick the model for the job and show you the choice in the fit check. For an assistant that must quote your content accurately, we usually use Anthropic's Claude or OpenAI's models through their official APIs; for a very high-volume store a smaller, cheaper model may handle the simple questions with a larger one behind it. The provider account is in your name, so you see the bill. Because the assistant reads your content at answer time rather than being retrained on it, switching provider later is a configuration change plus a re-run of the question list, not a rebuild.
Will it make things up?
Far less than a general chatbot, and never silently. The assistant is restricted to your content and your store data, it shows the source under each answer, and it is instructed to say "I don't have that" and offer a person when the sources do not cover the question. We test it against 40 to 100 real questions before launch and read the live conversations with you for 30 days. No system is perfect, which is why every conversation is logged where you can read and correct it, and why it can never state a price, a date or a refund you have not published.
Where does our data go, and who can see the conversations?
Your content and store data stay in your accounts. The assistant runs on your cloud account and calls your model provider under a business agreement that excludes training on your data. Conversation logs are stored in your account, not ours, and only people you grant access can read them. We sign an NDA before the fit check if you want one and a data processing agreement for the build. After handover we keep no copy of your customer data.
Which languages does it support?
Any language the model handles well, which covers English, the main European languages, and most others your customers are likely to use. The assistant answers in the language the visitor writes in, drawing on your content even when that content is in English only. If you sell in several markets we add questions in each language to the test list, because quality varies by language and we want to see it before you do.
What happens at handover? Do we depend on you afterwards?
At launch you hold everything: the code in your repository, the service on your cloud account, the model account, the question list and a short runbook that explains how to re-sync content, read the logs and change the instructions. Any engineer who knows Node or Python can maintain it. The care plan is optional and cancels any month; it exists for owners who would rather have a named engineer on call than learn the runbook.
How is this different from the chat tool my platform or helpdesk already offers?
Platform chat tools answer from a small FAQ you type in, or from generic training, and they live inside that platform. Ours reads your whole site and catalog, quotes your policies word for word, looks up orders in your store, and hands off into whichever helpdesk you use. It is a service you own, so it survives a platform move and a helpdesk change. If the platform tool already covers your questions, the fit check will tell you to keep it and save the money.
How long does it take and what do you need from us?
Two to four weeks from signed scope to launch, then 30 days of tuning. From you we need an hour to agree the question list, access to your site and store under a staff account you can revoke, and a decision about where hand-offs should land. Most owners spend three to four hours in total across the build, mostly reading test answers and telling us what they would have said differently.