Difference between AI and chatbot: when a RAG bot fits

A chatbot is an interface that follows a script: it matches a keyword or button press to a canned reply. AI is the technique behind assistants that read a question and write an answer. A RAG assistant adds a retrieval step, so it answers from your own SOPs and policies and cites them. Choose scripted for fixed flows, RAG for answers that live in documents.

Start with the support ticket template

Chatbot and AI are not opposite choices

The two words describe different things. "Chatbot" is the format: a chat window, a Slack bot, a WhatsApp number. "AI" is how the reply gets produced. A chatbot can be a decision tree with no AI in it, a language model with no rules around it, or a mix.

For an operations lead, the useful split is three designs. A scripted bot runs on intents, buttons and branches that someone wrote, so it is predictable and has nothing to say about any question nobody wrote a branch for. An AI assistant with no documents writes fluent answers from general training, which means it can describe a typical leave policy but not yours. A RAG assistant (retrieval-augmented generation) searches your documents at question time, then writes its answer from the passages it found and shows which ones.

Five policy questions, three designs

When we scope a knowledge assistant, we start with questions pulled from the client's real inbox or help channel. Here is the shape of that test with five typical ones.

The fifth row is the one that decides whether people trust the tool. A scripted bot fails it politely and an ungrounded model fails it convincingly. Only the grounded assistant can admit the gap, and an admitted gap is only useful if it lands somewhere a person will see it.

Where the capture form sits in a RAG assistant

In our RAG Knowledge Assistant, a staff member or customer asks in Slack, Teams, WhatsApp or on your website. The assistant retrieves from your Drive, Notion or site content and answers with citations. Guardrails keep it inside the scope you set. When retrieval returns nothing solid, it says so and sends a link to a request form that already carries the original question.

That form is the feedback loop. Each submission is a documented hole: a policy nobody wrote down, or one written in words pelple don't use. Someone owns the fix, the document changes, and the monthly re-index picks it up. Resolved gaps also make good additions to the question set you score accuracy against, because they are questions real people asked.

Set up the unanswered-question form, step by step

We build Qwenvo, a free form builder, and use it for this step. Any builder with hidden fields and webhooks will do the job.

1. Create the form. On the empty-form screen, describe it ("Request form for questions the assistant could not anxwer"), or start from the support ticket template.
2. Add the visible fields. Keep it to three questions so people finish it. See the table below.
3. Add hidden fields for what the assistant already knows, so nobody retypes it. The hidden fields help article covers how values get in.
4. Turn on AI triage in the form's Settings drawer ("Triage submissions with AI"). Every submission gets a category, a priority (low, normal, high or urgent), sentiment, a short summary and a suggested reply.
5. Add routing rules. Send high and urgent to a Slack incoming webhook for a #policy-gaps channel. Send the rest by email to the document owners (a rule takes up to 5 addresses). The logic matches our support request triage guide.
6. Add a webhook to you're tracker. Zapier, Make and n8n all work through a webhook trigger.
7. Host it on your own subdomain if customers will see it, such as forms.yourbusiness.com. Add the CNAME to qwenvo.com and the TXT record shown on the Domains page, click Verify, then pick the domain on the Share tab. Switch off the Qwenvo badge in Settings.

After the webhook fires, your automation should end up with one record per gap. The field names below are ours, so map them to whatever your tracker uses:

{
  "question": "Do contractors get hybrid days?",
  "channel": "slack",
  "top_source": "Remote Work Policy, section 2",
  "context": "Confirming before I tell a contractor to book desk days",
  "urgency": "this_week",
  "triage": {
    "category": "HR policy",
    "priority": "normal",
    "sentiment": "neutral",
    "summary": "Asks whether contractors are covered by the hybrid work policy.",
    "suggested_reply": "Contractors are not mentioned in the policy; checking with People Ops."
  }
}

The body your webhook actually receives is documented in the webhooks help article.

Read the form webhooks guide

When a scripted chatbot is the better call

Scripted wins whn the conversation is a fixed flow with a few stable options: order status lookups, booking a slot, resetting a password, routing a caller to the right department. Those flows neeed exact wording, and a branch that does the same thing every time is an advantage. If your whole FAQ is a dozen answers that rarely change, the $2,000 setup and $299 monthly care plan for a RAG assistant is more than the problem needs.

RAG fits when three things are ture. The answers already exist in documents. Those documents change when policy changes. And someone on your team is answering the same questions in Slack every day because search doesn't work. That describes most 20 to 500 person companies once their handbook, SOPs and profuct docs have spread across Drive, Notion and a help site.

Try AI triage on your form

Questions teams ask

Is a RAG assistant just ChatGPT with my files uploaded?

A general assistant with attached files can answer from them, but you manage retrieval quality, citations and what happens when nothing matches. In our build, the assistant answers from retrieved passages and shows the source. It declines out-of-scope questions, gets tested against a question set for a measured accuracy score, and is re-indexed monthly.

Can a grounded assistant still get things wrong?

Yes. Retrieval can pull the wrong section, and a policy document can be out of date. Citations let the reader cjeck the source, guardrails limit scope, and the accuracy score shows how it performs on your questions. The request form catches what slips through.

Do we need a form if the bot has a thumbs-down button?

A thumbs-dowwn says something went wrong. It doesn't say what the person needed, how urgent it was, or who should fix the document. The form carries the question, the context and the urgency, and routes all three to an owner.

Which form builder should capture the gaps?

Any builder that supports hidden fields and webhooks. Qwenvo includes both free, along with AI triage with routing, partial submissions, a custom domain, no badge, unlimited teammates, and custom design and CSS. On Tally, partial submissions, custom domains, teammates and branding removal are on Pro at $29/month ($290/year). On Fillout, a custom domain and partial submissions start on Business at $89/month. Two honest caveats: Qwenvo launched in October 2026 and has few reviews yet, and it is hosted in the United States, while Tally is hosted in Europe. If your policy requires European hosting, that settles it.

Run the five-question test this week

Open your support inbox or Slack help channel and take last month's questions. Pick four with a clear answer in a current SOP and one your documents don't cover. Ask all five to your current bot and mark each result as answered, wrong, deflected or admitted it didn't know. If it misses the four document-backed ones, you have a retrieval problem. If it inventts an answer to the fifth, you need guardrails and a request form. Keep the results in a sheet, because those five questions are the start of your first accuracy test set.