Support

Most of your support volume is five questions asked a thousand ways.

If your inbox carries the same handful of questions every week, those answers already exist somewhere on your site. They are just not reachable at the moment someone needs them. That gap is a content routing problem, not a headcount problem.

Before and after on a support siteThe same visitor question, handled by the current site and then by an assistant grounded in that site's own content.BEFOREVisitor searches the site,finds three pages, noneobviously appliesFills in a contact form,waitsSupport answers by pasting ahelp article linkAFTERVisitor asks in plainlanguage on the page theyare already onAssistant answers from yourdocumentation and cites thepage it usedIf the sources do not coverit, the assistant refusesrather than guessing
  • Private launchOnboarding selected teams now
  • Built by CreoglyphSeven client engagements behind the product

The problem is not volume, it is repetition

Support teams rarely drown in novel problems. They drown in the same five questions arriving through a contact form, at a rate that makes each one feel new to the person answering and routine to the person reading the thread afterwards.

Those five questions are almost always documented. There is a help article, a pricing note, a policy page. The visitor did not find it, or found it and could not tell whether it applied to their situation. Publishing more documentation does not fix this, because the failure is retrieval, not coverage.

The second cost is subtler. Because the routine questions arrive through the same channel as the genuinely hard ones, the hard ones wait behind them. Response time on the queries that actually need a person degrades because of queries that never did.

What changes

Today

  1. Visitor searches the site, finds three pages, none obviously applies
  2. Fills in a contact form, waits
  3. Support answers by pasting a help article link
  4. The same question arrives again next week from someone else
  5. Nobody logs that the page failed, so the page never improves

With an assistant

  1. Visitor asks in plain language on the page they are already on
  2. Assistant answers from your documentation and cites the page it used
  3. If the sources do not cover it, the assistant refuses rather than guessing
  4. The conversation escalates with the full transcript and the pages it read
  5. The failed question is logged as a content gap you can act on

What the assistant actually does with a support question

Reads the question

Not keyword matching. The question is embedded and compared against your indexed content by meaning, so 'can I get my money back' reaches a refund policy that never uses that phrasing.

Retrieves the passages

The relevant chunks come back ranked. Two or three go into the prompt, not twenty, because irrelevant context makes the answer worse rather than more thorough.

Answers or refuses

If the passages contain the answer, it is phrased from them and the source page is cited. If they do not, the assistant says so. A support assistant that cannot say no is a liability.

Escalates on a written rule

Repeated failure on the same thread, an explicit request for a person, or a topic you have marked as always human. Rules you can audit, not a judgement the model makes fresh each time.

Logs the gap

Every refusal and every low confidence answer is recorded by topic. That list is the most honest content roadmap you will get.

The three ways a conversation leaves the flowMost conversations end in a grounded answer. The exits are a logged refusal, an escalation to a person, or a capture before continuing.IN THE FLOWQuestionRetrieveGrounded answermost conversationsEXITSRefuse, log gapEscalateCapture, then continue

Deflection is the metric most often inflated

Vendors count every answered question as a deflected ticket. That assumes every visitor who asked would otherwise have written in, which is not true. Most would have left. Counting them as deflections turns a bounce into a saving on a slide.

The number worth tracking is narrower: conversations where the visitor asked something your support queue demonstrably receives, got a grounded answer with a citation, and did not then open a ticket in the following day. That is a smaller number and it is a real one.

The second number worth tracking is refusal rate by topic. A rising refusal rate is not a failing assistant. It is a map of what your documentation does not cover, sorted by how often people need it.

Where the data goes

Conversation transcripts are stored so you can read them and find content gaps. Visitor questions are sent to a model provider to produce an answer, along with the passages retrieved from your own content. Creobot holds no security certification and does not claim one. Security and data boundaries.

Questions

No, and a tool that claims it will is describing a different product. It removes the repetitive tier from the queue so the people you have spend their time on the questions that actually need judgement.

It says so and escalates. Grounding is a refusal mechanism before it is an accuracy mechanism. An assistant that answers everything is either exceptionally well sourced or quietly making things up, and on a support page the second one creates commitments you did not make.

No. It answers from your published content, so anything requiring a lookup against a customer record is outside what it can do. Those should be escalation triggers rather than attempted answers.

Two layers. Source selection controls what it can see, and escalation rules control what it must hand over regardless. Rules that matter commercially, like never quoting a price, belong in both.

Live handoff is planned for launch on the Scale plan. Today the honest description is that a conversation can be escalated with its transcript, not that a person is waiting on the other side of it.

The limiting factor is source selection, not setup. A site with good documentation and a clear sense of which pages are canonical is useful quickly. A site whose answers live in people's heads is not, and no tool changes that.

One credit covers one conversation regardless of how many messages it takes. See pricing for the full plan table.

Signal

Is this your situation?

Access is by waitlist, demo request or public-content pilot. Describe the question your visitors keep asking and we will tell you honestly whether it fits.