What the consultant should actually do
A visitor asks about delivery time, product compatibility, or a return policy and receives an answer within seconds. To an executive, this appears to be a direct way to reduce repetitive sales and support work. The harder question follows: did the answer come from the current policy, from a plausible phrase, or from a page the company already considers obsolete?
Search results already present several formats. AiMaxChat offers a packaged knowledge-grounded consultant with CRM integration, Atolko offers a website bot with scenario testing, and Amplify Lab describes a custom system with operator handoff and an on-premises option. These are suppliers’ claims. The useful buyer distinction is the boundary between routine configuration and a separate product.
An AI consultant usually does not keep the whole company “inside the model”. The application receives a question, retrieves several relevant passages from approved sources, passes them to the model, and asks it to compose an answer. Quality therefore depends on the documents, retrieval, instructions, model, and the rules that stop or hand off a conversation.
A packaged service covers a standard path
A packaged product is appropriate when the company needs a website widget, can supply documents as pages or files, gives all visitors the same information, and can send a difficult question to an operator. This route quickly reveals which enquiries are genuinely repetitive and whether visitors use the responses.
Before paying for a plan, walk through the complete cycle: upload material, update one document, remove an obsolete one, inspect answer history, accept a conversation as a person, and export statistics. Limits on pages, conversations, or operators can affect the budget more than an impressive demonstration on five selected questions.
A custom build is warranted when the answer depends on the customer’s identity, contract, stock, order status, or a calculation in an internal system. This is no longer retrieval from public pages. The application must verify the person’s right to the data, call a narrow CRM or accounting method, handle downtime, and show which fact supports the response.
A knowledge base is not an automatic safety control
The OWASP GenAI Security Project warns that retrieval and fine-tuning do not eliminate prompt injection. Text on a webpage, in an uploaded file, or in a hidden fragment may try to alter model behaviour. Retrieved content must not become an instruction merely because it lives in a company repository.
Sources need owners, effective dates, and access rules. A public returns policy and an internal exception sheet for employees must not appear in the same visitor answer. After a price or policy change, the company needs a known time by which the consultant begins using the new version and stops quoting the old one.
The model also needs permission not to answer. When no supporting passage exists, a clarifying question or human handoff is safer. Confident unsupported prose is more dangerous than a brief refusal because a visitor may treat it as a company commitment.
Quality is measured on awkward questions
The NIST AI Risk Management Framework core calls for documented scope and limits, human roles, test sets, measures, and production monitoring. For a website, this can become a concrete acceptance contract: which questions the consultant answers, clarifies, hands off immediately, or must never answer with protected data.
A working conversation is more than one accurate message. The operator needs the question, retrieved material, contact details, and handoff reason. If the visitor must repeat everything, automation has accelerated the opening response while making completion worse.
- Collect real questions from chats, calls, forms, and search logs after removing personal data.
- Add ambiguity, misspellings, obsolete names, and questions that have no answer in the knowledge base.
- Record the accepted answer, permissible variants, required source, and human-handoff condition.
- Measure correct answers, honest refusals, harmful errors, handoff time, lost context, and conversations that reach the intended action.
- Repeat the set after changing documents, model, instructions, or integrations instead of treating one successful run as a permanent guarantee.
Price depends on what happens after the answer
The simplest budget is a packaged subscription plus material preparation and configuration. Custom cost grows with authentication, several permissioned knowledge sources, CRM calls, action logs, on-premises hosting, high load, or authority to change data in another system.
pommeDeTerre’s current price list starts a knowledge-grounded AI consultant at RUB 150,000 and an assistant that performs approved actions in external systems at RUB 300,000. The boundary matters: reporting an order status and changing it require different controls, tests, and responsibility. A specific estimate follows selection of one process and examination of its sources.
A sensible first launch covers one website section, one language, and hours when a human can actually receive a difficult conversation. Then measure the cost of a usefully completed enquiry rather than the cost of a message: the visitor receives a verifiable answer, submits a valid lead, or reaches an employee with context intact. That chain reveals whether the company bought a consultant or merely another way to generate prose.