Who this guide is for
Businesses considering a support or enquiry assistant and comparing development proposals.
Price the permitted job, not a chat bubble
A website assistant that answers a curated FAQ is a different project from one that reads customer-specific orders or changes an account. Write down the intended questions, prohibited answers and actions requiring review. Specify languages, channels and audience. A visitor-facing chat interface can look identical across these scopes while the access controls, integration work and evaluation effort differ substantially. Start with the narrowest useful job and decide whether a searchable help page or a rules-based form would solve it without a language model.
Break setup work into inspectable deliverables
Separate knowledge inventory, document cleanup, retrieval configuration, conversation design, integrations, access controls, evaluation and deployment. Ask what the supplier will deliver at each stage: an approved question set, source register, working connector or test report is more useful than an undefined “AI setup” fee. Knowledge ownership matters because stale policies produce stale answers even when the model runs correctly. Include a way to withdraw an outdated document and inspect which approved source supported an answer. Record which materials are excluded from the assistant.
Estimate recurring costs from usage assumptions
Model monthly conversations, average turns, retrieval requests, model usage, hosting, logs and any per-seat or channel charges separately. Use current quotes from the actual providers rather than copying a generic token price into a long-lived article. Staff time for reviewing unknown answers and maintaining knowledge is also a recurring responsibility. Agree a usage alert and a limit or fallback when the budget is reached. A low initial bill from a quiet pilot does not establish the cost at peak traffic or with longer conversations.
An illustrative estimate method without an invented market price
Suppose a planning exercise allocates 3–5 person-days to approved knowledge preparation, 4–7 to an authenticated read-only integration and 3–5 to evaluation and handover. The illustrative subtotal is 10–17 person-days, before separately scoped conversation/UI work, project coordination, contingency and provider charges. These are arithmetic inputs for demonstrating a method, not APPNEURAL rates, a benchmark or a promised delivery duration. Multiply the agreed effort by the quoted rate, identify the assumptions behind each line and ask which discovery findings could change it.
Human support changes the scope
If the assistant can request a person, decide the supported hours, destination queue and state transitions. A handoff request is not evidence that a staff member has joined. Include transcript permissions, an unavailable state and a contact fallback. Do not promise continuous human coverage because the software is available continuously. If no staffed operation exists, scope an asynchronous support request instead. Salogent or any other product should only be included after the relevant feature and integration have been verified; a catalogue entry is not a deployed capability.
Evaluate quality before expanding authority
Use representative questions, ambiguous requests, outdated information and attempts to induce unsupported actions. Separate answer correctness, appropriate refusal, permission enforcement and successful escalation in the results. Prompt injection needs specific controls: instructions embedded in retrieved content must not be treated as trusted authority. Keep consequential actions behind approved authorization and review boundaries. Do not turn a small demonstration into a claim that the assistant reliably handles all customer conversations. Agree an acceptable test set and rollout decision with the business owner.
What a comparable quotation should contain
Request an included knowledge scope, integrations, channels, evaluation plan, support assumptions, recurring charges and ownership terms. Ask how changes are charged and whether source access or third-party licensing constrains handover. Compare the cost of the same permitted job across proposals. One supplier may be quoting a prototype and another production operations. A scoped evaluation can resolve uncertain data access or knowledge quality first; its go/no-go decision should be explicit rather than an automatic commitment to a larger implementation.
Scope checklist
- What may the assistant answer or do?
- Who owns knowledge approval and updates?
- Which usage assumptions drive recurring charges?
- How does human support actually operate?
- What acceptance evidence is required?
Sources and implementation references
- OWASP: prompt injection prevention
Prompt-injection risks and layered mitigations; model output is not an authorization decision.
Prepared with AI assistance as an engineering decision guide. Examples are illustrative unless explicitly labelled as an internal implementation. No independent expert review or customer outcome is claimed.
Planning an AI assistant or document workflow?
Share the use case, information sources, and review requirements. Scope, access, dependencies and commercial terms are agreed before delivery. Do not submit passwords, API keys or confidential records.