RAG chatbot — internal pre-quote checklist¶

Used by the Eidos Global team before a RAG chatbot quote leaves the building. Walks through every client answer and confirms we have the architectural decision and the cost line to match. If any row is "not sure", do not send the quote — book a follow-up call instead.

The structure mirrors the client questionnaire; the methodology behind each row is in the quote guide.

Pre-flight — who is in the room¶

Named technical owner from the client identified (not just a procurement contact). If absent, the project will stall — flag now, not later.
Named business sponsor with budget identified.
Eidos delivery lead nominated for the quote.
Eidos pre-sales engineer nominated for the quote.

Section 1 — Use case¶

Client answer	Architectural / commercial implication	Done?
Use case described in one sentence	Quote scope is narrow enough to ship; if the sentence has "and", consider phasing
Audience departments named	Identity / permission model can be drawn
Success metric named and measurable	We can write an acceptance criterion. If not, demand one
Regulator / governance committee identified	Add evidence-pack effort to the quote
Explicit no-go zones noted	Captured as guardrail requirements in section 6

Section 2 — Documents¶

Client answer	Implication	Done?
Document count bucket	Pick small / medium / large tier from quote guide section 10
Document locations	Each new source adds 1-3 days of connector work
Document formats	Scanned PDF or non-text formats add OCR effort and quality risk
"Are these the approved versions?" — honest answer captured	If no, add a content-cleanup workshop block (2-5 days) before ingestion
Approval owner named	Required for production ingestion pipeline
Change cadence captured	Scheduled vs event-driven ingestion decision made
New documents per month estimated	Variable embedding cost line populated
Multi-language?	Embedding model and LLM choice may need adjusting

Section 3 — Users¶

Client answer	Implication	Done?
Total users	Auth licence cost, token budget envelope
Daily active users	Token spend forecast
Peak concurrency	If self-hosted, GPU sizing. If frontier API, rate-limit tier
Permission policy (flat vs per-role)	If per-role, +5-10 days retrieval-layer work in the quote
Identity provider named	Entra = 1-2 days, anything else = 3-5 days
External users in scope?	Adds licence cost and complicates audit
Conversation retention policy captured	Audit log storage line item

Section 4 — Integration surface¶

Client answer	Implication	Done?
Surface(s) chosen (Teams / intranet embed / dedicated site / mobile)	UI effort block from quote guide section 4
Branding decision captured	Design days in the quote
Existing tools to integrate or replace?	Migration / coexistence effort

Section 5 — LLM hosting¶

Client answer	Implication	Done?
Position A / B / C / D captured	LLM choice decided; if D, propose A as default + B as alternative in the quote
If B or C: existing AI infra confirmed	If no existing infra, add hardware procurement lead time to wall-clock
Compliance regime named	Add evidence-pack effort. If HIPAA / PCI-DSS / regulated finance, escalate before quoting

Section 6 — Safety and governance¶

Client answer	Implication	Done?
Citations required?	Default yes; confirm it is in scope statement
Refusal behaviour captured	Default in scope; confirm wording with client
Topic refusals listed	Add to system prompt design; allow 1-2 days of refusal tuning
Audit log requirement captured	Audit log line item (3-5 days build, monthly storage line)
PII / no-mention list captured	Output-filter scope (3-7 days)

Section 7 — Operating model¶

Client answer	Implication	Done?
Day-to-day owner on client side named	If no name, downgrade to hand-off model in the quote and flag risk
Document maintainer(s) named	Required for ingestion-pipeline acceptance
SLA tier chosen	Monthly fee line
Operating model chosen	Monthly fee line and handover-pack scope
Go-live date captured	Forward calendar check: can we resource it?

Section 7b — Companion services¶

Reference the quote guide section 11 for the day-band per line.

Client answer	Implication	Done?
One login today, or many?	If "many", quote SSO rollout (11a) as a line item, not an assumption
App count behind separate logins	Drives 11a sizing tier (small / medium / large)
Recent audit / insurance push for SSO	Raises the priority and may unlock separate budget — name in the proposal
Legacy apps needing forward-auth	Per-app multiplier from 11b; quote as unit price not lump sum
Document home today	If SharePoint / shared drive / unknown, propose MkDocs as a foundation phase (11c)
Document format consistency	Drives the MkDocs migration band — quote scanned-PDF separately, it is the single biggest variance
Documents to migrate (bucket)	Migration-days line for 11c
Approval workflow exists?	If no, scope a "document governance" workshop into 11c discovery
Browseable site valued independently of chatbot?	If yes, MkDocs has its own ROI story and can be sold standalone — note this in the proposal
Bundle preference	Pick the matching row from quote guide section 11d; apply the 5-10% engagement-efficiency framing

Bundle pre-send checks:

When pricing a bundle, the individual line-item prices are also shown, so the client sees what they would pay for each piece separately and what the bundle discount is.
If the client picked "chatbot only" but the questionnaire shows scattered documents and no SSO, the proposal includes a brief "what we noticed" paragraph naming the risk — not as a sales push, as a delivery-honesty note. The chatbot will under-perform on a weak foundation and we should say so in writing.

Section 8 — Budget and procurement¶

Client answer	Implication	Done?
Implementation budget range disclosed	Scope to fit if possible; if not, propose a phased path that fits phase 1 in budget
Monthly budget range disclosed	Drives LLM-choice recommendation (frontier vs self-host)
One-time vs programme	Influences how much investment in reusable platform vs throwaway tooling
Procurement requirements captured	Security questionnaire effort, insurance evidence, supplier-onboarding lead time

Pre-send sanity checks¶

Before the quote leaves the building, the delivery lead confirms:

One-off cost is shown as a band, not a single number, unless the scope is unusually tight.
Monthly cost has both a fixed line and a variable line, with the variable line shown as a worked example for the client's size.
At least one alternative shape is offered (e.g. frontier vs self-hosted, or phase-1-only vs full scope). Clients should see a trade-off, not a single take-it-or-leave-it number.
Assumptions section lists everything the price depends on (user count, document count, formats, SSO, permission model). When any of these change, the quote changes — clients accept this if it is in writing up front.
Exclusions section is explicit: things the client expected might be included but are not (e.g. content cleanup labour, document digitisation, mobile app, integrations beyond the named ones).
An evaluation set commitment is in the quote: client provides 50-200 known-good Q/A pairs before UAT. Without these, we cannot sign off quality.
Phase 2 is named, even if not priced. Almost every successful RAG project grows; the client should know we expect it.
The number has been reviewed by someone who is not the author.

Anti-patterns — quotes we should not send¶

A single fixed price for an implementation where the document count is "we will let you know".
A monthly cost based on "average usage" without naming the daily active user and questions-per-day assumptions in writing.
A quote that promises a regulator-grade solution but is priced like a pilot. If the client has a regulator in the room, the quote includes evidence work; if it doesn't, we are storing up a fight at handover.
A frontier-API-based quote for a client who said "data cannot leave our premises". This sounds obvious; it has happened.
A quote with no named client owner. The system will decay; we will be blamed.