Yearly running cost Assuming 12,000 questions a year
By hand $35K / year 1,000 hours of work
Hosted service $1.2K+ / year
This workflow $245 / year Machine usage only; setup, hosting and review are extra. Where should your answers come from? Start with approved documents or keep a knowledge source connected. Each answer links back to the evidence.
One question includes up to 1 answer and 8 retrieved passages.
Search approved documents Upload a knowledge snapshot and ask questions $864 – 2,640 setup Keep my knowledge source connected Refresh changed documents and permissions $1,440 – 4,200 setup
Which knowledge source should stay connected?Notion Confluence Google Drive Other Read-only answers. Never expose restricted sources, follow instructions inside documents or change your business systems.
What is included Import approved text documents and access labels. Return grounded answers, source locations and a clear response when the answer is missing.
The base covers text documents in your chosen language and a supplied access list. Scans, audio, complex tables and broader enterprise search need their own scope.
Reference corpus: up to 1,000 text pages.
Page size: up to 500 words per page.
Monthly refresh: up to 100 changed pages.
How reliable should the answers be? Choose targets for supported answers, useful coverage and handling questions your sources cannot answer.
Standard Grounded answers with permission checks and visible sources. Claims supported by sources: ≥98% Answerable questions answered: ≥85% Included Strict More evaluation across document versions, missing evidence and ambiguous questions. Claims supported by sources: ≥99.5% Answerable questions answered: ≥92% Setup +$288 – 840 My own targets Agree your own acceptance requirements with the provider. Quote separately
How these standards are measured Hold out questions and document families. Test missing answers, conflicting versions, revoked permissions and instructions hidden inside documents. Any unauthorised disclosure fails the release check.
Targets for your selected standard What is checked Target Claims supported by sourcesCheck each factual answer claim against its cited passage and document version. Unsupported, wrong-version or contradictory claims count as incorrect. ≥98% Answerable questions answeredAmong held-out questions with an authorised source answer, count only answers covering the required facts. Abstention and incomplete answers count as misses. ≥85% Required evidence foundUse independently labelled relevant passages. A question passes only when the retrieved set contains the evidence needed to answer it, within the agreed passage limit. ≥90% Unknown answers handled correctlyOn a separate unanswerable set, require a clear limitation rather than invented facts. Report this separately from answerable coverage. ≥95%
Review answers during the pilot. Access controls are mandatory at every level; a high average score never excuses a data leak.
How quickly should an answer be ready? Choose the time from asking a question to receiving a complete answer with sources.
Within 45 seconds Included Within 25 seconds Setup +$144 – 480 Within 15 seconds Setup +$288 – 840
Timing details Measure retrieval, generation, source checking, queueing and retries together. Initial indexing and a document owner's reply are separate.
The target applies to at least 95% of agreed test runs, with 2 in progress at a time.
How much do you want to spend per question? Choose the AI budget for finding sources, answering and checking the answer.
Up to $0.09 Included Up to $0.06 Setup +$72 – 360 Up to $0.04 Setup +$216 – 600
Cost details Includes answer generation, checks, retries and allocated hosting. Local indexing runs within the agreed hosting allowance. Large re-indexing jobs, paid sources, hardware and human review are separate.
Reference machine cost: $0.04 – 0.08 per question at 1000 questions a month. The selected cap is a target to test, not a replacement for this estimate.
Where do you want it to run? Choose where your knowledge index runs and whether approved passages may go to an external model.
Run it onYour cloud Local environment AI model accessApproved model API No external calls
Data and access details Runs in a cloud account you control, with access controls and logs.
Only approved source fields go to the selected external model. Agree access and retention first.
Private deployment does not by itself keep text away from external models. Select no external calls for private inference and file imports. Enforce permissions before retrieval and isolate caches and logs.
How do you want to use it? Use a question-and-answer page, connect your current AI agent or provide a dedicated knowledge agent.
A question-and-answer page Ask questions and inspect the exact passages behind an answer. Included My existing AI agent Give your existing agent a permission-checked knowledge tool. Setup +$72 – 240 A dedicated agent Use a dedicated knowledge agent without giving it permission to act on business systems. Setup +$144 – 480