An embedded AI knowledge assistant, built by a two-person team
A two-person AI team shipped a production, document-grounded AI assistant for a 50+ office professional-services network, embedded in the platform its users already work in. On the delivery estimates, it landed at roughly 3× the speed and about a third of the cost of a conventionally-staffed team.
Before
- Documentation scattered across PDFs, Word files, SOPs, and unsearchable training videos.
- Users interrupted work to search folders, open files one by one, or ask a colleague.
- Many users did not know the documentation existed; support load grew as the platform added features.
- Conventional estimate: about 43 working days with a business analyst, developer, and QA.
After
- A governed, document-grounded assistant embedded in the platform users already work in.
- Instant answers with inline citations and deep links; it refuses rather than guesses; zero data retention.
- Live in production across all 50+ offices, with self-service handling routine questions.
- Delivered in about 15 working days by a two-person AI team: roughly 3x faster at about a third of the cost.
Knowledge everyone had, no one could find
The client runs a feature-rich, evolving web platform used daily across a network of 50+ franchise offices. Its documentation had grown across formats (PDFs, Word documents, SOPs, and video training recordings) stored without a searchable, unified interface. Users either could not find what they needed or did not know the documentation existed. Every operational question meant interrupting work to search folders, open files one by one, or ask a colleague: friction at exactly the moment guidance was needed, and growing support load as the platform added features.
A grounded assistant, embedded in the workflow
Rather than a general-purpose chatbot, the team built a deliberately constrained, document-grounded assistant, embedded as a persistent widget inside the client's platform, using the existing session, with no separate sign-in. It answers only from approved internal documentation and is instructed to refuse, not guess, when information is missing.
- Grounded answers, every time. A short direct answer, a supporting explanation, and a Sources section with document titles and deep links to the exact passage.
- Whole knowledge base searchable. An ingestion pipeline indexes PDFs, DOCX, SOPs, structured tables, diagrams (via OCR), and video transcripts, with scheduled incremental sync.
- Trust by design. Explicit refusal for out-of-scope queries, an AI-limitations disclaimer on every response, and a zero data-retention policy for all AI-vendor interactions.
- Continuous improvement. Like/Dislike feedback plus an admin panel for conversation logs, feedback, and prompt configuration.
Development began in a sandbox so the client's operations team could validate answer quality, tone, and source coverage before a line of production code was written. Built model-agnostic on Claude, the assistant shipped MVP-first, then expanded to video indexing, table/image processing, inline citations, and guided in-product workflow assistance.
A two-person AI team, ~3× faster at ~⅓ the cost
The delivery was run by a two-person AI team, an AI Product and an Forward Deployed Engineer, rather than the conventional business analyst, developer, and QA trio. On the delivery estimates for the two staffing models:
- ~3× faster. About 15 working days for the AI team, against an estimated 43 working days conventionally.
- ~⅓ the cost. Roughly a third of what the conventionally-staffed team was estimated to cost (about 2.7× cheaper).
- Quality held. The same production outcome: a live, governed, citable assistant deployed across the network.
These are delivery estimates comparing the two staffing models, not a parallel measured build.
Live across the network
The assistant is deployed to production and accessible to users across all 50+ offices. Typical knowledge queries return in under five seconds; every answer is grounded in approved documentation; no conversation data is retained by the AI vendor. Routine support questions are now handled by self-service, freeing operations staff for higher-value work, and new users onboard faster with guided assistance available from day one. A content-governance framework keeps responsibilities clear: First Line Software owns ingestion and the platform; the client owns content and domain approval.
AI-native delivery, in production. A compact, AI-augmented team delivering a governed, document-grounded product faster and for less, in the shape of the RACE Programming framework.