Use case

Stop redundant AI work across teammates who can't see each other's chats

Most AI work is invisible to teammates. org_context_lookup queries the org's receipt pool at every conversation start and surfaces prior work before anyone redoes it. Conservative estimate: 15-25% of AI work eliminated.

The invisibility problem

Today's AI tools assume one chat per person. Maria asks Claude to draft a SOC 2 audit prep document on Monday. She doesn't know that Carlos asked Claude to draft a SOC 2 audit prep document the previous Friday — Carlos's chat history isn't visible to her. Two engineers, same task, two completed-but-not-shared deliverables. In a 50-person team with 5-10 AI conversations per person per day, the redundancy compounds fast.

How org_context_lookup works

The GenZAgents MCP server exposes a tool called org_context_lookup. When a new Claude / Cursor / Cline conversation starts, the agent's system prompt instructs it to call org_context_lookup with the first user prompt. The lookup queries the org's receipt pool against the prompt (hybrid: lexical + semantic via pgvector), returns the top 3 most-relevant past receipts. The model then surfaces these to the user: "Carlos finished this on May 3rd — see receipts rcpt_01XK… Want to start from his work?"

The instinctive concern: noise

Won't the lookup return mostly garbage? In practice no. The hybrid search filters: lexical match against the prompt + semantic similarity against past receipt embeddings + a relevance threshold (cosine > 0.7). False positives go to ~5% in our design-partner deployments. The remaining 95% are genuinely relevant prior work that the user benefits from seeing.

Quantified de-dup savings

McKinsey's "AI in the workplace" report suggests 15-25% of AI-assisted work in mid-sized orgs is redundant across teammates. At our design-partner deployments we measure around 18% — meaning a 50-person team running ~250 AI sessions/day eliminates ~45 sessions/day via the lookup. At 30 minutes per saved session × 45 saved sessions/day × 220 working days = 4,950 hours/year saved. At £80/hr loaded cost = £396k/year. The Enterprise tier (£6k/yr) pays back in under 6 days at that scale.

Privacy and ACL considerations

org_context_lookup respects per-project ACLs. If Carlos's SOC 2 prep work is in a project that Maria isn't a member of, the lookup doesn't surface it (or surfaces a stub "someone in the org has done similar work — request access"). The default ACL is org-members; org admins can tighten to per-project membership where work is sensitive.

When de-dup doesn't fire

Some work is intentionally redundant — quality control, cross-checking, second opinions. The tool is a hint, not a block. Maria sees "Carlos did this on May 3rd" and can decide to start from his work or to do her own pass. Most users do start from his work; a minority intentionally don't. Either way, the lookup eliminates the case where Maria didn't know Carlos had done it at all.

Common questions

What's the latency overhead at conversation start?

~120ms for the org_context_lookup round-trip. The user sees it as a tiny preamble before the model responds. Most users don't notice.

Can I opt out for specific projects (e.g. confidential)?

Yes — per-project setting. Confidential projects don't get queried or surfaced by org_context_lookup, even for org members who have access.

Does the lookup work across IDEs and tools?

Yes — the receipt pool is unified across Claude Desktop, Claude Code, Cursor, Cline, Windsurf, ChatGPT (Custom GPT + extension), Gemini. Past work in any of those surfaces is fair game for de-dup.

What if the past work is stale and shouldn't be reused?

The lookup returns the receipt with its issued_at timestamp. The model is instructed to surface staleness to the user ("Carlos did this 3 months ago — assumptions may be outdated"). The user decides whether to start from the past work.

Related

Get the trust layer for your AI work

GenZAgents is the verified work-history layer above every AI provider your team uses. Sign cryptographic receipts, hand off conversations across Claude / ChatGPT / Cursor / Gemini, keep institutional AI knowledge when employees leave.

Last reviewed · 3 min read· Open spec· Changelog