Building an AI relationship-intelligence platform from 0→1.
Bring Your Own Network went live · network of ~1,000 agencies
Professional networks have the signal. The problem is that signal is fragmented — across email threads, meeting history, company context, and connection data — and none of it is connected to what you actually need right now. Collective OS is an AI platform that turns disconnected relationship data into actionable intelligence: who to reach, why they matter, and what path gets you there.
This is an anonymized reconstruction of real product work. The workflow UI, data, names, and examples are synthetic; the stated outcomes reflect the original programme.
Finding the right problem to solve.
From Signal to Bet
17 competitors. 14 feature dimensions. One insight.
A full competitive landscape across CRM, relationship intelligence, network tools, and AI assistants revealed that most tools either required manual input (CRMs), ignored relationship depth (network tools), or automated contact without context (AI assistants). The gap: a product that combined passive signal capture with active, contextual intelligence.
We cut from 29 features to 23 — then down to 6 MVP epics and 7 core capabilities — based on one filter: does this strengthen or dilute the core value loop?
Ship Fast + Measure
Building measurement into the loop from day one.
The outcome of AI-assisted work is not just the model response — it’s whether the human acted on it, and whether that action created value. I defined the measurement framework before implementation, not after.
- Time to useful shortlist — from opportunity intake to acted-on shortlist
- Recommendation acceptance rate — shortlisted candidates where outreach was initiated
- Evidence coverage — recommendations where the user rated evidence as useful
- Repeat usage — users returning within 7 days of completing a workflow
Building a system, not a feature.
System Model
Every capability — Network Intelligence, Reconnect, the Recommendation workflow — depends on the same underlying data: relationship signals, normalized and scored. Building that shared layer correctly meant every downstream product could be built on a consistent foundation.
Signal normalization — Product / System Design Requirement
Different sources, one relationship model.
Decision: Create a normalized relationship/signal layer so downstream capabilities did not need to interpret every raw source independently. Network Intelligence, Reconnect and future agent workflows all need a consistent representation of relationship context.
Gmail / BYON integration lifecycle
CONTRACT
What information enters/leaves the boundary?
STATE
What state is the integration currently in?
ACCESS
What has the user authorized?
FAILURE
What happens when an external dependency fails?
RECOVERY
How does the user/system recover?
TELEMETRY
Is it broken, unused, stale, or functioning?
Partial failure handling — Product / System Design Requirement
What happens when one dependency fails?
YES
Surface available evidence, explain limitations, human review
NO
Request more context / show unavailable state
Degrade gracefully. Don’t hide missing context behind a confident AI response.
Permissions as product experience
- What access is actually required for the workflow to work?
- What does the user explicitly consent to sharing?
- What happens if the user grants partial access?
- How does the user review or revoke access later?
- What does the user understand about how their data is used?
- What happens to stored data if the user disconnects?
Permissions are part of the product experience — not just an engineering concern.
Where AI helps. Where humans decide.
SYSTEM / RULES
- Normalization
- Signal scoring
- Matching rules
- Ranking logic
- State management
AI
- Capability extraction
- Seniority inference
- Context enrichment
- Interpretation narrative
- Evidence explanation
HUMAN
- Opportunity definition
- Recommendation review
- Shortlist decisions
- Override / dismiss
- Outreach initiation
Quality at every stage — click a stage to see who is responsible
System / Rules
Receives structured opportunity from the user. Validates required fields. Routes to the interpretation layer.
AI
None. This is a deterministic intake step.
Human
Fills the opportunity form — role, seniority, required skills, context. The quality of the input shapes the quality of all downstream results.
End-to-end recommendation workflow
- 01Opportunity
- 02Interpret
- 03Search network
- 04Rank relevance
- 05Explain
- 06Human review
- 07Next action
Try the prototype.
Enter a real or synthetic opportunity and watch the system interpret it, match your network, and surface ranked recommendations with supporting evidence. All data is synthetic — no real people or connections.
Intake & requirements
Relationship map
Click a node to see relationship path and evidence. Dashed edges are second-degree connections.
Audit trail
No activity yet. Analyze an opportunity to begin the trail.
Agentic workflow — how it works behind the prototype
AGENT SKILL
Opportunity Interpreter
Receives the opportunity. Extracts required capabilities, infers seniority, enriches context.
AGENT SKILL
Network Searcher
Queries the normalized signal model. Returns candidates matching required capabilities within the network.
HUMAN GATE
Evidence Review
Human reviews AI interpretation and supporting evidence before shortlisting.
AGENT SKILL
Path Finder
Identifies warm introduction paths. Scores by relationship strength and relevance.
AGENT SKILL
Explanation Builder
Generates recommendation narrative: why this person, what evidence, how to reach them.
HUMAN GATE
Outreach Decision
Human decides who to shortlist, who to contact, and what path to use. No automated sending.
What shipped. What I learned.
How we measured success
- Time to useful shortlist
- Recommendation acceptance rate
- Shortlist-to-introduction rate
- Evidence coverage
- Repeat usage within 7 days
Where it landed
BYON live
Bring Your Own Network went live
~1,000
Agencies in the Collective OS network
Measurement notes
No outcome figures are claimed for this product. The measures listed alongside describe how success was designed to be tracked, and the agents behind the recommendations are covered in the agents case study.
Key decisions
INSTRUMENT THE VALUE LOOP EARLIER
Define instrumentation alongside MVP definition rather than after feature definition. Knowing whether the loop was working should have been designed in from day one.
VALIDATE RELATIONSHIP QUALITY SOONER
Network size is useful, but relationship utility matters more. Test earlier whether surfaced relationships led to meaningful action — not just whether users engaged with the list.
KEEP REDUCING SCOPE
AI products accumulate possibilities quickly. Every additional capability should prove it strengthens the core value loop — or it gets cut.
“The goal was never to automate relationships. The goal was to make relationship intelligence — the kind professionals spend years developing manually — available at the moment it was actually useful.”