Complete technical specification of the seven-agent AI investment pipeline, rebuilt around the compound risk architecture. Each agent is specified with its inputs, processing logic, outputs, and integration with the two-axis scoring model and the investment chain mechanism.
The pipeline is not sequential. Agents A-01 and A-02 run in parallel. A-03 runs as soon as A-02 completes. A-04 is triggered conditionally by A-03's score. A-05 waits for both A-03 and A-04. A-06 and A-07 run post-investment. End-to-end latency: under 48 hours for auto-approve candidates; 72–96 hours for candidates requiring A-04 interview.
A-01 and A-02 run simultaneously and independently. A-02 does not see A-01's outputs before completing its research — this prevents A-01's conversation from anchoring A-02's income estimate. The two outputs are compared in A-03 as independent signals.
Conducts a 20–35 minute natural language conversation via WhatsApp. Language is auto-detected from the first message and the entire conversation proceeds in that language. The agent does not ask about finances first — it asks about the business, the community, the people. Financial questions emerge from the conversation rather than leading it. Average completion rate: 84%. Average duration: 28 minutes.
The conversation builds a structured profile covering: business type and history, community relationships, group membership, income sources, existing financial relationships (mobile money, savings groups, any MFI history), and the candidate's own assessment of what investment would enable.
Runs in parallel with A-01, querying four independent data sources to produce a structured research dossier. The income baseline is triangulated from three signals — declared (from A-01 output), transactional (from mobile money history), and peer-validated (from NGO programme records where available). The baseline is the median of the three; significant divergence between signals is flagged.
New in the compound model: A-02 now also queries the asset catalogue to score asset-income fit for the top 3 most appropriate assets for this candidate's franchise type and geography. This output is passed directly to A-03 for the asset coverage axis.
A-03 is the most important agent in the pipeline. It takes the outputs of A-01 and A-02, scores the candidate on two independent axes, positions her in the investment matrix, and produces a specific recommended capital structure — not just an approval or decline.
Composite score from five weighted signal groups. Auto-approve above 75 on this axis alone (if asset coverage also above 40%). Below 55 requires either high asset coverage or human review.
Score reflecting both the proportion of working capital deployable as title-retained assets and the quality of those assets as structural protection. Zero asset coverage = trust-only model.
A structured 15–20 minute video conversation via WhatsApp or equivalent. Not a form — a dynamic conversation that adjusts based on responses. The AI analyses response latency, narrative consistency, vocal stress patterns, and cross-references statements against A-01 and A-02 data in real time.
A baseline is established in the first session. Subsequent monthly check-ins (A-07) track deviation from this individual's baseline rather than from a population average — controlling for cultural and individual variation in expression style.
Three people nominated by the candidate — savings group members, regular customers, neighbours — each receive a short independent AI interview (8–12 minutes). The AI checks for narrative consistency across peer accounts and against the candidate's own account, without revealing what any other interviewee said.
A-05 takes the final compound score (from A-03, updated by A-04 where triggered) and produces the complete investment recommendation. It does not recalculate the score — it structures the investment to the score. The recommended structure from A-03 is confirmed or refined based on any A-04 updates.
Critically: A-05 also activates the chain nomination process. The candidate is informed that she will nominate three people from her community before receiving the working capital tranches — and that those three people's access to investment depends on her completing the revenue share relationship. The nomination request is framed as a partnership, not a condition.
Generates a bilingual Stage 1 Pre-Formalisation Agreement (English + candidate's language) from the A-05 output. The agreement is plain-language by design — comprehensible to someone who has never seen a legal document. Title retention language is explicit: assets listed by name and description, ownership conditions stated clearly, transfer trigger defined as the break-even threshold amount.
WhatsApp delivery. Candidate acceptance recorded with timestamp and message hash as execution evidence — legally valid in Rwanda, Kenya, Uganda, Tanzania, and Ghana without physical signature.
A-07 does not produce a single risk flag. It produces a compound risk profile update — both trust score and asset status updated simultaneously. A missed payment is interpreted differently depending on the current trust score and asset coverage status (see Appendix A for three worked examples).
Chain trigger evaluation runs monthly: when cumulative revenue share reaches the break-even threshold AND the current trust score meets the minimum set at investment, the chain is released — three nominee intake conversations opened automatically, no human intervention required.
LP-grade KPI reports generated automatically as a byproduct of monitoring. No additional data collection required. Reports delivered to management company dashboard continuously; LP-formatted reports generated quarterly.
The asset catalogue is a structured database of available productive assets by franchise type and geography. A-02 queries it during research. A-03 selects from it during structuring. A-06 instructs procurement. A-07 monitors condition. The catalogue is updated quarterly as new geographies and franchise types are activated.
Essential for savings group coordinators, market traders, and any franchise requiring mobile money at full functionality. MDM profile installed at disbursement.
High fit for community kitchens (refrigeration, lighting), learning companions (evening sessions), and any franchise requiring reliable evening operation. Eliminates kerosene cost immediately.
Sewing machine (tailoring), grain mill (food processing), refrigerator (community kitchen / pharmacy). Direct income generation. Asset-income fit typically 0.90+.
Market traders, agricultural collectors, delivery service operators. Replaces 2–4 hours daily walking. Directly increases daily trading capacity and geographic reach.
Repair collectives, health workers, agricultural collectors in large-radius geographies. Highest value single asset in catalogue. Vehicle registry title retention is legally cleanest enforcement mechanism.
Repair tools (diagnostic equipment, hand tools), healthcare equipment (TENS machine, treatment table), documentation equipment (camera, audio recorder for cultural memory keepers). Directly enables the income-generating activity.
The investment chain is not managed by a separate system. It is embedded in the pipeline — each stage has a specific role in the chain mechanism. A-05 activates nominations. A-06 opens nominee intake. A-07 monitors the chain trigger. The chain grows through the same pipeline that processes every investment.
"The chain is not a referral programme. Every component runs through the standard pipeline. The only difference is that the disbursal of working capital tranches is conditional on nomination, and the release of nominees' investments is conditional on the referrer's break-even. Trust flows through the same system that processes every investment."