Mothusi uses AI to make every operator-to-platform interaction productive. The MGS framework, not the model, decides what those interactions mean. This page documents the boundary, the commitments, and the audit posture that lenders, DFIs, and institutional partners read before procurement.
Written for institutional procurement, data-governance teams, and the methodology committee.
Every signal carries an evidence grade. The operator sees it. The funder sees it. The programme officer sees it. The same record reads the same way to everyone.
Recorded as a self-reported figure for March. The growth record will not display this as verified revenue until a bank statement or invoice trail is shared. The MGS evidence grade stays at self-reported for now.
Each commitment is visible in product behaviour and citable in procurement reviews. Programme officers, funders, and operators can verify each one against the live platform, not against a slide.
Mothusi is the conversational layer that captures and structures evidence. MGS is the published framework that grades that evidence and assigns stage. The model never freely invents a stage, a confidence grade, or a risk band. Scoring is rule-driven, citable, and reproducible without the model in the loop.
Every signal carries an evidence-confidence grade visible on the operator's growth record. A claim made in conversation is not equivalent to a document, a live data feed, or a third-party verification. Funders, programme officers, and operators read the same record and see exactly which signals are verified and which are not.
Stage promotions, funding-readiness signals, and risk band changes surface to programme officers and operators for review. No consequential decision is finalised inside the model alone. Mothusi flags. Humans confirm. The audit trail captures both.
Every signal Mothusi captures about a business is visible to that business. The operator can review their growth record, dispute incorrect evidence, request re-verification, withdraw consent to share, and remove data they did not consent to capture. Transparency is the default, not an opt-in.
We monitor model behaviour across language, region, business stage, and operator gender. Discrepancies are reviewed by the methodology committee. The bias-monitoring audit log is open to institutional partners under contract and summarised in the public methodology release notes.
Operator data is not used to fine-tune underlying foundation models. Mothusi uses context engineering and retrieval, not retraining. Where institutional partners contract for bespoke evaluation on aggregate, de-identified data, that is a separate written agreement with explicit, revocable consent.
The boundary between AI and human judgement is operational, not rhetorical. Below is the specific division of labour for the consequential decision types on the platform.
| Decision area | Mothusi does | Mothusi does not |
|---|---|---|
| Scoring and grading | Mothusi captures the signals; the MGS framework, not the model, assigns the grade. | No model invents a stage, a confidence grade, or a risk band. The rules are published. |
| Funding decisions | Mothusi prepares the readiness pack and surfaces the missing evidence. | No model decides whether to lend. Funders decide, with the operator's consent to share. |
| Compliance verdicts | Mothusi tracks document presence, expiry windows, and renewal nudges. | No model issues a compliance certificate or replaces a regulator's verdict. |
| Risk signals | Mothusi surfaces forward-looking signals against the evidence the operator has shared. | No model assigns a credit score, predicts default, or substitutes for a credit bureau. |
| Programme matching | Mothusi recommends interventions based on the operator's current MGS state. | No model enrols an operator into a programme without programme-officer review. |
| Learning and content | Mothusi tailors guidance, summaries, and drafts to the operator's context. | No model awards a credential, certifies completion, or grades a third-party assessment. |
Mothusi is not a black-box scoring model. It is a transparent evidence layer governed by a published framework, an audit trail, and human review at every consequential decision.
Institutional users do not buy AI promises. They buy a posture they can review and a record they can defend. Four operating commitments back the audit story.
The MGS framework is open to read in full. Every release carries a version, a changelog, and a methodology-committee minute. Operators and institutions can audit the rules that produced any score.
Where Mothusi outputs influence a consequential decision, the contributing signals, their evidence grade, and the framework rule applied are written to a structured audit trail on the operator's record.
The growth record belongs to the operator. They can export it, share it under consent, and revoke that consent. The record is portable; the model is not what makes it credible.
For lenders, DFIs, and regulated procurement teams, a model-risk-management briefing covers model selection, evaluation methodology, prompt and context engineering, fall-back behaviour, and incident response.
Responsible AI is one of four trust pillars. The others are data and security posture, methodology governance, and open citation. Each has its own page.
From business support to measurable enterprise development. Across sectors, countries, and real operating environments.