Governed AI
AI that computes before it speaks
CoaleTech builds AI for finance, audit, and operations leaders who cannot afford a wrong number. Every figure is computed deterministically in SQL and Python. The language model only narrates. Every action is classified green, amber, or red. Every invocation is written to an immutable audit log.
How auditable AI is engineered, not asserted
Deterministic compute first
Every figure starts in SQL and Python against your own ledger. The LLM never guesses a number; it only explains what has already been computed.
Governance gate
Before anything acts, a governance gate classifies it green, amber, or red. Green runs automatically, amber waits for a human to confirm, red is advisory only and can never write.
Immutable audit log
Every agent invocation, input, and output is written to an immutable Agent Action Log. Auditors see the full chain: who asked, what computed, what was gated, and what changed.
Graceful degradation
When AI is switched off or a provider fails, deterministic template rendering keeps the product running. Governance and compute do not depend on a model key.
The governance gate
Three classifications stand between the model and your data
Nothing the AI proposes reaches your records unclassified. The gate decides what may run on its own, what needs a person, and what the model is never allowed to touch.
Green: auto
Low-risk, reversible, fully deterministic. Runs without a prompt and lands in the audit log.
Amber: confirm
Material or hard-to-reverse. Proposed to a named human, applied only on explicit confirmation.
Red: review-only
Judgement, disclosure, or clinical calls. The AI advises; it is structurally blocked from writing.
Here is one amber action, end to end: the agent proposes, the gate holds it, a person confirms, the ledger posts, and every step is already in the audit log.
- 09:41agentproposepayment batch · 14 invoices · KES 1,412,000
- 09:41gateclassifyAMBER · human confirmation required
- 09:52controllerconfirmbatch approved
- 09:52ledgerpost14 Payment Entries · AAL-0918 · immutable
AI flagship
Close the books on time. Then let governed AI explain them.
Coale Finance runs a structured, checklist-driven, deadline-tracked close on ERPNext, then layers 30+ analytics agents that compute answers from your own data, classify every action green, amber, or red, and record everything in an immutable audit log. Built by finance engineers, not generic software vendors.
Structured period close
Checklist templates, task dependencies, SLA deadlines, and approvals keep the month-end close on track and auditable.
Automatic period lock
Back-dated postings into closed periods are blocked across accounting, stock, and payroll transactions.
30+ analytics agents
Cash flow, gross margin, anomaly detection, tax compliance, order-to-cash, procure-to-pay, and operational agents run on your data.
Morning CFO brief
A scheduled 06:00 health-scored digest lands in your inbox: close status, cash-flow outlook, anomalies, and key margins.
Four principles behind every CoaleTech AI feature
Sovereign, bring-your-own models
Use OpenAI, Anthropic, OpenRouter, Ollama Cloud, or a local endpoint. Your data stays under your control; model choice stays under yours.
Human-in-the-loop by design
AI is advisory. Your validated rules and your people stay the source of truth; any model-versus-computed mismatch is flagged for a human, never applied silently.
Deterministic and guardrailed
Narration runs at temperature 0 against fixed schemas. Every deployment adds usage quotas, response caching, and full audit logging.
Agentic domain specialists
Thirty-plus specialised agents cover finance, audit, manufacturing, supply chain, HR, and operations, all orchestrated under one governance layer.
In production
One pattern, many products
The deterministic-first, human-in-the-loop, fully logged discipline runs across our AI-enabled products. Different domains, different models, the same governance.
Deterministic compute → governance gate → optional governed write → LLM narration. Every step lands in an immutable Agent Action Log, across 30+ analytics agents.
AI finding reviewer and multi-agent control testing. Opt-in per engagement, with human sign-off before any conclusion stands.
Laboratory AI reviewer on your own infrastructure. Reads sample metadata, cross-checks against reference ranges, and flags discordant results without sending data off-premise.
Legal document review and clause extraction with a deterministic-first reviewer. Optional WhatsApp voice assistant for client intake and deadline reminders.
Task-scoped commerce agent inside WhatsApp. Catalogues, carts, M-Pesa payments, and order status queries stay inside lanes and hand off to humans when they leave them.
Three-layer recommendations: SQL, then SQL-ML (RFM and collaborative filtering), then LLM ranking. The maths is deterministic. The model only orders the shortlist.
Deterministic ML computes, LLM narrates. Complexity-based model routing, a daily AI quota, and a full AI Query / Usage audit trail.
See governed AI running on your own data
We will walk through the deterministic core, the green/amber/red gate, and the audit log on a live ERPNext environment.
FrappeVerse Africa 2026
Intelligent reports for top-level management
Our conference talk on putting AI-assisted reporting in front of executives: how numbers are computed deterministically in the ERP and narrated for leadership, the same discipline that runs the 06:00 CFO brief in Coale Finance and the 17 specialist agents in Insights v3.0. Recorded on the official Frappe channel.
Explore Insights v3.0