Free download · reference model · no email required
Most AI programs stall at pilots. This free, five-document reference model shows owners and executives of 1–500-person organisations what to build, in what order, under which controls, and how to prove the value—without the hype.
Document 01 · target architecture
The architecture document answers the hardest question first: what sits between a model and a business transaction? A six-plane reference model wraps a governed execution fabric around the SaaS systems you already run—without replacing them.
the six planes
Experience & collaboration; workflows, agents & action control; models, context & evaluation; data, knowledge & integration—with identity, security and governance cutting across all of it. Sized for 1–20, 21–100 and 101–500-person organisations, with Google- and Microsoft-anchored reference stacks.
the safety boundary
Models emit typed proposals. Deterministic policy checks, authenticated approvals and an idempotent executor decide what actually reaches your ledger, your CRM, or your customers. Systems of record stay authoritative—always.
execution modes are assigned per action class—and earned
Read permitted sources; search, classify, summarise. No writes.
Produce the draft or recommendation; a person independently commits it.
Execute the exact typed action an authorised person approved.
Exceptional: narrow, reversible, low-impact actions inside hard limits, continuously monitored.
Includes the 18 architecture decisions every adopter must record—so the model becomes your model, on the record.
Document 02 · phased implementation plan
The plan runs up to 24 months, calibrated to readiness and risk—a small firm may reach its steady state in 6–12. Each phase must earn the next: baselines before build, evaluation before reliance, sustained evidence before autonomy.
a portfolio, not a parade of pilots
A weighted selection framework, a reference catalogue of 15 starter workflows with control ceilings, and kill criteria agreed before anything is built. A workflow can be retired at any gate—failed experiments stop consuming licences.
start monday morning
Appoint the owners, inventory the AI already in the building, fix identity and sharing basics, baseline 10–20 task families, and ship one or two low-risk workflows through evaluation and shadow mode—then hold the first evidence gate.
With readiness assessment, resourcing and budget guardrails by company size—calendar progress never substitutes for gate evidence.
Document 03 · data, security & AI governance
Governance exists to make useful adoption repeatable and safe. The governance document keeps control effort proportional to consequence: a policy and register set an SME can actually run, mapped to current OWASP, NIST and ISO practice.
Every sanctioned AI system, data source and production workflow has a business owner and an operational owner.
Deterministic policy and transaction controls sit between model reasoning and any side effect.
A four-level classification, permission hygiene, minimisation and retrieval controls—credentials never enter a prompt.
Approvers have competence, context, authority, time and a practical way to intervene—no rubber stamps.
Controls for prompt injection, tool misuse, data poisoning and runaway cost, aligned to the OWASP LLM and agentic Top 10.
No-blame reporting, containment authority, drills—and 2026 legal watchpoints for Australia and the EU.
Registers, intake templates and checklists included—a spreadsheet is acceptable; keeping it current is the control.
Document 04 · benefits realisation
The benefits document is the discipline most AI programs skip: baselines before business cases, capacity and cash in separate ledgers, benefits banked only when a bill, contract, licence or planned hire actually changes—and the full cost of ownership counted, not just the licences.
separate ledgers
Hours released are reported as capacity and allocated once—to growth, quality, or a documented workforce route. They become cash only through evidenced conversion: reduced external spend, avoided hiring, retired software. Finance co-signs what gets banked.
a worked example
A fully worked example with conservative, base and stretch cases—including a plausibly negative Year 1. Plus fifteen benefit claims a finance team should reject on sight, starting with “minutes saved × every employee”.
Published productivity effects span roughly −20% to +35% depending on task, user and process fit. The guide refuses to plant a generic number in your business case—it shows you how to earn your own.
Free download · no email required
Plain markdown, deliberately. Version it, redline it, ground your own AI assistant on it—the model is only useful once it becomes yours. Document 00 orients the set and includes a minimum viable pattern for firms of 1–20 people.
the complete set
The full reference model in a single markdown file—easy to read, search, print, or share with your board.
Written for organisations of 1–500 people on ordinary SaaS estates. Grounded in current good practice as at August 2026—NIST AI RMF, ISO/IEC 42001, OWASP LLM & agentic Top 10, ASD guidance and the EU AI Act timeline—a reference model you adapt and own, not legal advice.
From the model to the machine
Governed memory, explicit mandates, and one execution surface across the systems you already use—the architecture the guide describes, as company-controlled infrastructure. We are speaking with founders, executives, and boards now.