These principles sit underneath the AI-ICM's six non-negotiables. The non-negotiables apply to the service you specify. These working principles apply to how you produce the specification. Select each principle to see what it means in practice and a quick self-check.
A principle says "we will be transparent." A specification says exactly how transparency works: in which channels, with what controls, against what standard, signed off by whom.
The course produces specifications. Any sentence in your work that could sit unchanged in a values statement is too abstract.
Could a vendor build from your draft, or only quote it in a values statement?
The Domain Audit Map in Week 1 sets the tone. Audits that produce comfort do not produce useful specifications.
Record what is actually happening: the AI you are using, the governance you are not, the exposure you have been glossing over. Discomfort in the audit is the work, not a sign of failure.
If a peer team read your audit, would they say it is honest, or polite?
Regulatory references name the framework and the section. Evidence references name the source. Practice descriptions name the tool, the team, and the context.
Where AI helped you produce the work, you note it. The same standard your specifications will require of the services you design.
Pick three claims in your draft at random. Can you point to the source for each?
Source visibility, consent-first personalisation, accessibility at display, abstain-and-escalate, risk-and-action pairing, change control.
Your plan can require more, and often will, where your sector demands it. It cannot require less. A draft that falls below the floor is incomplete, not minimalist.
For each of the six, can you point to where your draft satisfies it?
Interventions can widen gaps as easily as they close them. The framework calibrates to that risk.
Equity in your specifications shows up as operational rules with measurable outputs: segmentation criteria, frequency caps, accessibility checks at display, equity-of-reach indicators. Values statements about equity do not survive an audit. Operational rules do.
Show your equity claim to a sceptical colleague. Can they tell what it requires you to do, or only what you believe?
Many AI-enabled services launch with a plan for day one and nothing for day two hundred. That is how drift happens, and drift is what fails an audit.
Every plan you produce in this course has an integrated review cadence: drift detection, indicator reporting, feedback loops, crisis protocol. Without it, what you have built is a launch document, not a governed system.
If your service ran for six months without a launch event, what would catch the drift?