Dear {{ first_name | reader }},
Over the last week I spent most of my time building something most consultants haven't thought about yet. Not a new service. Not a proposal.
A sovereign AI infrastructure — self-hosted, EU-based, and with no Big Tech dependency — so that when I use AI tools on client work, I can demonstrate exactly what runs where, what touches what, and what never leaves the secure system. Full transparency. Documented accountability. A transparency card for my own practice is in effect.
I'm a solo consultant. If I can build this, large organisations have no excuse for not having thought about it.
And then, while I was deep in server configurations and cron jobs 😵💫, a 91-page governance framework landed from CDAC Network, the Alan Turing Institute, and Humanitarian AI Advisory.
SAFE AI — the Standards and Assurance Framework for Ethical AI. Built for humanitarian organisations deploying AI in refugee camps and conflict zones.
Nobody in the private sector communications world seems to have noticed.
That's the story for this week.
Enjoy!
WAG THE DOG NEWSLETTER | ISSUE WEEK 23, 2026
KEY TAKEAWAYS
The humanitarian sector built a governance framework you should steal. SAFE AI was designed for refugee camps and conflict zones. The accountability logic applies identically to any organisation running AI systems that affect people. The operating context is different. The core question is not.
A Transparency Card is not a policy document. It is a living record – updated when the model changes, when risks shift, when incidents occur. Your responsible AI policy is not this. Your vendor checklist is not this. Most organisations do not have this.
The cost of inaccuracy matters as much as accuracy. Section 2 of the SAFE AI framework forces you to calculate not just whether your AI system works but also what happens to the people on the receiving end when it does not. Most private sector deployments skip that calculation entirely.
Accountability needs a name, not a function. A shared inbox is not accountability. A team is not accountability. The SAFE AI framework demands a single named individual whose job it is to be called when the system fails. If that name does not exist in your organisation, that is the gap.
Legal permissibility and reasonable expectation are not the same question. Your legal team will tell you the data use is covered by the terms of service. The communication question is different: if the people whose data trained this model knew exactly how it was being used, would they consider it fair? The gap between those two answers is where your reputational exposure lives.
Table of Contents
What is the SAFE AI Framework and why does it matter?
Somewhere in a refugee camp in northern Kenya, a humanitarian organisation is using an AI system to help determine who gets food assistance. The system makes recommendations. Staff act on them. And until last month, there was no shared sector standard for what happens when it gets it wrong.
That changed last May.
CDAC Network, The Alan Turing Institute, and Humanitarian AI Advisory published SAFE AI — the Standards and Assurance Framework for Ethical AI. Ninety-one pages. Two years in development. Fieldwork in Kakuma Refugee Camp with 193 community members across structured focus groups. Eight communities of practice. A roundtable at the UK Foreign Commonwealth and Development Office attended by ninety people from thirty institutions. Five authors spanning humanitarian leadership, AI accountability, responsible AI consultancy, and international policy.
The humanitarian sector built this because it had no choice. Its legitimacy depends on accountability to the people it serves. You do not deploy AI that shapes access to food assistance and then discover, when something goes wrong, that you have no governance architecture.
Read it. Then ask the question the framework forces you to ask. Does your organisation have the equivalent?