Newsroom
Notes, methods, and points of view from the Galahad team.
-
Access is an institutional fact, not a prompt
How do you stop an AI agent from showing a user data they are not cleared to see? A system prompt, a guardrail or an audit log will not do it. Access rights belong to the institution, and the agent must inherit them, down to the individual field.
-
AI Act and DORA: what the deployer documents
A bank deploying an AI tool in a credit decision must document it three times: for the AI Act, for DORA and for the GDPR. What each file requires, who writes each part, and why the work is lighter when the vendor holds none of the data.
-
An agent that writes to the system of record
Can an AI agent be allowed to write to the ERP, the CRM or the ledger? Yes, on one condition: a named person sees the exact change before it is committed. Why a log is not a control, why bulk approval approves nothing, and why undo is a second write.
-
Answers across four systems, no warehouse
The answer to one question is spread across four systems, and the usual remedy is a data warehouse: a year of work, a copy to reconcile, and the question still open. How to get answers across systems without migrating anything.
-
Build or buy the agent platform
Build the agent platform in-house or buy one from a US vendor? Both leave the intelligence rented and the model of your business in someone else's hands. What each option leaves you owning after three years, and a third option.
-
Ontology, semantic layer, knowledge graph
Semantic layer, ontology, knowledge graph: three terms vendors use for what sits between an AI agent and your data. What each one does, what an agent needs from each, and why the ontology has to be yours.
-
The security architect's checklist
Eight requirements for any AI agent vendor before it touches a regulated system: where it runs, who operates it, whether it runs isolated, where the copies are, who approves. Each with a test you run yourself and the answer that should disqualify the vendor.
-
The trail that holds up
An AI system took part in a decision, and a regulator now asks how. What the audit trail must allow: name the person, see the data as it stood, replay the decision, trust the order of events. And three kinds of log that fail the test.
-
Why the pilot hallucinated
Most generative AI pilots deliver nothing measurable, and the model gets the blame. The failure is upstream: a copy of the business that was already wrong when the question was asked, and an answer nobody could trace. What a pilot that reaches production does differently.
-
Machines of Consequence
Sight, traceable power, and the continental century. Forty-nine theses on why the frontier that decides the next twenty years is consequence, and why the position everyone agrees is lost is merely vacant.
-
AI deployments that will not fail in 2026
95% of generative AI pilots return nothing measurable. The problem is not the model, it is rented intelligence: why an ontology you own is the only AI asset that stays yours.
-
Getting your money's worth
Consultancies, labs and software vendors bill hours, tokens and seats: three ways to charge at the door. Why Galahad bills only on the value created, payable on production.
-
Own the decision space: the case for sovereign, governed AI
Why the next advantage for governments and enterprises is sovereign, governed AI: agents that read across every data silo and act under human control, on-premise, with a complete audit trail.