Can we trust AI to act on its own?→Governance and guardrails
Who allowed the AI to do that?→Identity travels with every request
Can someone talk the AI into something risky?→Risk-graded autonomy
Can we see who did what, and why?→One thread from request to result
Where are the keys kept?→One vault, no shared secrets
Why did the AI bill jump, and who spent it?→One counted doorway
Is every task getting the right model?→Right model, right task
What if our AI vendor goes down or raises its price?→No single point of supply
Are we paying for tools that do the same job?→One tool per job
Does the AI forget everything between conversations?→Governed shared memory
Is this producing real work, or just answers?→The digital factory
Which work should run itself?→Managed, not scripted
Can we get more from the Copilot licences we already pay for?→Copilot proposes, the platform verifies
Will people actually use it?→Meet people where they work
How much time do people lose sorting inbound information?→Intake to insight, untouched
Can the same platform run a real business?→One governed foundation for every product
What happens when it breaks at night?→Detect, fix, escalate
Where is the list of what is open, and who owns it?→One list, closed with evidence
Could we recover if we lost everything?→Proven by restoring
Can you prove it to an auditor?→Challenge before acceptance
Does it get better over time, or just older?→Every lesson becomes a rule
Will this still work after the next upgrade?→Clean Core for AI
Do our documents match what is actually running?→Documentation as a by-product
Portfolio / The journey
The program story
How one owner and a team of AI agents built an enterprise-style AI platform, and the decisions that shaped it.
The overviewSix AI questions, and how we answered themSee all six answers
In one screen
The goal was an AI platform used for real work and a working demonstration of how to lead AI transformation, as close to enterprise-grade as an employer would recognize, without big-company bureaucracy.
A fast build was followed by a run of “it looks fine but isn’t” discoveries, and then by governance. The lasting rule: nothing counts as working until it has been made to happen and watched.
How it evolved
An earlier project ran many agents on the same server; it produced activity, not reliable outcomes.
AI-drafted architecture documents multiplied and disagreed; the design was redrawn several times in days.
The agent, the websites, dashboards and the first backups came up, with quietly broken things hiding underneath.
A code repository became the record, backups were proven by restore, and controls had to show they were alive.
Audits with challengers, decision records, one task board and one home for documents.
Brain, executor and Microsoft 365 specialist meet at contracts, with self-healing and documents that rebuild themselves.
Decisions
Instead of: A second server or a second agent for guests
Kept the platform understandable and affordable; high availability was a conscious trade-off at this scale.
Instead of: Web chat and other channels with full access
Teams carries a verified person, which made the whole access model possible.
Instead of: Rented models through subscriptions
The server could not run useful models fast enough, so they were replaced by rented models with fallbacks.
Instead of: One brain directing an executor
It became the executor; the brain stayed in Hermes, and replaceability became the reason to keep both.
Instead of: Pay-per-use access
It saved real money but cost weeks of investigation into a confusing error.
Instead of: One reviewer's conclusion
Challengers corrected or overturned many first conclusions.
The build team
Claude was the builder, investigator and auditor; Hermes is the platform’s main agent and is taught every skill Claude uses; Microsoft Copilot served as an independent auditor; I owned the decisions. No single agent’s first answer held up on its own — the method did: verify by making it happen, research before building, and challenge every conclusion.
The full record stays in the private edition because it contains operational and security detail: decision records, incident write-ups, audit reports and the task board.
This story was checked by independent review lanes that started with no memory, by live tests of the critical processes, and by a challenger who tried to break the conclusions.
Contact
I would be glad to walk you through any part of this live and talk about the role you have in mind.
j.walters@erpaiintegration.com
Go deeper: Lessons · The journey · How it’s built · Live health
