AI Operations
What happens after the AI answers
The AI gives you the answer. It is a good answer. Now what? Who approves it? Who executes it? Who records that it happened? And who is accountable when it goes wrong?
Adam Cowles1 March 20265 min read
The AI gives you the answer. It is a good answer. Now what? Who approves it? Who executes it? Who records that it happened? And who is accountable when it goes wrong?
The last mile problem
AI models are astonishingly good at generating answers. Claude can write a deployment script. GPT can draft a database migration. Gemini can analyse your error logs and suggest a fix. The intelligence is real and it is improving every quarter.
But between the answer and the outcome, there is a gap. The AI suggests restarting a service. Should it? The AI recommends changing a firewall rule. Who approves that? The AI generates a customer communication. Who reviews it before it goes out?
This is the last mile problem in AI operations. The model is the easy part. The infrastructure between the model's output and your production systems is the hard part. And almost nobody has built it.
Five things that need to happen
Between the AI producing an answer and that answer having an effect on your business, five things need to happen. Today, most organisations handle zero of them systematically.
- Policy check. Is this action within the agent's authorised scope? A customer service agent should not be modifying infrastructure. A code review agent should not be deploying to production. Scope boundaries need to be defined and enforced before execution, not discovered after.
- Human approval gate. Some actions should require a human to say yes. Not all of them. Not none of them. The boundary between "auto-approve" and "require approval" needs to be explicit and configurable. A refund under a hundred pounds: auto-approve. A refund over ten thousand: human in the loop.
- Structured execution. The action needs to be executed through a controlled channel, not pasted into a terminal by a developer who read the AI's suggestion. Controlled execution means the system knows what was done, by whom, at what time, with what parameters.
- Evidence recording. Every step of the chain needs to be recorded in a way that cannot be altered after the fact. Not a log file that can be edited. A structured, append-only, cryptographically linked record. The kind of record that survives an audit.
- Accountability chain. When someone asks "who authorised this?", there needs to be an answer. The AI suggested it. A policy allowed it. A human approved it (or the auto-approval rule did). The system executed it. Each link in the chain is traceable.
This is not about distrusting AI
The models are good. Claude is genuinely remarkable. GPT-4 changed what was possible. Gemini, Llama, Mistral, each brings something different. The intelligence is real and it deserves to be used.
But intelligence without infrastructure is advice without accountability. A brilliant consultant who whispers suggestions in your ear is useful. A brilliant consultant whose suggestions are automatically executed on your production systems, with no approval process, no audit trail, and no way to reconstruct what happened: that is a liability.
The answer is not to use AI less. It is to build the infrastructure that makes using it safe. Approval gates. Policy enforcement. Audit trails. Accountability chains. The same governance infrastructure that every other critical business process already has.
The neutral layer
The governance layer cannot be owned by any one AI provider. If Anthropic builds the governance for Claude, it does not cover your GPT integrations. If OpenAI builds it, it does not cover your local models. If Google builds it, it is designed to keep you in Google Cloud.
What is needed is a neutral layer. One that sits between your AI models (all of them) and your business operations. One that enforces the same policies regardless of which model produced the suggestion. One that records the same evidence regardless of which API the request came through.
This is the architectural bet behind Quox. Open protocols for agent messaging, governance, and evidence. An orchestration layer that is model-agnostic. An audit trail that does not care whether the decision came from Claude, GPT, Gemini, or a local model running on your own hardware. The trust layer is the same. The evidence is the same. The accountability is the same.
The cost of not building it
Without this infrastructure, organisations end up in one of two places. Either they slow down AI adoption to the point where it delivers no value, wrapping every AI interaction in so much manual review that they would have been faster doing it by hand. Or they move fast and accept the risk, deploying AI agents with no governance and hoping nothing goes wrong.
Both options are losing positions. The first wastes the technology. The second invites the 2am phone call.
The third option is to build the last mile. Policy gates, approval workflows, structured evidence, accountability chains. Not as an afterthought. As infrastructure. The same way you would not run a payment system without reconciliation, you should not run an AI agent without governance. The question is not whether your organisation needs this. It is whether you build it before or after something goes wrong.
Build the last mile before you need it.
Quox is a model-agnostic governance layer: policy checks, approval gates, and cryptographic evidence for every AI-suggested action, regardless of which model produced it.