What is the Agent Operations Framework?
AI Operations is running a business's operational systems with AI embedded in them, with named ownership, measured outcomes and a human accountable when something goes wrong. The Agent Operations Framework scores that maturity across five layers to produce an Agent Sprawl Index from 0 to 100, so you always know where you stand.
Why AI operations fails in most businesses
The technology works. The implementation works. What fails is sustained attention and accountability. Most businesses have no owner for their AI, no idea what it costs, and no way of knowing whether it still works. This framework exists to make that visible, measurable, and fixable — in that order.
The five layers
Each layer scores 0 to 5. The five scores combine into a single 0–100 Agent Sprawl Index.
"What AI is in use, who owns it, what it costs monthly."
Operational Context
Shadow AI and fragmented software acquisition create massive operational risk. Teams rapidly procure standalone LLM accounts, custom GPT wrappers, automated webhooks, and third-party SaaS extensions to solve local friction. Without a centralized register, the leadership team loses visibility over operational expenses, data flow, and system access points.
A complete inventory establishes line-item transparency for every agent, script, API connection, and vendor subscription. It reconciles active seats against operational output to eliminate redundant billing.
"nobody can list what is running."
"every tool and agent has a named owner, a monthly cost and a stated purpose."
"tools bought personally on individual credit cards, invisible to the business."
"Is it still working, and how would you know?"
Operational Context
Automated workflows and autonomous agents degrade over time as underlying API structures change, prompt instructions drift, and input formats evolve. When error handling is missing, automation pipelines fail quietly. Upstream tools continue passing corrupted inputs to downstream CRM systems and database tables without raising alerts.
Operational integrity requires automated health checks, routine response audits, output verification logging, and designated staff trained to intervene when model performance drops below acceptable benchmarks.
"nobody has checked since it was set up."
"there is a review cadence, known failure modes, and an escalation path."
"silent degradation. The automation still reports success while quietly returning wrong answers."
"Who is accountable, and what does policy require?"
Operational Context
Governance transforms high-level compliance principles into operational ground rules. When autonomous systems process customer information, modify live records, or send external correspondence, clear ownership lines are required. Unclear authority leads to legal liability, customer friction, and systemic data leaks.
Effective governance defines clear risk tiers, operational rollback procedures, data retention policies, and explicit authorization thresholds for model deployment and decommission.
"no policy, no owner, no record."
"documented decision rights, risk tiers, and a defined answer to who decides when a system is switched off."
"policy exists on paper but no one has operational authority to enforce it."
"What should be automated, and what must stay human?"
Operational Context
Premature automation locks broken processes into digital stone. Standardizing underlying SOPs must always precede technical execution. High-stakes actions—such as financial approvals, client contract changes, and sensitive escalations—require explicit human-in-the-loop controls to maintain strategic context and service quality.
Orchestration maps complete process pathways, setting clear handover rules between automated agents and human operators to ensure smooth operations without compromising oversight.
"automation applied wherever it was easy."
"explicit boundaries, with human approval points at anything consequential."
"automating a decision before the process underneath it was ever standardised."
"Is the team actually using it, or is it shelfware?"
Operational Context
Software license acquisition does not equal operational transformation. Without structured onboarding, workflow integration, and ongoing performance measurement, internal teams revert to legacy manual workarounds. Modern AI investments become expensive shelfware that bloats monthly recurring software budgets.
Adoption monitoring measures weekly active workflows, output completion speed, and user feedback loops to verify that technical deployments yield tangible operational efficiency.
"licences paid for, nothing used."
"measured active use, with adoption tracked over time."
"tools deployed and never adopted, quietly costing money every month."
The Agent Sprawl Index
The five layer scores combine into a single 0–100 figure.
- Most tools known but nothing reviewed
- No one accountable on paper
- Automation applied where it was easy
- A third of licences unused
The score is not a grade, it is a map. It tells you what to fix first.
"A twelve-person property management company scoring Inventory 4, Integrity 2, Governance 1, Orchestration 2, Adoption 3 — producing an index in the twenties."
What your score means
0–25 Unmanaged. AI is running, nobody is accountable, and the cost is unknown.
26–50 Inconsistent. Some ownership exists, but coverage is partial and nothing is reviewed.
51–75 Governed in part. Ownership is assigned and some controls exist, but measurement is patchy.
76–100 Operated. Every agent has an owner, a cost, a review cadence and a documented failure mode.
Get your score
The scorecard is free and takes about five minutes. You answer twenty-five questions — five per layer — and receive a scored readiness map plus a prioritised 90-day roadmap by email.
Why this framework exists
This framework is mine. I use it in every engagement, and it is what my work is measured against. It is published openly because a diagnostic you cannot inspect is a sales document, not a method. Run your own estate through it — if you score above 75 without my help, you do not need me.
Framework questions
Find out what your AI estate is actually costing you.
Book a 30-minute call. We diagnose first, then match the engagement to what we find.