The Executive Guide to Agentic Operations

The problem: automation everywhere, ownership nowhere

Automation tends to arrive bottom-up. A capable person builds a workflow to save their team time. Another team buys a tool. Someone connects an AI model to a shared inbox. Each step is sensible. Together they create an operation nobody can see whole.

The symptoms are recognisable. Nobody can list every automation that touches customer data. When something breaks, the person who built it is the only one who can fix it. AI tools are used individually, with no shared record of what they did. And when leadership asks whether any of it is paying off, the answer is a collection of anecdotes.

Adding agents to that picture without changing it multiplies the problem. Agents act, at speed, across systems. The question is no longer only "does this tool work?" but "who decided this agent may do that, and what happens when it is wrong?"

Concept: what agentic operations means

Agentic operations is a way of running work in which people and AI agents carry out missions together inside an operating layer that makes ownership, boundaries and evidence explicit. Three ideas carry most of the weight:

• Missions, not tools. Every piece of agent work exists to achieve a defined outcome for someone, owned by a named person.
• Boundaries before power. Every agent has a role, an identity, approved tools, explicit limits, evaluation criteria and a named person accountable for it before it acts. The Guild deploys every agent this way, with a named Human Wizard.
• Autonomy is earned. An agent's freedom to act grows only as evidence of its performance accumulates, and only by a person's decision.

The most important move is from stage 2 to stage 3, because that is where agents start to act rather than suggest.

When to move

You are ready to move to governed agents when:

• a workflow is frequent, costly when it goes wrong, and owned by someone who wants it fixed;
• the decisions in it can be written down as rules, thresholds or named approvals;
• the data it needs is accessible, permitted and reasonably clean;
• you can measure today's performance, so you will know whether the agent improved it.

If none of your workflows meets these conditions, the first job is process design and data access, not agents. Choosing the First Automation helps pick the candidate; if you are unsure where you sit on the map, the Guild Assessment is a guided starting point, after which a Human Wizard reviews your answers and replies with a written brief.

Architecture: the operating layer

Agents need an operating layer around them in the same way employees need an organisation. In the Guild that layer is the Citadel, organised in eight layers. The names are ours; the functions apply to any organisation.

1. Mission — outcomes, scope, owners, milestones.
2. Wizard — human direction and approvals.
3. Agent — roles, tools, memory, limits.
4. Knowledge — approved context and sources.
5. Integration — systems, APIs and credentials.
6. Guardrail — permissions, policies, escalation.
7. Observability — health, cost, outcomes, audit.
8. Improvement — evaluation and optimisation cycles.

A leader does not need to know how each layer is built. A leader does need to know that each exists, and who owns it.

Components: the roles around every agent

The layers are held together by people. Every mission needs the same small set of roles, written down:

• Executive sponsor — sets priorities, chairs the portfolio review, and decides on autonomy changes above an agreed level of risk.
• Mission owner — accountable for the outcome; usually the person who runs the process today.
• Accountable reviewer — approves deployment and reviews exceptions. In the Guild this is a named Human Wizard.
• Builders — design, integrate and evaluate the agent, and change it when the evidence says so.
• Approvers — the people who make the decisions the agent routes to them, each with a backup.
• The agent — carries out its mission within its boundaries and records what it did and why.

In a small organisation one person may hold several of these roles. That is fine. What matters is that every role is held by someone, and that an agent is never the reviewer of its own work.

Implementation: one mission, then a portfolio

Six decisions only leaders can make

None of these should be delegated to a vendor or left to the people building the agents.

1. Which missions. The few workflows where agents should start, chosen for value, feasibility, risk and readiness.
2. Who owns each one. One accountable person per mission, with a backup.
3. What stays human. The decisions that always reach a person — money, commitments, personal data, exceptions to policy.
4. How far each agent may go. Its level of autonomy: observe, assist, operate within limits, or optimise (The Citadel Trust Model).
5. What evidence counts. The measures and thresholds that decide whether an agent is ready, and whether it may do more.
6. How the portfolio is reviewed. The rhythm and forum in which evidence is examined and autonomy widened, narrowed or withdrawn.

The gates

Begin with one mission, not a strategy document, and take it through the same gates every time:

1. Discover — map the workflow, people, systems, risk and value.
2. Design — define the agent's role, tools, knowledge, boundaries and human checkpoints.
3. Build — create the workflows, integrations, evaluations and controls.
4. Integrate — connect to approved systems, data, identities and people.
5. Deploy — release with monitoring, escalation, audit and operating playbooks.
6. Evolve — measure outcomes, review exceptions, improve, and raise autonomy deliberately.

Each gate ends in a decision with an owner and evidence. When the first mission runs well, keep its operating habits — owners, boundaries, approvals, reviews — and reuse them for the second. By the third or fourth you are managing a portfolio, and one operating view becomes essential.

Evaluation: what to ask to see

An executive does not need to read evaluation logs. An executive should ask for a scorecard that answers four questions:

• Is it right? Task success, factual accuracy, citation quality, retrieval accuracy.
• Is it safe? Tool selection and execution, escalation accuracy, policy adherence.
• Does it run? Hallucination rate, exception rate, latency, cost per mission.
• Does it help? Human intervention, customer satisfaction, business outcome, regression results.

In the Guild, Athena owns the mission definition, Merlin the experiments, Oracle the scorecard and Sentinel the boundaries; a Human Wizard owns the decision to deploy. Whatever your structure, each of those responsibilities needs a name. Evaluation Kit 001 sets out the full scorecard.

Security and governance

Six controls make agent autonomy visible, reviewable and reversible:

• Identity and access — role-based authority, isolated credentials, least privilege, revocation.
• Approved knowledge — source ownership, freshness, permissions, citations, retention.
• Controlled tools — allowlisted systems, bounded actions, safe defaults, test environments.
• Evaluation — quality thresholds, policy tests, drift signals, outcome measures.
• Human escalation — clear approval gates, exception routes, decision records.
• Audit and improvement — action history, cost visibility, incident learning.

These are also what procurement, auditors and boards will ask about, and what public frameworks such as the NIST AI Risk Management Framework and ISO/IEC 42001 expect an organisation to be able to show. Being able to hand over an architecture overview, an access matrix, an evaluation report, an operating runbook, a vendor register and audit evidence is a sign that governance is real rather than declared. A board-ready AI control model shows how to present it.

Cost considerations

The cost of agentic operations has four parts:

• Build — design, integration and evaluation for each mission.
• Run — model usage, infrastructure and monitoring, which grow with volume.
• Oversight — the time people spend approving, reviewing exceptions and examining evidence.
• Change — re-evaluating when models, prompts, tools or data change.

Oversight is the cost most easily left out of a business case, and the one that falls as agents earn autonomy. Measure cost per completed mission rather than cost per call, and compare it with the cost of the same outcome done the old way.

Example

Illustrative example, not a client engagement.

A regional distributor has a dozen automations built by different people, a sales team using an AI assistant, and a backlog of order exceptions handled by email. Leadership starts by listing every automation with an owner; the ones nobody will own are switched off. It writes an AI use policy and moves the assistant onto approved knowledge.

It then picks one mission: preparing order exceptions for review. An agent gathers the order, stock and customer records and proposes a resolution; a person approves each one. The mission owner reviews the decisions monthly, and when the evidence supports it, leadership decides whether routine exceptions below an agreed value move to exception-only approval. Whatever it decides, the operating rhythm — owner, boundaries, approvals, review — is ready for the second mission.

Failure modes

• Starting with the platform. Buying an agent platform before choosing a mission.
• Pilots without owners. An initiative owned by an innovation team has nobody to carry it into daily operations.
• Autonomy by default. Agents given broad access "to see what they can do".
• Activity metrics. Messages sent and tasks touched reported as results.
• Evaluation once. A launch test that is never repeated as things change.
• Invisible portfolio. Nobody can list every agent, its owner and its limits.

Checklist

• Every automation and agent has a named owner and a record.
• There is a written list of decisions that always stay with a person.
• The first mission was chosen for value, feasibility, risk and readiness, and has a baseline.
• Each agent has a role, identity, approved tools, explicit limits, evaluation criteria and a named human.
• Each agent's level of autonomy is recorded, with the evidence needed to change it.
• A scorecard covers quality, safety, operations and business outcome.
• Leadership reviews the portfolio on a fixed rhythm.
• You could hand an auditor the access matrix, evaluation report and audit evidence today.