Governed Execution: Managing Agentic AI — Article 6 of 13
The prompt is the starting point for most organizations, which is understandable. The prompt is visible and easy to alter. It feels like an instruction. When the output is disappointing, the first step is to improve the prompt by being more specific, adding context, clarifying the role, specifying the tone, listing rules, asking for citations, requesting step-by-step reasoning, and reminding the model not to invent. Sometimes this makes a difference. That is still not governance.
A prompt shows the system what the user desires. Governance determines what the system is permitted to do, what it must keep, what it must not do, what evidence it must provide, and when it must cease. These are separate responsibilities.
The difference is important since agentic AI no longer only generates text; it now carries out tasks. It searches, classifies, compares, summarises, ranks, recommends, invokes tools, and moves information as part of a workflow. In this situation, an improved prompt may enhance the instruction but does not necessarily ensure controlled execution.
The prompt is a request. The mandate is the control.
The prompt problem
Return to the stage-gate memo. A project team asks the agent to verify the requirements, link them to the approved source documents, identify gaps, and prepare a recommendation for the gate. A prompt can clearly outline this task by stating that it should use only approved documents, flag missing evidence, preserve mandatory controls, escalate unresolved dependencies, and avoid unsupported assumptions. That is a good prompt.
It does, however, create a problem for management. The prompt is generally used at the start of the task. It might affect the answer and influence the agent’s behaviour. Yet when action starts, the agent remains free to make various intermediate decisions such as which documents to retrieve, which excerpts to consider relevant, which ambiguities to resolve, which missing items to highlight, which exceptions to escalate, and how to turn all that into the final memo. That is where drift enters.
The more important the workflow is, the insufficient the original instruction becomes. Although a prompt may state the rule, it doesn’t show that the rule remained active throughout execution. The distinction lies between telling and governing.
Why longer instructions do not solve it
The right thing to do is to make the prompt longer by adding more rules, examples, forbidden behaviors, formatting requirements, warnings, a checklist at the end, language such as “do not proceed unless…”, and reminders that the work is important. Clarity is acceptable, yet poor instructions lead to poor work.
Lengthy prompts can create a false sense of security because thorough instructions feel like control. It might not be. A very long prompt can end up pasted into a text box and, although it may include the correct words, still fail as a control mechanism. The problem is not whether the directive was given once, but whether it was during execution as the work progressed.
That is how organizations proceed when engaged in serious work unrelated to AI. A project charter is not just an inspirational paragraph. A risk policy is not simply a reminder to be careful. A contract is not just a suggestion of desired behavior. These documents establish authority, boundaries, obligations, evidence, and remedies. They are more than just means of communication; they are tools of governance.
Agentic AI also requires this kind of change; the organization must go from merely improving the wording to exercising explicit control.
What the mandate must contain
A governing mandate goes beyond the task description. Simply being asked to ‘prepare the memo’ is not sufficient. ‘Validate the requirements’ is not enough. ‘Be accurate’ is also insufficient. It must specify the conditions under which the work will be considered valid.
Regarding the stage-gate memo, the objective is to validate the requirements for the governance committee; the acceptable sources are the approved project documents, not plausible external substitutes; the constraints are that the mandatory controls must continue to be mandatory; the rules on evidence state that any missing evidence must be noted as missing and not inferred; the escalation triggers are that unresolved dependencies must be brought to the surface; and the conditions for completion are that the memo is not considered complete just because it reads well, but it is only complete if the evidence and exceptions have been dealt with in line with the mandate.
A semantic contract is the kind of unit typically not included in the prompt itself. The organization does not require the model to follow an instruction; it needs an object the workflow can be directed by.
The semantic contract
A semantic contract is the explicit operating mandate for AI-executed work.
It converts the organization’s meaning into terms the execution process can use: it specifies the objective, the source boundaries, the required constraints, the evidence obligations, the escalation rules, and the completion criteria. It goes beyond being merely a better prompt; it is the governance object that links authorization to execution.
The term “semantic” matters because the relevant obligations are generally not simple numerical thresholds; instead, they are based on meaning. A requirement has to be linked to an approved source; a control must continue to be mandatory; a missing document must stay missing; if a dependency prevents validation it must be escalated; and a recommendation must not take the place of a committee’s decision. They are not simply matters of style. They are commitments that shape the work’s meaning.
The term “contract” matters because the object in question is not informal. It specifies the actions the agent is permitted to carry out and what the organization must later be able to inspect. It does not make the agent responsible in a human way; the agent still cannot care, respond, or take ownership of the results. However, it does provide the organization with a structured method of governing the agent’s execution.
A prompt asks for behavior. A semantic contract defines authority.
What changes when the mandate is explicit
When the mandate is made clear, the management question then changes. Without it, reviewers are left to judge whether the memo is plausible. Do the citations seem believable? Does the recommendation make sense? Did a human approve it? This lets reviewers ask a better question: whether the execution has preserved the mandate.
It doesn’t mean review vanishes; rather, it becomes less showy and more focused. Rather than asking Maya to approve a refined product without knowing how it was produced, the organization can show her the parts where the work remained within the mandate, the areas where evidence was missing, the places that tested the constraints, and the instances where escalation was needed.
This has a similar effect on how teams plan their AI workflows; instead of viewing governance as something added at the end, they now see it as something to establish before action is taken. The agent is not just instructed to “be careful”. It is given limited authority. The organisation beforehand determines what the system may resolve, what it must bring to the surface, and what it must not deal with. It is in this way that governance moves upstream.
The failure of prompt-as-policy
It is understandable, but dangerous, for many organizations to put policy into prompts and treat the problem as solved. Policy included in a prompt is fragile because adherence is not guaranteed. Context may weaken it, later instructions may override it, subsequent directives may dilute it, or the final output may only superficially satisfy it. The model might generate language that seems compliant even though execution departed from the policy earlier.
This is why endpoint inspection alone is insufficient. By the time the final memo is produced, significant substitutions may have already occurred. The absence of a source may have been resolved. A mandatory control may have been weakened. A recommendation may incorporate an exception. The final output may conform to the instruction’s surface while violating the work the instruction was meant to control. This is precisely why we should not treat the prompt as the means of control; the prompt is a component of the interface, not the governance architecture.
Governing before the agent acts
The first step in controlled execution is to establish the mandate before the agent takes any action; not after the draft seems suspicious, not after the reviewer requests support, not after an audit problem appears, but before execution starts.
It does not have to be a huge mandate. In many workflows, it can be concise. What matters is not creating bureaucratic showiness but determining what must remain true for the work to be authorized.
- What is the objective?
- Which sources are allowed?
- Which constraints cannot be weakened?
- What counts as missing evidence?
- What must be escalated?
- When is the work complete?
These are not prompt-writing questions. They are control questions.
Such questions are standard management questions. Agentic AI makes them urgent since the individual acting can now act swiftly without passing the matter through meetings, drafts, and reviews.
The organization doesn’t have to slow everything down; it just needs to know what must not be lost as work speeds up.
From Act I to Act II
The first five articles identified the issue: AI is turning into an executor. Useful output does not have to be governed output, and execution can diverge from the mandate. It is impossible to motivate the agent to take responsibility. When human oversight comes too late to see much, it ends up absorbing the blame. This is the Act I problem.
The Act II answer begins here: governance must move upstream of the output. The organization must define the work before the agent performs it. It must turn implicit expectations into an explicit mandate. It must stop treating the prompt as enough to govern delegated machine execution. Prompts will still matter. Clear instructions will still matter. Better interfaces will still matter. But when AI executes consequential work, prompting is not the discipline. Governance is.
For Assiduity, this is where governed execution begins. The semantic contract is not an improvement to the prompt. It is the operating mandate that lets an organization define what the agent is authorized to do before the work begins. Without that mandate, the firm is left trying to govern execution from the final artifact. With it, the work is measured against the authority that allowed it to proceed.
The next article turns to the evidence that makes this governance usable. Once the mandate is explicit, the organization still needs to see whether execution preserved it.
Next: Process Evidence Is the New Management Control.
Part of Governed Execution: Managing Agentic AI — a series on the management discipline required when AI executes work, but firms still answer for it.