AI playbooks
AI needs more than a prompt.It needs the rules of the work.
A business process is rarely just an instruction. It contains evidence requirements, decision rules, exceptions, permissions, stop conditions and people who remain accountable for consequential decisions. TEMRIK uses the term playbook for the structure that brings those operating rules together around AI-assisted work.
Controlled operating pattern
Playbook
Purpose · evidence · rules · exceptions · permissions
Agent
Assemble · check · prepare · recommend
Evidence
Current sources and required records
Proposed action
Structured output, not uncontrolled release
Human authority
Approve · reject · amend · escalate
The playbook provides the operating boundary. The model or agent remains one component inside that boundary.
The distinction
Prompting is not process design.
A prompt can tell a model what you want. It does not, by itself, establish evidence standards, corporate authority, tool permissions, exception handling or the conditions under which the system must stop.
Prompt
A request or instruction given to a model for a task or interaction.
Policy
A rule that constrains what is permitted, required, restricted or escalated.
SOP
Documented instructions describing how people or teams should perform repeatable work.
Playbook
An operational structure combining purpose, evidence, rules, exceptions, permissions, authority and audit for a defined class of work.
Workflow
The executable sequence or state transitions through which work moves.
Agent instruction
Configuration that tells an AI agent its role, objectives, constraints and available capabilities.
A prompt tells AI what you want.A playbook tells the business how the work should be done.
Playbook anatomy
Encode the work around the model.
The structure below is a TEMRIK architecture concept. It is intentionally provider-neutral: the operating logic should survive changes in model, agent framework or tool provider.
Purpose
What business outcome is this playbook for?
Trigger
What event or request starts the work?
Inputs
What information is allowed into the workflow?
Required evidence
What must be present before the work can progress?
Decision rules
What business logic determines the next step?
Exceptions
Which situations require a different route?
Tools
Which systems or capabilities may be used?
Permissions
What may each actor read, prepare, recommend or execute?
Stop conditions
When must the AI pause, refuse, fail safely or escalate?
Approver
Who owns consequential authority?
Output
What artefact, recommendation or action should result?
Audit requirement
What evidence must be retained to reconstruct the run?
Worked example
Construction variation playbook
Revised drawing received. What happens next?
A variation assessment is a useful example because the work is evidence-heavy, commercially consequential and often time-sensitive. AI can reduce preparation effort without inheriting the commercial manager's authority.
Trigger
Revised drawing or instruction received.
Human owner
Commercial manager.
Required evidence
AI may
- Assemble evidence
- Identify missing records
- Compare revisions
- Calculate or summarise impacts from supplied data
- Prepare a draft notice
AI may not
Issue the contractual notice without the defined human approval.
The exact boundary should be configured to the contract, customer policy and implementation scope.
From SOP to executable control
Capture
Document how the work is actually performed, not just how the SOP says it is performed.
Evidence
Identify mandatory inputs, records and provenance.
Rules
Separate deterministic business rules from model judgement.
Authority
Define what AI can read, prepare, recommend and execute.
Exceptions
Specify stop, escalation and missing-information paths.
Observe
Record material events, approvals, released actions and outcomes.
The goal is not to turn every procedure into autonomous software.The goal is to make the operating boundary explicit.
Decision rules
Keep deterministic controls deterministic.
Some decisions are better expressed as explicit software rules than inferred from natural-language instructions. Thresholds, required approvals, permitted tool scopes and mandatory evidence can often be enforced outside the model.
Model judgement
Classify, summarise, compare, extract, draft, reason over supplied context.
Policy rule
Allow, restrict, require approval, set threshold, route exception, deny.
Human decision
Accept commercial risk, waive a control, release a consequential action.
Audit event
Record who or what proposed, approved, changed and released the outcome.
Agent relationship
The playbook is not the agent.
Layer 1
Playbook
Defines the business operating boundary: evidence, rules, exceptions, permissions, authority and required outputs.
Layer 2
Agent
Uses model capability and tools to perform permitted reasoning and work inside the playbook.
Layer 3
Control plane
Applies identity, data access, tool rights, approval gates and audit around the agent and playbook.
See how this fits into TEMRIK's agentic AI architecture, the AI security boundary and the human control layer.
Evidence before action
A playbook can require evidence before the agent is allowed to progress.
This turns missing information into an explicit workflow state rather than an invitation for the model to invent, assume or silently continue.
Evidence complete
Continue within authority.
Evidence incomplete
Request missing information.
Evidence conflicts
Flag inconsistency and escalate.
Authority exceeded
Stop and request approval.
Tool unavailable
Fail safely or route to an alternative defined path.
Human authority
Approval is part of the playbook, not an afterthought.
Human review is most useful when the reviewer receives a structured evidence package, the proposed action, the rule that triggered approval and the exception state—rather than a vague request to “check the AI.”
Why this matters
Repeatability
The same class of work starts from the same operating structure.
Auditability
Required evidence, approvals and material actions can be reconstructed.
Safer autonomy
Authority can expand by workflow rather than by granting the model broad access.
Provider independence
Business rules remain company assets even when the model or agent framework changes.
Operating architecture
The TEMRIK AI control plane is designed to place company rules, evidence and human authority around models, agents and tools.
Research basis
Grounded in current agent architecture, risk governance and process-control sources.
These references support the distinction between model instructions, agent configuration, organisational governance and process design. They do not imply endorsement of TEMRIK or certification of this architecture.
Anthropic
Building Effective AI Agents
Distinguishes workflows from agents and argues for matching system complexity to the task.
Open primary sourceOpenAI
Agent definitions and configuration
Documents agent instructions, tools, guardrails, handoffs and structured outputs as explicit runtime configuration.
Open primary sourceNIST
AI RMF — Govern
Calls for AI risk policies, processes, procedures and practices to be implemented and documented across the organisation.
Open primary sourceISO
The process approach in ISO 9001
Describes processes as interrelated activities with inputs, intended outputs, checks, planning and controls.
Open primary sourceOMG
Business Process Model and Notation
Provides a standard notation for specifying business processes and bridging process design with technical implementation.
Open primary sourceTEMRIK
Public playbook repository
Public repository for TEMRIK playbook examples and operating-rule artefacts.
Open primary sourceRelated control layers
How TEMRIK works
See the wider operating model around data, AI and business action.
Agentic AI
Understand tools, orchestration, delegation, memory and bounded agency.
AI Security
See how data, identity, model access and action boundaries fit together.
Human Control
Define where people remain the decision authority.
Free field guide
27 Rules of Peace
Better AI starts with clearer operating rules.
TEMRIK's free field guide explores decision rights, escalation, evidence and keeping people in authority as AI enters real operational workflows.
Start with one workflow
Turn one SOP into an AI playbook.
Map the trigger, evidence, decision rules, tools, permissions, exceptions, stop conditions and human approval point before expanding autonomy.
Playbook execution, integrations and automation boundaries are implementation dependent. This page describes TEMRIK's operating architecture and does not claim that every illustrated control is universally deployed.