Report Title: Human-Directed AI Agents and AI-Enabled Workflows Guidance Report
Version: 1.1 | September 2026
Central Proposition: AI agents should operate as directed digital collaborators. They may gather, organise, draft, analyse, and complete approved tasks within defined boundaries. Human staff retain accountability, exercise judgment, approve significant outputs, and control consequential actions.
Executive Summary
Specialist consultancy agencies depend on knowledge work: gathering client information, analysing evidence, preparing deliverables, coordinating projects, and communicating recommendations. AI agents and AI-enabled workflows can reduce the burden of repetitive and time-consuming work, but their value depends on disciplined task selection, clear operating boundaries, and visible human responsibility.
Recommended Operating Rule: Humans set objectives and boundaries. AI prepares or performs the assigned work. Humans verify evidence, apply judgment, approve consequential outputs, and remain accountable for use.
Key Recommendations
- Begin with work and workflows, not fashionable tools or artificial job titles.
- Classify every task as AI-executed with review, AI-supported and human-led, or fully human-led.
- Use narrow, descriptive agent scopes and explicit “do not act” boundaries.
- Keep consequential approvals, commitments, and professional judgments with authorised staff.
- Curate knowledge sources and assign an owner, review cycle, and version to each source.
- Test agents and workflow components separately before connecting them.
- Introduce autonomy gradually and apply least-privilege access.
- Maintain an agent register, decision log, change control, and retirement process.
- Connect validated specialist agents through a hub or handoff only when cross-domain demand justifies it.
1. Purpose and Scope
This report is a reference and implementation guide for a specialist consultancy agency. It applies to internal operations and client-facing delivery where AI may assist with research, project coordination, evidence synthesis, document preparation, quality checks, knowledge retrieval, and routine communications. It treats AI as a managed capability within a human-led service model.
| Audience Group | Primary Focus |
|---|---|
| Agency Leaders & Service Owners | Strategic oversight, service accountability, and portfolio governance. |
| Consultants & Project Managers | Day-to-day workflow integration, evidence review, and quality control. |
| Operations, Finance & Support Teams | Task selection, process execution, and administrative automation. |
| AI Agent Makers & Designers | Bounded agent creation, prompt architecture, and system integration. |
| Governance, Security & QA Teams | Access control, testing verification, audit logging, and risk mitigation. |
2. Operating Philosophy: Directed Digital Collaboration
2.1 The Human Role
- Define the intended business outcome.
- Set the agent’s scope, authority, and exclusions.
- Provide or approve source information.
- Decide the acceptable level of automation.
- Review evidence, assumptions, and uncertainty.
- Approve consequential outputs and external use.
- Monitor performance and intervene when needed.
- Remain accountable for decisions and service quality.
2.2 The AI Role
- Retrieve authorised information.
- Organise and classify supplied material.
- Draft standard content and working documents.
- Summarise evidence while identifying gaps.
- Perform defined analytical or administrative steps.
- Recommend options using stated criteria.
- Execute approved low-risk actions through permitted tools.
- Escalate ambiguity, conflicts, and out-of-scope requests.
The Accountability Boundary: Delegating a task to an agent does not transfer accountability. Organisational users remain responsible for reviewing, validating, and approving AI-supported work. Describe agents by the service they provide (e.g., “Proposal Quality Check Agent”), avoiding titles that imply professional or managerial office.
3. Task and Workflow Inventory
Mapping required service work prevents agent portfolios from being driven by technology availability rather than operational need.
| Too Broad Definition | Better Task Definition |
|---|---|
| Manage client delivery | Compile the weekly engagement status pack from approved project records. |
| Lead marketing | Turn an approved case study into three channel-specific draft posts. |
| Handle finance | Check an expense submission for missing mandatory fields. |
| Do research | Create a cited evidence summary against the agreed research question. |
| Manage quality | Check a draft report against the approved deliverable checklist. |
Recommended Task-Register Fields
- Task & Intended Outcome: Defines the work and what completed means.
- Process Owner: Names the accountable human.
- Frequency & Volume: Shows repetition and potential value.
- Inputs & Sources: Identifies required information.
- Output & Recipient: Clarifies use and audience.
- Current Effort / Pain Point: Explains the improvement opportunity.
- Data Sensitivity & Impact of Error: Supports access control and shows potential consequences.
- Ease of Verification & Review Point: Clarifies whether review is practical and makes human control explicit.
- Proposed AI Role: Records the intended level of support.
4. Task Suitability Assessment Framework
| Criterion | Questions to Ask | Favourable Signal |
|---|---|---|
| Repeatability | Does the task follow a stable pattern? Are inputs and outputs recognisable? | A standard process occurs frequently. |
| Impact | What happens if the output is wrong, incomplete, or late? | Consequences are limited or controlled by review. |
| Error Detectability | Can a reviewer compare the output with reliable evidence? | Errors are visible and inexpensive to correct. |
| Time Sensitivity | Does faster completion create real value? Is there time to review? | Speed helps without bypassing control. |
Classifying the Level of AI Involvement
- AI Executes, Human Reviews: Performs a repeatable bounded task and prepares the result. Human checks and approves before significant use (e.g., compiling weekly status summaries from authorised records).
- AI Supports, Human Leads: Organises evidence, drafts options, or highlights patterns. Human interprets, decides, and shapes the final output (e.g., preparing options for implementation recommendations).
- Fully Human-Led: May provide limited administrative assistance only. Human owns reasoning, decision, approval, and communication (e.g., approving client commitments or resolving high-impact trade-offs).
5. Designing Coherent Agent Functions
| Agent Name | Bounded Functions | Human Boundary |
|---|---|---|
| Engagement Briefing Agent | Compile updates, decisions, actions, and missing inputs from approved records. | Engagement lead validates context and decides emphasis. |
| Evidence Synthesis Agent | Organise sources, create cited summaries, and identify information gaps. | Consultant assesses relevance and forms conclusions. |
| Proposal Drafting Agent | Create first drafts from approved briefs, case material, and service catalogues. | Account owner approves claims, scope, price, and commitments. |
| Deliverable Quality Agent | Check structure, required sections, terminology, and acceptance criteria. | Quality reviewer decides readiness and signs off. |
| Client Communication Agent | Prepare routine updates and tailored draft messages from approved facts. | Named staff member approves external communications. |
| Operations Coordination Agent | Prepare task lists, reminders, packs, and routine internal summaries. | Operations owner controls priorities and exceptions. |
📋 COPY-AND-ADAPT PROMPT: Specialist Agent Scope Design
Design a bounded AI agent for [FUNCTION]. Define its purpose, users, allowed tasks, excluded tasks, approved knowledge, permitted actions, required inputs, output format, human approvals, escalation rules, limitations, success measures and test scenarios. Do not imply that the agent holds decision authority.
6. Designing AI-Enabled Workflows & Controls
- Trigger: A person or approved event starts the work.
- Input Validation: Required fields, source identity, and permissions are checked.
- Context Assembly: Authorised knowledge and current task data are retrieved.
- AI Step: The agent classifies, drafts, summarises, analyses, or proposes.
- Deterministic Checks: Rules validate required fields, ranges, states, or permissions.
- Human Checkpoint: An authorised person reviews evidence and resolves exceptions.
- Approved Action: A workflow creates, updates, sends, or stores the approved result.
- Logging & Feedback: Inputs, outputs, approvals, and actions are logged for continuous improvement.
| Workflow Stage | Mandatory Control |
|---|---|
| Before Data Access | Authenticate user and apply least-privilege permissions. |
| Before AI Processing | Validate source, purpose, and permitted data classification. |
| Before Recommendation | Require source references and explicitly identify missing inputs. |
| Before External Communication | Require named-user sign-off and approval. |
| Before Financial / Contractual Action | Keep ultimate authorisation exclusively with an authorised human. |
| After Action Execution | Create an auditable log entry and enable exception review. |
7. Knowledge, Data, and Permissions
- Authoritative Ownership: Name an owner, approved version, and review date for every knowledge source.
- Curation over Volume: Use small, highly curated collections rather than uncontrolled repositories, removing obsolete material.
- Least Privilege: Grant only the minimal data and tool permissions required for bounded functions, preferring read-only access.
- Security Boundary: Treat external content and tool outputs as untrusted inputs, keeping prompt instructions strictly separate from retrieved data.
Context Layer Setup: Learn how to manage curated project references and coaching chats in our Copilot Notebooks Guidance.
8. Human Review Checkpoints and Non-Delegable Rights
| Review Type | Core Reviewer Question |
|---|---|
| Evidence Review | Are the sources complete, current, and correctly represented? |
| Assumption Review | Has the agent filled a gap or confused fact with inference? |
| Technical Review | Did rules, calculations, formatting, and integrations work correctly? |
| Professional Review | Is the interpretation appropriate to the client and context? |
| Approval Review | Is the output ready to send, publish, commit, or action? |
Decision rights that MUST remain fully human-led:
- Final client recommendations and strategic advice.
- Contractual commitments, pricing, discounts, and scope changes.
- Acceptance of legal, regulatory, or reputational risks.
- Personnel decisions and formal quality sign-offs.
9. Testing, Validation, and Deployment
Test topics, tools, workflows, and prompts in isolation before connecting them. Validate using synthetic or representative non-sensitive data across normal requests, incomplete inputs, ambiguous prompts, edge cases, and prompt injection attempts.
Progressive Autonomy Principle: Responsibility should expand in small increments. First allow the agent to draft, then to prepare a proposed action, and only later permit approved low-risk actions where controls and monitoring are proven.
10. Connecting Agents as the Portfolio Grows
| Pattern | Use When | Design Focus |
|---|---|---|
| Hub | A request spans domains or requires a blended response. | Routing accuracy, shared context, conflict handling, and synthesis. |
| Handoff | One specialist agent should own the request end-to-end. | Clear triggers, context transfer, ownership, and return paths. |
| Workflow Orchestration | The sequence is process-driven and predictable. | Deterministic steps, approvals, retries, and audit records. |
11. Governance, Lifecycle Management & Project Setup
To prevent agent sprawl and security risks, maintain a central agent register capturing business owners, technical owners, status, knowledge sources, tools, permissions, and review dates.
Each substantial agent project should be executed using our standardised Microsoft 365 setup framework:
- SharePoint Folder: Controlled project files, approved prompts, designs, and evidence.
- Excel Project Log: Formal records of tasks, decisions, risks, testing, and handover evidence.
- Copilot Notebook: Curated project context, research, requirements, and coaching chats.
- Microsoft Planner & Agent: Approved operational tasks, progress tracking, and prioritisation.
Project Execution: To structure the workspace, task logs, and governance files for a new agent project, follow our 4-part Microsoft 365 setup framework.
12. Copy-Ready Workflow Prompts
📋 COPY-AND-ADAPT PROMPT: Human-Directed Workflow Design
Design a human-directed AI-enabled workflow for [TASK]. Show the trigger, required inputs, data validation, AI step, deterministic checks, human review, approval, action, logging, exception handling and feedback loop. Identify what AI may do, what requires confirmation, what requires human approval and what must remain human-led. Use the minimum necessary permissions and do not invent missing business rules.
📋 COPY-AND-ADAPT PROMPT: Agent Register Portfolio Audit
Review this AI agent register. Identify overlapping scopes, agents without clear owners, duplicated knowledge, excessive permissions, missing review dates, weak human checkpoints, untested dependencies and candidates for consolidation or retirement. Separate observations from recommendations and do not approve changes.
13. Quick-Reference Deployment Checklist
- ☐ Business task is specific and the accountable human owner is named.
- ☐ Task suitability assessment is complete and AI role is classified.
- ☐ Scope exclusions and “do not act” boundaries are documented.
- ☐ Knowledge sources are curated, owned, and versioned.
- ☐ Permissions follow least privilege (read-only where possible).
- ☐ Human review points and decision boundaries are explicit.
- ☐ Realistic test set passes (normal, edge cases, and injection tests).
- ☐ Logging, audit trails, and rollback routes exist.
- ☐ Agent is registered in the central agent register with a review date.
📌 Recommended Operating Summary
The most effective agent strategy is a portfolio of bounded capabilities designed around real work. Efficiency comes from assigning repetitive preparation to AI; trust comes from keeping human staff visibly in control of decisions and client outputs.