Many organisations track active users, account numbers, weekly prompts and training attendance, but while these measures show that employees are using AI, they do not answer management’s central question: where is the business value?
AI may help customer-service staff draft accurate emails in one minute instead of five, generate polished reports and presentations in minutes, or produce professional sales messages, yet these outputs do not necessarily improve customer loyalty, management decisions or conversion rates. Indeed, cheaper and faster content creation can sometimes conceal shallow thinking or result in generic communication.
This creates a value gap in which an organisation is AI-active but not AI-productive: employees are busy using AI, while profit, revenue, customer satisfaction and decision quality remain unchanged. An “AI Enterprise 2.0” must therefore move beyond assistant-level productivity and focus on generating measurable business value.
The strategic shift
Agentic artificial intelligence marks a move from passive, prompt-driven assistance to goal-directed participation in business workflows. An AI assistant typically completes a single task after receiving an instruction. An AI agent can pursue a defined goal, plan several steps, use authorised information and systems, execute actions within limits, monitor results and escalate exceptions.
This distinction changes the management question. The issue is no longer simply whether employees use AI, but how work should be divided between people and digital agents, how decision rights should be assigned, and which outcomes demonstrate value. Agentic AI is therefore as much an organisation-design and governance challenge as a technology deployment.
Activity is not the same as value
Adoption measures—active accounts, prompts per week, generated documents and training completion—show activity, not necessarily performance. Faster email drafting may save time without improving customer loyalty. A polished report may be produced quickly while contributing little to decision quality. Generic sales messages may look professional without lifting conversion.
A value-led programme links agent performance to business results such as revenue, profitability, customer satisfaction, risk reduction, decision quality and response time. Time saved remains useful, but it should be treated as an input to value creation rather than the final result.
Design roles, not isolated tasks
Durable agent design begins with a role. “Summarise calls” or “check invoices” describes an activity; “customer insight agent” or “procurement compliance agent” defines a continuing responsibility that can be integrated into a team.
Each agent role should specify four elements:
- Accountability — the result for which the agent is responsible.
- Access — the systems, data, documents and tools it is permitted to use.
- Authority — the decisions or actions it may take, including thresholds and prohibited actions.
- Performance — the business outcome and quality measures used to judge success.
This approach resembles onboarding a new team member: define the role, provide only the necessary access, explain internal rules, test performance and review the work. A well-written prompt cannot compensate for unclear accountability or ambiguous business policy.
A modular operating model
Multi-agent systems are easier to understand and govern when responsibilities are separated. Specialised agents can be tested, improved, replaced and reused independently. An orchestrator may later coordinate the workflow, but it should generally be introduced only after the underlying roles, hand-offs and controls are proven.
| Management stage | Agent contribution | Human responsibility |
| Signal | Monitor events, detect exceptions and collect relevant inputs. | Define what matters and validate source coverage. |
| Decision | Interpret, classify, score or recommend within stated policy. | Set criteria, challenge assumptions and decide ambiguous or high-impact cases. |
| Action | Draft, notify, update records or execute authorised steps. | Approve material actions and remain accountable for outcomes. |
| Learning | Review results, errors and changes in the operating environment. | Adjust policy, thresholds, access and performance expectations. |
Governance before autonomy
The principal risk is not factual error alone. An agent can use accurate information and still make a poor business decision when goals, priorities or constraints are unclear. Governance should therefore translate management judgement into explicit operating boundaries.
- State the goal together with priorities, budget limits, service levels and prohibited actions.
- Segment customers, transactions or scenarios so that different rules can be applied consistently.
- Set confidence, financial, risk or materiality thresholds that trigger human review.
- Consider reversibility: lower-risk decisions are generally easier to delegate when they can be corrected without disproportionate harm.
- Maintain a named human owner, an escalation route, an error taxonomy and a review cadence.
Graduated decision authority
Authority need not be binary. It can increase as evidence and controls mature:
| Level | Authority | Control principle |
| 1 | Observe and alert | The agent identifies events or anomalies; a person decides what follows. |
| 2 | Prepare | The agent gathers and organises information for human use. |
| 3 | Recommend | The agent proposes a decision or action with supporting rationale. |
| 4 | Execute within boundaries | The agent acts under defined thresholds and escalates exceptions. |
| 5 | Orchestrate | The agent coordinates other agents while preserving controls and escalation. |
Thresholds should be calibrated through pilots. If they are too strict, routine cases repeatedly return to people and the expected value disappears. If they are too loose, the organisation accepts excessive risk. Human review should concentrate on exceptional, ambiguous and high-impact cases rather than duplicate every routine step.
Illustrative functional architectures
Sales
A signal agent identifies tenders or prospect activity; a briefing agent researches the organisation and industry; a scoring agent assesses urgency, value, deadline and fit; and an action agent prepares material after human approval. The human salesperson validates the opportunity, interprets context and develops the relationship.
Finance
A data-quality agent checks consistency; a variance agent compares periods; a scenario agent investigates causes and options; and a briefing agent assembles management materials. Analysts shift from data extraction towards challenge, judgement and advice.
Legal and professional services
An intake agent checks completeness; a precedent agent retrieves similar matters; a policy or research agent gathers relevant internal and external materials; a risk-classification agent flags higher-risk files; and a drafting agent assembles material for professional review. Human judgement and accountability remain essential, particularly where consequences are significant.
Operations
An inspection agent monitors operating signals; a diagnostic agent investigates anomalies; and an improvement agent evaluates corrective actions. An orchestrator may coordinate the agents once the workflow is understood. Automatic implementation should be limited to actions that fall within tested operating boundaries.
Technology and data considerations
Agents may draw from internal documents, messages, meetings, operational data and authorised external sources. Access should follow least-privilege principles, and the provenance and completeness of sources should be understood. When an external source blocks automated access or requires verification, organisations should use an authorised alternative, licensed access or a human review step rather than assuming that a partial result is complete.
Platform selection, licensing, integration and security remain important, but they should support—not substitute for—clear workflow design. The strongest controls begin with defined roles, decision rights, business rules and accountability.
A practical 30-day management exercise
- Select one important workflow with a visible business outcome and manageable risk.
- Map its signals, decisions, actions, hand-offs, approvals and learning points.
- Define one or two narrow agent roles, including authorised data and prohibited actions.
- Begin at observe, prepare or recommend authority unless evidence supports more.
- Name the human owner and establish review cadence, error categories and escalation routes.
- Compare performance with the current process across quality, speed, cost, risk and business outcome.
- Expand authority or add an orchestrator only after the evidence and controls justify the change.
| MANAGEMENT CONCLUSION
The aim is not the same output at greater speed. It is a better-performing human–AI system: faster management response, stronger decisions, improved service and scalable execution within responsible boundaries. |