Many companies have now trained employees to use AI. They have enabled copilots, written internal policies and tested their first agents in sales, service or operations. Yet the economic impact often remains below expectations.
In my experience, the cause is rarely a lack of technology. The real problem is that companies place new AI capabilities inside old work models. Job descriptions, decision paths, performance metrics and handovers remain unchanged. AI accelerates individual tasks, but not the complete value stream.
This is the next major management challenge: Companies must not only introduce AI. They must design work so that people and AI can produce better outcomes together.
Deloitte describes this gap clearly. Nearly 60 percent of workers intentionally use AI at work, but only 14 percent of surveyed leaders say their organisation is adept at designing human-AI interactions. Fifty-nine percent still take a technology-focused approach and layer AI onto existing systems and processes.1 Microsoft’s 2026 Work Trend Index similarly finds that organisational factors such as culture, manager support and talent practices are associated with more than twice the reported AI impact of individual factors.2
The implication is clear: AI readiness is becoming a work-design question.
Why training alone does not solve the problem
AI fluency matters. Employees need to understand what a model can do, how outputs should be reviewed and where risks arise. But training does not redesign a process.
A sales representative can be excellent at prompting and still spend every day assembling information from three systems. A service employee can use an AI assistant and still require five approvals for every exception. A controller can create an analysis in minutes while the organisation still needs four weeks to make a decision.
Deloitte reports that many organisations prioritise AI fluency and upskilling, while far fewer redesign roles, workflows and career paths. Only 34 percent of surveyed organisations are deeply reimagining products, processes or business models; 37 percent use AI more superficially, with little or no change to existing processes.3
| Technology-led adoption | Human-AI Work Design |
|---|---|
| “Which tool should we give the team?” | “Which outcome must the process deliver reliably?” |
| AI assists isolated tasks | People and AI shape an end-to-end value stream |
| Existing roles stay unchanged | Roles, decision rights and exceptions are redesigned |
| Success is measured through usage | Success is measured through quality, speed, cost and risk |
| Control happens afterwards | Controls and escalation are built into the workflow |
Optimising only the left-hand side may improve individual productivity. Designing the right-hand side changes the performance of the enterprise.
Work must be redesigned from outcomes backwards
Traditional job descriptions bundle activities: answering emails, creating reports, qualifying leads, reviewing contracts or handling customer enquiries. This view is too broad for an AI-capable operating model.
The better unit is the outcome. In sales, that might be a qualified opportunity with complete CRM context. In service, a correct resolution within a defined time. In finance, an auditable forecast with documented assumptions.
Only after the outcome is clear can the work be decomposed meaningfully:
| Layer | Example: customer enquiry |
|---|---|
| Input | Email, form, meeting note or support ticket |
| Interpretation | Identify intent, urgency, customer and context |
| Knowledge work | Combine policies, product knowledge and history |
| Decision | Choose response, offer, escalation or rejection |
| Execution | Send communication and update systems |
| Control | Monitor quality, risk and outcome |
| Learning | Feed back errors, exceptions and new insights |
This decomposition avoids two common mistakes. First, it prevents companies from automating everything that is technically possible. Second, it turns human involvement from a vague promise into an explicit design choice wherever judgment, accountability, empathy or negotiation is required.
The Fassbender Human-AI Work Design Canvas
For practical implementation, I recommend seven management decisions. Together, they connect process design, operating model and governance in a system an SME can actually run.
| Design decision | Management question | Concrete outcome |
|---|---|---|
| 1. Outcome | Which business and work result should improve? | Target, baseline and quality definition |
| 2. Work Decomposition | Which activities, decisions and exceptions form the process? | End-to-end work map |
| 3. Interaction Mode | What relationship should exist between human and AI? | Defined interaction mode for each step |
| 4. Decision Rights | What may AI propose, decide or execute? | Decision and delegation matrix |
| 5. Controls & Escalation | Which thresholds limit autonomy? | Review rules, approvals and escalation paths |
| 6. Roles & Capabilities | Which human roles and capabilities change? | Role model, RACI and enablement plan |
| 7. Learning System | How are impact, errors and improvements fed back? | KPI, review and improvement cycle |
The canvas deliberately does not begin with a tool. It begins with the outcome and ends with a learning system. Technology is selected only when its role in the operating model is clear.
Five interaction modes instead of a generic “human in the loop”
“A human reviews the output” sounds responsible, but it is not yet a robust design. Companies need to define more precisely how people and AI interact.
AI as assistant means AI prepares information while a person performs the work and makes the decision. AI as collaborator describes an iterative process in which both parties develop drafts, analyses or options. AI as coach provides guidance and feedback without taking the decision. AI as executing agent performs defined steps within explicit boundaries. AI as supervised operator manages a process largely autonomously while humans monitor thresholds, exceptions and system performance.
Microsoft distinguishes similar modes along human and agent intensity. The important question is not which mode appears most advanced. It is which mode fits the work, its risk and the required level of judgment.2
A high-risk process may use AI for research and consistency checks while retaining final approval with an accountable person. A standardised internal data transfer may allow AI to execute within boundaries and escalate only deviations.
Decision rights are the core of the new operating model
As soon as AI moves beyond drafting text and begins operating systems, setting priorities or triggering actions, it moves from tool toward organisational actor. Autonomy, persistence and delegated rights change the management task.4
A practical delegation matrix can distinguish four levels:
| Level | AI authority | Human role | Suitable for |
|---|---|---|---|
| Propose | AI creates options | Human decides completely | new, sensitive or insufficiently tested work |
| Execute after approval | AI prepares the action | Human approves before execution | financial, legal or customer-facing steps |
| Execute with escalation | AI acts inside defined boundaries | Human handles exceptions | stable processes with measurable thresholds |
| Autonomous within mandate | AI decides and acts in a narrow scope | Human monitors system performance | highly standardised, reversible and observable workflows |
The concept of human on the loop is useful here. People do not approve every transaction. They define goals, boundaries and escalations, monitor patterns and intervene when the system deviates. This is not weaker control. It is a different form of control that can scale with transaction volume.
Accountability nevertheless remains human. Deloitte reports that only about one in five companies has a mature governance model for autonomous AI agents.3 This is why decision rights and control architecture cannot be postponed until after a pilot.
The roles an SME actually needs
A mid-sized company does not immediately need a large AI centre of excellence. It does need explicit accountability. One person may cover several roles, but none of the functions should remain invisible.
| Role | Core accountability |
|---|---|
| Executive Sponsor | Objective, priority, budget and conflict decisions |
| Process Owner | End-to-end outcome, process performance and change |
| Business AI Owner | Functional configuration and business value of the AI system |
| Quality Steward | Quality criteria, tests, samples and failure patterns |
| Data Owner | Access, quality, purpose limitation and data accountability |
| AI Operations Owner | Operations, monitoring, cost and technical escalation |
| Risk/Compliance | Risk class, control requirements and evidence |
These roles also change management. Managers no longer allocate tasks only. They define outcomes, set quality standards, create room for experimentation and decide when a new way of working becomes mandatory.
Microsoft’s research illustrates why this level matters. Only 26 percent of surveyed AI users say leadership is clearly and consistently aligned. Sixty-five percent fear falling behind if they do not adapt quickly, while 45 percent say it feels safer to focus on current objectives than to redesign work with AI. Only 13 percent say work redesign is rewarded even when an experiment does not immediately succeed.2
If performance metrics reward only the old system, no training programme will embed the new one.