Executive attention has shifted from Agile transformations to AI. Yet approval queues, unclear priorities and cross-functional handoffs remain. AI will not repair those constraints by itself. It can remove appropriate routine work and accelerate learning, but only when the organisation also redesigns workflow, decision rights and human oversight.
It is tempting to declare that the Agile era is over and the AI era has begun. In day-to-day operations, changing the label does not change the system.
A team may produce a draft within an hour using AI. If it then waits ten days for three approvals, the business is not faster. It has filled the queue sooner.
DORA's 2025 research on AI-assisted software development describes AI as an amplifier of an organisation's existing strengths and weaknesses. That is a more useful starting point for leaders than a promise of automatic productivity. AI does not replace the organisational ability to decide what matters, test assumptions with customers and change direction when the evidence changes. Agility is supposed to build that ability.
In brief
- “AI versus Agile” is a false choice. AI works on tasks; agility should improve how the organisation prioritises, learns and delivers value.
- Agile transformation theatre appears when ceremonies change but priorities, decision rights and the flow of work do not.
- AI should be applied only after the workflow has been mapped and unnecessary steps have been removed.
- People must retain context, judgement, accountability, relationships and decisions with material consequences.
- Scaling belongs after a bounded experiment, human validation and Proof of Change against a baseline.
Two transformations, the same friction
Agile transformations promised shorter feedback loops, greater autonomy and faster value delivery. In some organisations, success was measured mainly by rolled-out roles and regular ceremonies. The decision system remained unchanged.
AI can follow the same path. An organisation buys licences, announces an initiative and begins counting use cases without identifying the constraining workflow, its baseline or who can change the rules.
The difference between Agile transformation theatre and genuine agility is the ability to discover quickly that an assumption is wrong and change direction before investing too much.
The May 2025 Kanban Guide defines Kanban as a strategy for optimising value flow through workflow definition, active management and improvement. If a team cannot see work in progress, ageing work, blockers and queues, it cannot know whether AI is accelerating the right thing.
AI cannot repair decisions the organisation refuses to make
AI can summarise documents, compare options, prepare a first draft, detect inconsistencies or convert unstructured input into a workable format. It cannot independently determine the legitimate priority between growth, risk, customer experience and regulatory obligations.
Nor can AI repair five systemic problems:
- Priorities without real choices. When everything is a top priority, AI merely increases the volume of work in progress.
- Governance without risk differentiation. Applying the same approval chain to a low-risk draft and a consequential decision creates avoidable queues.
- Handoffs without an outcome owner. Faster output from one function can overload the next.
- Incentives focused on local performance. Each function optimises its own step while the end-to-end result remains unchanged.
- Leadership without delegated authority. The team carries accountability but must wait for management whenever reality deviates from the plan.
If these conditions persist, the organisation generates more content and more automatically created tasks. That is not productivity. It is AI noise.
What AI can support and what must remain human
Ask “Which task can we safely eliminate, automate or augment?” rather than “Which role can we replace?”
| Suitable candidate for AI support | Human responsibility |
|---|---|
| Summarising extensive material with traceable sources | Decide which sources are trustworthy and relevant |
| Producing a first draft of a document, test or solution option | Judge quality, consequences and acceptability |
| Detecting repetition, inconsistencies and missing fields | Decide on exceptions and remain accountable for the impact |
| Preparing inputs for a retrospective or planning discussion | Lead sensitive dialogue, conflict and shared learning |
| Moving routine information between approved systems | Set data boundaries, access rights and controls |
The ILO's May 2025 study drew on a representative sample from 29,753 tasks in an occupational classification. Its conclusion is deliberately measured: because most occupations contain tasks that require human input, transformation of jobs is the more likely impact of generative AI than wholesale replacement.
That finding creates a leadership obligation. Changing tasks changes roles and professional identity. If people do not know whether AI is removing unnecessary work or evaluating them indirectly, they may resist or conceal experiments. Human Leadership requires clarity about purpose, authority, accountability and safe error reporting.
Six gates before scaling AI
Use six decision gates before launching another enterprise-wide transformation. Each may lead to continuing, changing course or stopping.
1. Map the flow of work
Choose one specific outcome and follow the work from request to value. Record waiting, returns, manual re-entry, exceptions and decision points. The map must describe reality, not the prescribed process.
2. Remove unnecessary work
Do not automate a report that nobody uses. Do not teach AI to copy data between two systems if the handoff can be removed. Elimination comes before automation.
3. Select a bounded routine
A strong candidate is repetitive, time-consuming and governed by understandable quality criteria. The organisation must also know which data may be used, who validates the output and what happens when the system is wrong.
4. Run a small experiment
Define the baseline, the outcome owner and a limited test period. Do not measure only the speed of the isolated task. Track waiting across the entire flow, rework, quality, exceptions and the effect on downstream roles.
5. Preserve meaningful human validation
The NIST AI Risk Management Framework addresses trustworthiness across the design, development, use and evaluation of AI systems. Human oversight cannot therefore be a signature at the end. The validator needs the competence and authority to challenge, stop and escalate the output.
6. Require Proof of Change
Compare the result with the baseline. Did the end-to-end flow improve, or only local activity? Did waiting or rework fall? Did quality and safety remain within acceptable boundaries? Scale only after that review. If the evidence is weak, adjust or stop the pilot.
Hypothetical enterprise example: supplier onboarding
This scenario is illustrative, not a Lemon SYNERGY client reference.
A multinational company introduces AI to prepare supplier-onboarding documentation. The model can extract information from contracts and populate a working draft quickly. The overall process does not become faster because each request still moves through procurement, security, legal and finance. Every function uses a different form, and ambiguous cases return to the beginning.
The team first maps the flow. It finds that some data is collected twice and that low-risk cases wait in the same queue as exceptions. The duplicate step is removed, a shared intake is introduced and the team defines which cases require expert judgement.
AI then receives a bounded role: extracting approved fields, checking completeness and preparing the working file. A person validates sensitive data, decides exceptions and remains accountable for approval. The pilot compares total lead time, returns and quality, not merely extraction speed.
Agility in this example is not a sprint calendar. It is the ability to test a change at limited scale, gather feedback and improve the system. AI does not replace that capability. It is one lever within it.
Risk is not an appendix to the pilot
A poorly led AI initiative can damage trust before it creates value. Surveillance appears when usage data is quietly repurposed to evaluate individuals. Deskilling becomes a risk when a role merely accepts AI output for long enough to lose the ability to assess it independently. Ambiguous accountability leaves everyone blaming the tool or the previous approval when something fails.
Data security requires approved tools, permitted data classes, retention rules, access rights and an escalation path. Current European Commission guidance on AI literacy emphasises context, people's knowledge and human oversight. A generic course disconnected from real roles is not enough.
Eight executive questions
- Which specific business outcome should change, and who owns it?
- Where does work currently wait, return or require re-entry across the end-to-end flow?
- What can be eliminated before anything is automated?
- Which bounded task is suitable for AI, and how will we recognise an acceptable output?
- Which decisions must remain human, and who can stop an AI-supported action?
- Which data, security, legal and intellectual-property boundaries must the pilot respect?
- How will the change affect roles, skills, trust and accountability?
- Which baseline and Proof of Change will determine whether we scale, adjust or stop?
Do not begin by asking what AI will replace
Begin where performance is being lost. The answer may be to remove an approval, clarify a priority or change a decision right. You may find a routine where AI returns capacity, or discover that the technology does not yet help. Learning that early is valuable.
AI will not replace agility as an organisational capacity to learn. It can, however, expose whether the organisation genuinely has that capacity.
If you need to select one workflow, identify the risks and frame a measurable pilot question, an AI Opportunity Scan may be an appropriate first step. It is a diagnostic approach for deciding whether to pilot, not a promise of implementation, compliance or ROI.