Why traditional cost-cutting logic, premature workforce decisions and the loss of institutional knowledge could undermine the very AI transformations companies are trying to achieve.
A recent conversation with a friend, a senior finance executive at a large multinational, left me thinking about a contradiction at the heart of many AI transformation initiatives.
Her organization is preparing to move significant parts of its finance operations toward an agentic operating model. The ambition is substantial: better quality, greater productivity and significantly lower operating costs. So far, nothing unusual. These are precisely the outcomes that make AI transformation so attractive to business leaders.
What troubled me was how the transformation was being translated into management decisions.
My friend had already been given an ambitious headcount reduction target, yet she had very limited visibility into the future operating model that was supposed to make those reductions possible. She didn’t know which activities would be performed by AI agents, which decisions would remain human, or how exception handling, internal controls and accountability would work. Most importantly, she had no clear picture of what capabilities her future team would actually need.
At the same time, she was expected to replace some of her experienced employees with new talent bringing fresh capabilities into the organization.
Her dilemma was painfully practical. Where do you find people who understand both the company’s complex internal finance operations and AI-driven automation? And how do you identify the skills you need when you haven’t yet defined the operating model those people are supposed to support?
I kept thinking about one question: What if some of the people being let go are precisely the people the organization needs to make the transformation work?
I don’t know the full context behind this particular company’s decisions. There may well be detailed transformation plans elsewhere in the organization. But the conversation revealed a problem in its own right: a senior leader expected to deliver the transformation had little visibility into the assumptions behind it.
More importantly, it illustrates a broader management risk that I believe deserves much more attention.
We’re trying to build AI-native organizations using management practices designed for the pre-AI world.
1. Workforce reduction: an input or an output?
I spent a significant part of my career in corporate finance. I understand the pressure to deliver productivity improvements, reduce costs and commit to financial targets. There is nothing inherently wrong with setting ambitious efficiency objectives for AI transformation. In fact, without clear business objectives, transformation programs can easily become expensive technology experiments.
The problem begins when we treat the expected financial outcome as though it were already an established operational fact.
In a traditional cost reduction exercise, management might set a savings target and ask business units to develop plans to achieve it. This approach assumes that the underlying operating model is relatively stable and that the organization can identify opportunities to reduce costs within that model.
But agentic transformation is fundamentally different. It potentially changes the architecture of work itself.
Consider a finance process involving data collection, validation, reconciliation, exception handling, approvals and reporting. Different parts of that process may be suitable for different forms of automation. Some activities can be handled through deterministic workflows, others may benefit from AI-assisted analysis, while certain tasks might be suitable for autonomous agents operating within carefully defined boundaries. Still others will continue to require human judgment, accountability or intervention.
The relevant question is therefore not simply how many activities can be automated. It is how the entire process will function once work is redistributed between people, software and AI agents.
This matters because automation does not necessarily eliminate complexity. It often redistributes it.
A process may become faster at one stage while creating a bottleneck elsewhere. An automated decision may eliminate manual effort while introducing new monitoring, validation or exception-management requirements. An agent may reduce processing time while increasing dependencies on data quality, system availability and access management.
These changes don’t necessarily weaken the business case for automation. They determine what the real business case looks like.
And they have direct implications for workforce planning. A process that becomes 70% automated does not automatically require 70% fewer employees. The remaining work may involve different tasks, capabilities, control responsibilities and capacity requirements.
Workforce requirements should be derived from the redesigned operating model, not inferred directly from an automation percentage.
There is also a sequencing problem. Reducing capacity before the new operating model has been validated may undermine the very transformation intended to generate the savings. During implementation, organizations often need additional capacity to document processes, validate outputs, handle unexpected failures and operate old and new systems in parallel.
In other words, the timing of workforce reductions matters just as much as their eventual magnitude.
2. Are we discarding the knowledge we need most?
The second issue is particularly troubling because it exposes a common assumption about AI transformation: that building an AI-enabled organization necessarily means replacing existing employees with people who possess new technological skills.
Sometimes external recruitment is essential. New capabilities will undoubtedly be needed. But treating this primarily as a workforce replacement exercise overlooks the value of institutional knowledge.
Think about the finance professional who knows why a particular reconciliation regularly fails, the controller who understands which apparently routine transaction requires additional investigation, or the team member who knows that two systems use slightly different definitions of the same business concept.
Much of this knowledge is not formally documented. It exists in people’s experience, workarounds, informal controls and understanding of how the organization actually operates.
In traditional process improvement, this knowledge is valuable. In agentic transformation, it may be indispensable.
When organizations automate processes based primarily on documented workflows, they risk reproducing the formal process while overlooking the exceptions, dependencies and informal decisions that make it work in practice. These gaps may remain invisible until the automated system encounters a situation that experienced employees previously handled almost instinctively.
This is particularly relevant when moving from conventional automation to more autonomous AI systems. An agent may be able to execute a task, but the organization still needs to understand the boundaries within which that task can be performed reliably, which exceptions require intervention, and how to recognize when an apparently successful outcome is actually wrong.
The people who understand these distinctions are often the employees who have spent years working with the processes we now want to automate.
The people who know where the process breaks are often the very people you need to redesign it.
Of course, not every existing role will survive, and not every employee will have the capabilities or motivation to transition into a new one. But the decision should be based on a deliberate assessment of future work and individual capabilities, not on a general assumption that new technology requires entirely new people.
The real challenge is not choosing between existing talent and fresh talent. It is determining how to combine domain expertise, process knowledge, technological capabilities and the ability to manage AI-enabled operations.
And doing so before losing knowledge that may be extremely difficult to recover.
3. We’re hiring for jobs we haven’t designed yet
This leads to the third contradiction: organizations are preparing to recruit new talent for roles they may not yet fully understand.
We hear increasingly frequent references to AI specialists, automation experts, agent supervisors and AI-enabled finance professionals. But what do these roles actually require?
Does a future finance professional need to know how to build an AI agent, or how to define business rules and evaluate its performance? Do they need programming and integration skills, or the ability to investigate exceptions, challenge AI-generated recommendations and exercise professional judgment when automated processes encounter uncertainty?
The answer depends on the operating model.
A highly automated accounts payable process will have very different human capability requirements from an AI-augmented financial planning function. Even within the same process, the necessary skills will vary depending on the degree of autonomy, the nature of the decisions involved and the risks associated with errors.
This is why I believe strategic workforce planning needs to move beyond comparing today’s job descriptions with tomorrow’s technology capabilities.
We need to understand the work itself: which activities disappear, which are transformed, which become more important, and what entirely new human responsibilities emerge when AI assumes a larger share of execution.
We also need to distinguish between capabilities required at the individual level and those that can be distributed across a team.
An organization does not necessarily need every finance professional to become an AI engineer. It may instead need a combination of domain experts, process owners, AI specialists, control professionals and managers who understand how to orchestrate work across human and AI resources.
The right combination will depend on the specific process, technology and risk environment.
Only after understanding these requirements can we make informed decisions about reskilling, redeployment, recruitment and potential workforce reductions.
The question isn’t whether we need new skills. We certainly do. The question is whether we’re designing the work before deciding who should do it.
4. Governance belongs inside the operating model, not on top of it
There is another dimension to this problem that often gets lost in discussions about productivity and workforce transformation: governance.
The term tends to bring compliance, policies, documentation and regulatory requirements to mind. These are important, but they represent only part of what governance needs to accomplish in an AI-enabled organization.
In an agentic operating model, governance becomes inseparable from operational design.
Consider a relatively straightforward finance automation scenario. An AI agent is authorized to process transactions, reconcile information across systems and initiate corrective actions when discrepancies are detected.
Who determines the agent’s authority? Which transactions may it execute independently, and which require approval? What information can it access? Under what circumstances must it escalate to a human? Who is accountable for the outcome, and how does the organization detect errors that may not be immediately visible?
These are not questions that can be satisfactorily addressed by adding a governance framework after implementation.
They influence which activities should be automated in the first place, how much autonomy is appropriate, what technical capabilities are required and how the organization must operate.
They also have economic consequences.
A process requiring intensive human verification may have a very different cost structure from one operating reliably within predefined automated controls. Equally, the choice between preventive controls, continuous automated monitoring and human intervention can significantly affect processing capacity, operational resilience and workforce requirements.
In other words, governance decisions directly influence the business case.
Governance is not merely a constraint on AI transformation. It is one of the inputs required to design an operating model that can deliver value reliably.
This also challenges the traditional separation between business transformation, technology implementation, risk management and compliance. In an agentic organization, decisions made in any one of these areas can fundamentally change the requirements in the others.
Designing them sequentially, in separate organizational silos, creates unnecessary risks and can lead to expensive redesign later.
5. The real problem may be how we manage transformation
Perhaps the most important lesson from my friend’s experience is not about technology or headcount. It is about management.
Large AI transformation programs typically involve several parties. Senior executives establish financial and strategic objectives. Business leaders are responsible for operational delivery. Technology teams and external vendors evaluate and implement solutions. HR develops workforce plans, while risk and compliance functions establish requirements and controls.
Each group may make perfectly reasonable decisions within its own area of responsibility.
But the decisions themselves are deeply interconnected.
An automation choice changes the process. The redesigned process changes the risk landscape. The control architecture determines which activities can operate autonomously and which require human involvement. These decisions influence workforce requirements, skills, organizational structure and management responsibilities. Together, they determine the investment, operating costs, implementation risks and realizable benefits.
If these decisions are made independently, an organization can end up with a transformation plan that appears coherent on paper but contains fundamental inconsistencies.
The technology roadmap may assume levels of autonomy that the control environment cannot support. The workforce plan may anticipate headcount reductions before the corresponding processes are operational. Recruitment may focus on technical skills without considering the business knowledge required to operate the redesigned processes. And financial targets may be incorporated into budgets before the underlying assumptions have been sufficiently validated.
The problem is not necessarily that any one party made a bad decision. It is that the organization failed to connect decisions that can no longer be made independently.
This is also why I believe AI transformation requires a different kind of management discipline.
Not less financial rigor, but more integrated financial, operational and organizational planning.
Not less accountability, but clearer accountability across the entire transformation.
And not less ambition, but a more rigorous understanding of the conditions under which that ambition can be realized.
6. What would a better approach look like?
I would start with the work, not the technology or the headcount target.
The first step is understanding how current processes actually operate, including their data flows, exceptions, dependencies, bottlenecks, controls and capacity requirements. This requires going beyond formal process documentation and involving the people who understand how work gets done in practice.
The next step is exploring alternative automation scenarios rather than assuming that the maximum possible degree of autonomy is necessarily optimal. Different solutions may produce different combinations of efficiency, quality, resilience, cost and risk. In some cases, conventional automation will be sufficient. In others, AI assistance or agentic execution may create substantially greater value.
Each scenario should then be evaluated as a potential future operating model, not merely as a technology implementation.
How would the end-to-end process change? Where would bottlenecks move? Which new risks would emerge, and which existing risks might be reduced? What controls and governance mechanisms would be required? How would responsibilities be distributed between humans and AI, and what capabilities would the organization need to operate the resulting system?
Only then can we build a workforce transition plan that distinguishes between activities being automated, roles being transformed, capabilities that can be developed internally and those that genuinely require external recruitment.
The financial business case should reflect these interdependencies. It should distinguish theoretical automation potential from achievable productivity improvements and, critically, from savings that can actually be realized within a defined timeframe.
And the implementation roadmap should recognize that the transition itself requires resources. Process expertise, testing capacity, change management and parallel operations are not peripheral activities. They are part of what makes the transformation possible.
Importantly, none of this means that financial targets must wait until every operational detail has been resolved. Ambitious goals can and should provide direction. But there is a critical difference between setting a target and treating the target as a validated transformation plan.
The purpose of integrated planning is to make the assumptions, dependencies, trade-offs and uncertainties visible before management commits to decisions that may be difficult to reverse.
7. AI transformation needs a different management mindset
We are entering a period in which AI may fundamentally change how organizations operate. Yet many of our management practices were developed for a world in which work, roles and organizational structures were relatively stable.
We set annual targets, allocate budgets, assign headcount numbers and hold leaders accountable for delivery. Those disciplines still matter. Having spent much of my career in finance leadership, I would be the last person to argue otherwise.
But they are no longer sufficient when the transformation itself changes the nature of the work, the division of responsibilities and the capabilities required to deliver business outcomes.
The challenge is not simply to make existing organizations more efficient. It is to understand how they should operate when human expertise and increasingly autonomous AI systems work together.
That requires a different sequence of decisions, a different level of integration between business functions and a much more deliberate approach to managing organizational knowledge.
Above all, it requires us to resist the temptation to treat projected efficiency gains as realized savings before we understand the operating model that will make them possible.
Otherwise, we risk optimizing the cost base of an organization that no longer exists while undermining our ability to build the one that should replace it.
We’re trying to build AI-native organizations. Perhaps it’s time we started managing their transformation differently.
And that leaves me with one question:
Are we actually redesigning organizations for AI, or are we simply using AI to justify the cost reductions we already wanted to make?
Redesign the work before you resize the workforce.