I have a real problem with the way organisations are justifying AI investment right now.

The dominant logic goes something like this: AI can automate cognitive work; therefore, we can reduce headcount; therefore, the investment pays for itself. It is clean, it is measurable, and it is, in my view, a serious mistake — for reasons that are practical, economic, and ultimately ethical.

The Wrong Measurement

Think back to the ERP rollouts of the last two decades. Organisations spent millions on modernising their financial infrastructure. Did early ERP vendors promise back-office consolidation and headcount savings? Absolutely. But the measure of success was mainly functional: faster close, cleaner data, better forecasts.

Did successful ERP deployments eventually affect headcount? Yes, but mainly through not needing to hire additional people as the business scaled, not through eliminating existing roles. The experienced leaders understood that drawing conclusions about staffing before the infrastructure was stabilised was a recipe for operational failure. The capability had to be proven first.

With AI, the logic runs in complete reversal. Cut first, prove later. Some companies have already discovered what this leads to — rehiring staff they eliminated because the AI could not handle the operational reality without them. Because AI agents can take autonomous action in a way an ERP never could, the temptation to treat headcount reduction as the primary metric is understandable. But it is the wrong measure at the wrong moment.

The better question is: should AI be measured by headcount reduction at all? Or should it be measured by the contribution it makes beyond what humans could do alone — the speed, the scale, the pattern recognition across data volumes no team could process manually? That is a completely different conversation. And it is the one most organisations should be openly discussing.

How Could It Apply to FP&A Specifically?

FP&A is the only function that genuinely sits in the middle of everything. It simultaneously touches commercial performance, operational costs, capital allocation, and strategic planning. What FP&A leaders decide and what they measure shapes decisions across the entire organisation, not just within a single function. That is not a small thing. It is also precisely why FP&A is where AI investment should be evaluated differently.

The baseline question for any AI deployment in FP&A should not be how many analysts were saved. It should be: can we now do something we were never able to do before?

That might mean the ability to analyse internal performance against external market data in real time — finding correlations between historical patterns and market fluctuations that were previously too complex to model manually. It might mean building a clearer picture of what is happening to your most important customers — what market forces are affecting their businesses, and how that will translate into your own revenue trajectory before it shows up in the numbers. It might mean scenario planning that is genuinely forward-looking rather than a repackaging of last year's assumptions.

The measurement framework shifts accordingly. Success is not headcount reduced. It is predictive accuracy — how capable the function becomes of anticipating what is happening around the organisation, not just within it. How much earlier can strategic decisions be made? How much closer FP&A gets to being a genuine input into company strategy, rather than a reporting function that describes the past as a proxy for a future that may look quite different.

In practical terms, this means FP&A should measure AI not only by headcount reduction, but by whether predictive accuracy improves, strategic decisions can be made earlier, scenario planning becomes more forward-looking, AI usage costs are understood, and the level of human review required is clear.

The Economics Are Not as Clean as They Look

It is simply impossible to predict with confidence whether AI is cheaper than a human. The cost of running AI at scale is more volatile and less predictable than most headcount reduction models assume.

AI is not billed like traditional software. API connections are charged per token - every query, every back-and-forth between systems, every agent action costs money. The more integrations you build, the more complex the workflow, the higher the bill. Gartner warns that while unit token costs are falling, overall inference costs may still rise because consumption is rising faster than prices are dropping. Agentic models require between 5 and 30 times more tokens per task than a standard AI chatbot. This means running AI at scale can quickly rival or surpass core engineering payroll expense, yet most organisations still lack the frameworks to forecast and control these volatile infrastructure costs accurately.

Optimum Partners, citing the FinOps Foundation’s 2026 State of FinOps report, notes that 73% of organisations reported AI costs exceeding original projections. This uncertainty is not an argument against AI. It is an argument for deploying it where it adds the most demonstrable value first, rather than using it as a blunt instrument for cost reduction before its economics are properly understood.

AI Is Not Yet Ready for Messy Reality

There is a paradox at the heart of current AI deployment. AI holds enormous information. What it cannot reliably do is operationalise that information in a specific, messy, and context-dependent business environment without human intervention. Someone has to correct it when it hallucinates. Someone has to handle the edge cases it drops. Someone has to understand the business logic well enough to know when the output is wrong.

That work is being done right now, inside finance teams across every industry, largely without acknowledgement. People doing it are not just using AI — they are making it usable. They are feeding it the institutional knowledge, the exception handling, the contextual judgment that the model cannot supply for itself.

The Redeployment Headlines Are Louder Than the Plans

Announcements about headcount reductions tend to be loud and specific. The details behind redeployment tend to be thin.

Most organisations are better at communicating what they are cutting than explaining what the transition looks like: what roles will exist, what skills will be needed, what investment in development will be made, and over what timeline. In the absence of that detail, redeployment is not a plan.

And it creates a future problem that most organisations are not yet pricing in.

The Regulatory Reality

The EU AI Act exists. Enforcement deadlines for many high-risk systems are being phased in from December 2027, with further timelines extending into 2028, but the direction of travel is clear. Requirements around bias elimination, model reliability, auditability, and third-party assessments are coming.

Governing autonomous AI systems to stay compliant will require a combination of technical understanding and institutional context. That expertise does not yet exist at scale. When it does, external specialists will command a significant premium.

Again, people best placed to develop that capability internally are the domain experts who understand the underlying process, the exceptions, and the business logic. Organisations eliminating those people today will find themselves competing in a talent war for someone who understands AI, compliance, and their specific business context simultaneously. Those people will cost multiples of what it would have taken to develop the expertise already inside the business.

And no, your internal team does not need to become deep tech engineers. You can always rent external developer talent or use third-party providers for specific technical knowledge. But those developers will be useless without internal handlers.

How FP&A Can Address the Mess, Redeployment, and Regulation

Here is an example of how the three issues described above could be addressed, specifically using the FP&A function.

Even if AI is successfully rolled out, as a layer on top of your existing FP&A tool, for example, there is an urgent underlying gap that is not being discussed enough. It is not just about clean data. It is also about someone owning the logic layer beneath AI. Someone needs to define what the system is authorised to conclude, from what data, using what assumptions, and, more importantly, when it all goes stale because your business has changed. Without that governance decision, sometimes referred to as ‘mandate’, clean data and good architecture are not enough. The system will continue producing confident, coherent outputs from a frame that no longer reflects reality.

FP&A is better placed than most to own this. The function already sits at the intersection of commercial, operational, and financial data. And it already operates as the connective tissue between business functions, making it the natural bridge between those functions and IT when mandates need defining, maintaining, or revising.

This does not make FP&A a technology function. It makes it more strategically relevant.

The mandate question is also a potential answer to the redeployment gap, the governance challenge, and the question of whether agentic AI was the right tool in the first place.

The Obligation That Also Needs to Be Discussed

There is a deeper issue that most articles on this subject do not name directly.

The regulatory response to AI displacement has not kept pace with the speed of deployment. The proposed mechanisms for broader societal benefit — sovereign wealth participation and the idea that AI will make everyone richer — remain largely theoretical and unproven at the pace and scale required.

This creates an acute organisational-level obligation. The same employees who are actively making AI viable are the ones most exposed to displacement from the system they are improving. They helped build the capability that AI is now being layered on top of.

Organisations have responsibilities to multiple stakeholders. Shareholders matter. They are not the only ones who matter. For any progressive business, especially those certified as B-Corps, the reasonable expectation is that our people's work evolves alongside technology. A redundancy package is not an evolution plan.

Deploying AI responsibly means being honest about the economics, realistic about what redeployment requires, serious about governance, and clear-eyed about the fact that the people making AI viable in your specific context deserve more than a ‘thank you, goodbye’.

Sources

1. Gartner: AI coding costs to surpass developer salary by 2028 - Gartner Press Release

2. Gartner: LLM inference costs to fall 90% by 2030 - Gartner Press Release

3. Optimum Partners. “AI Token Costs and How They Might Wreck Your Budget.” 19 May 2026. Citing the FinOps Foundation’s 2026 State of FinOps report. Optimum Partners

4. Policy-Insider.AI. “EU AI Act Timeline: Key Compliance Deadlines for 2027–2028.” 27 May 2026- https://policy-insider.ai/eu-ai-act-timeline-key-compliance-deadlines-for-2027-2028/