The following is a guest article by Diana Mugambi, Senior Manager of FP&A Operations at GE Vernova. The views expressed are solely those of the author.

Over the past year, almost every finance meeting and webinar deck I have reviewed has featured similar slides. These presentations typically include AI and automation topics for finance teams, emphasizing promises of productivity gains and the urgent steps needed to achieve them. CFOs are under immense pressure to embrace automation or risk being left behind.

For modern finance leaders, the challenge lies in determining which areas should not be included in the automation loop. These choices often do not remain at an abstract level but surface later during forecast updates, financial report reviews, internal control walkthroughs, or audits by the audit team—at which point no one can point to a single decision owner.

I support automation, but not the way it is currently being marketed. My concern is that human accountability in financial decisions is being quietly eroded—because outputs are system-generated, they are automatically accepted.

When Speed Quietly Replaces Context

Many AI automation pitches and productivity discussions treat human judgment in financial decisions as a flaw—biased, inconsistent, and slow. This naturally should be "engineered away." AI is objective, unaffected by fatigue, never takes vacations or sick days, and scales quickly. This narrative is appealing, but it is also incomplete.

Judgment does not disappear when processes are automated. Instead, it shifts from being exercised by humans to being embedded in computer programs. Design decisions may be made only once, but models use these judgments daily to produce output decisions. Once embedded for months or even years, these judgments become harder to challenge. In practice, financial decisions—such as reserves, impairments, and revenue recognition—are rarely binary, and the human element relies on contextual assessment of inputs, not just computational accuracy.

When automation treats these decisions as deterministic, it does not eliminate risk but rather centralizes it. We are locking judgment into AI-driven processes, often without the same level of scrutiny applied to human decisions. Human risk is visible, surfacing in management reviews and audit findings. Automated errors, however, often do not self-disclose, and due to scale, their potential harm can be greater.

If AI applies flawed assumptions and executes them consistently, the consequences are more severe, and the impact manifests later. Ultimately, the CFO's signature does not come with an AI or software license disclaimer. The fiduciary responsibility for the integrity of financial statements remains human. If an algorithm "hallucinates" a forecast or amplifies a flawed assumption, the consequences could be dire. You cannot delegate legal accountability to a vendor's black-box algorithm.

Are There Downsides?

One unintended consequence of aggressive automation that I failed to fully appreciate early on is that it rarely considers the long-term erosion of financial judgment in the development of finance talent. Financial judgment is not intuition; it is built through repeated exposure to ambiguity, errors, and the consequences of poor decisions.

Early in my career, these moments were inevitable, especially during close and budget cycles when forecasts did not reconcile cleanly. Estimates were debated, and exceptions triggered uncomfortable conversations during reviews. This friction, though inefficient, was formative. It taught teams to recognize when numbers were technically correct but conceptually unreasonable.

As more validation moves into automated loops with built-in logic, fewer assumptions are challenged. This friction disappears, and this "memory" is no longer exercised, passed on, or challenged. Forecasts converge too neatly, exceptions are filtered out before they become issues, and assumptions go unquestioned because the model has been "cleared."

Over time, the finance function becomes an efficient data-processing machine, but its historical depth diminishes, and its ability to detect when results diverge from reality weakens. The cost of this atrophy rarely shows in steady-state operations but surfaces during disruptions, crises, market shocks, and restructurings. In those moments, models trained on stable patterns encounter realities they were never designed to handle, and decisions must be explained to boards, auditors, regulators, or markets. No one asks an algorithm to explain itself. Accountability and judgment for financial results remain human.

Drawing Hard Boundaries

When developing a finance automation strategy, the discussion should not be about whether a process can be automated, but whether failures can still be detected and challenged. While we all agree that not all finance processes should be treated equally, some low-impact, repetitive processes do benefit greatly from AI's productivity gains.

However, in areas with significant impact that typically rely on human judgment, caution is necessary. This does not mean AI has no role, but AI should be used to identify risks and exceptions, simulate scenarios, and enhance judgment. This is not resisting innovation but safeguarding the credibility of the finance function and ensuring we have clear ownership of our decisions.

When Slowing Down Feels Like Risk-Taking

In practice, restraint is harder than acceleration. I have seen automation decisions driven more by inertia than conviction: automation roadmaps have been approved, consultants arrive with polished decks promising transformation, vendors only show "happily ever after" stories, and peers showcase achievements at industry conferences. In this context, pausing almost seems irresponsible, especially when CFOs face pressure to justify headcount reductions to demonstrate returns on these significant technology investments.

The CFO's role is to ensure that automation enhances the institution's intelligence, not quietly erases it. Execution can and should be automated, but understanding why numbers behave as they do cannot. When circumstances change, institutions do not fail due to a lack of data but due to a lack of human memory needed to interpret that data. We must deliberately cultivate the human expertise required to oversee model inputs and outputs.

Execution can be automated, but explanation cannot.