Anthropic's Economic Index [1] shows that business and finance jobs have a 94.3% theoretical AI exposure. The question for FP&A leaders is how the function continues to add value when automation takes over existing tasks. In this article, we will explore the emerging role of FP&A Agent Managers.

The FP&A Agent Manager

Joint research by Harvard Business School and Salesforce [3] describes a new role emerging across industries: the agent manager, responsible for orchestrating how AI agents perform, learn, and collaborate with human colleagues. The most effective agent managers come from domain expertise, not technical AI backgrounds. Applied to finance, the agent manager is the FP&A professional whose work is to interrogate how AI systems are producing output, whether their assumptions hold against current commercial reality, and whether the outputs they generate, predictive analytics, scenario analysis, represent an accurate view of the business.

Zillow [2] is what the absence of this capability looks like. In 2021, Zillow shut down its algorithmic home-buying business and took over $500 million in write-downs. The pricing model kept bidding aggressively as the housing market turned, on assumptions about resale value that were no longer holding. Nobody whose role was to interrogate those assumptions and stop the system intervened in time. When AI agents impact finance, producing forecasts and scenarios that shape capital allocation or operational commitments, oversight becomes crucial.

HBR [3] argues that in the agentic era, business units take ownership of the agents that power their workflows: designing them, testing them, governing them. For FP&A, this means CFOs own the governance of AI systems whose outputs affect misstatement risk, forecast accuracy, and the commercial grounding of the analysis the business is about to act on. That ownership cannot be delegated to IT, because the consequences of agent failure are financial, not technical.

BCG's 10/20/70 research [4] frames why this matters commercially: 70% of AI value comes from people, processes, and culture. Designing the FP&A function around agent management is the core of the finance function's AI return on investment.

Designing Roles & The FP&A Agent Manager

MIT Sloan research [5] shows that not all tasks should be automated. For FP&A, data collection, variance analysis, reconciliation, and drafting commentary are easily replaced by AI, but contextual judgement, stakeholder communication, exception handling, and analytical framework design still require humans. Finance leaders must decide which tasks to augment or replace to define the future of finance roles.

FP&A Agent Manager Samuel’s day starts by reviewing a funding request for an AI-powered supply chain system to cut inventory by 20-30%. The business wants rapid approval to capture the projected savings. He reviews the proposition and informs the CFO that the AI system introduces probabilistic forecasting risk that could materially affect inventory valuation assumptions. He recommends combining AI with rules-based controls, flags the tool as "high risk," and requires extra financial oversight. Automated filters limit inventory reductions to 15% below historical safety stock, sending exceptions for human review. These measures support quick AI adoption while protecting COGS reporting accuracy.

The FP&A agent manager profile is built on a stack of capabilities, each of which enables the next. None of them sits at the centre of finance qualifications today. Each layer answers a simple question: can the analyst understand, challenge, and control what the model produces? The agent manager profile demands all four.

Figure 1. The Four Layers of FP&A Agent Management

1. Data and tooling literacy. This is the essential base. Data lineage tracks the journey from source system to AI output, revealing where errors may arise. Database fundamentals, such as joins, keys, classification, and source-system behaviour, allow analysts to trace anomalies. Tooling competence enables analysts to use deployed platforms; without it, they can't examine or understand data.

2. Probabilistic model behaviour. AI outputs are estimates, not exact calculations. Analysts must be aware of model drift, data bias, and common failure modes. Instead of checking outputs like facts, they should assess them for plausibility.

3. Probability theory. FP&A professionals will increasingly need to understand probabilistic outputs rather than treating AI-generated forecasts as precise answers. That means interpreting confidence ranges, recognising bias, and understanding when models may be directionally wrong despite appearing statistically confident.

4. Control design. The financial application layer of the role: applying AO/IC principles to AI-enabled workflows, defining review points, automated checks, exception thresholds, segregation of duties, and downstream reconciliations. This distinguishes the FP&A agent manager from a data scientist: the latter understands the model, while the former ensures it is controlled so misstatements do not reach the financial statements.

The four layers are sequential, not parallel. Each requires the one below it to function. That sequence is also the diagnostic test for whether a function has the capability: an organisation that has invested in tooling without probability literacy has built the foundation without the analytical layer above it. An organisation that has trained on AI risks without a well-designed control framework has built the analytical layer without applying it.

Monday Moves: From Insight to Action

Three actions for FP&A leaders who want to start designing the FP&A Agent Manager role this week.

1. Map workflows where probabilistic output reaches financial reporting or business decisions. The exercise reveals where the agent manager role needs to sit operationally.

2. Identify who currently owns AI governance for these workflows. If the answer is IT or "no one specifically," the role gap is concrete. CFOs cannot delegate this without delegating consequences that land on their financial statements.

3. Audit one AI-supported output against the four operational responsibilities. Pick one workflow. Test whether agent performance is being monitored, whether anomalies trigger root cause analysis, whether handoffs to human analysts are designed in, and whether the impact is being quantified. The gaps are the role's first deliverables.

In the next article, we will examine the disruption of skills development when AI automates entry-level tasks and how to rebuild the talent pipeline.

Sources

1.  Anthropic. (2026, March). Anthropic Economic Index report: Learning curves. Anthropic. https://www.anthropic.com/research/economic-index-march-2026-report
2. Kiger, P. J. (2021, December 21).     Flip Flop: Why Zillow’s Algorithmic Home Buying Venture Imploded. Insights by Stanford Business. https://www.gsb.stanford.edu/insights/flip-flop-why-zillows-algorithmic-home-buying-venture-imploded
3. Srinivasan, S., Wei, S. (2026, February 12). To Thrive in the AI Era, Companies Need Agent Managers. Harvard Business Review. https://hbr.org/2026/02/to-thrive-in-the-ai-era-companies-need-agent-managers?autocomplete=true
4. Boston Consulting Group. (2024, December 12). The Leader’s Guide to Transforming with AI. https://www.bcg.com/featured-insights/the-leaders-guide-to-transforming-with-ai
5. Loaiza, I., & Rigobon, R. (2025). These human capabilities complement AI's shortcomings. MIT Sloan School of Management. https://mitsloan.mit.edu/ideas-made-to-matter/these-human-capabilities-complement-ais-shortcomings