The previous generation of AI in finance was good at answering. Ask a question, get a number. Ask a follow-up, get another number. Useful, but still a chat. The orchestrator is what turns the answer into a posted journal, an approved scenario, a refreshed forecast, a board-ready commentary, a routed approval, and a logged audit trail. Insight without action is a slide. The orchestrator is what makes the action happen.

For an FP&A function, three workflows define what good looks like.

Automated data gathering

Agents retrieve and validate data from ERP, CRM, treasury, HR, and source systems. The manual reconciliation burden that has historically consumed the first three days of every month-end disappears. The agent reads the sub-ledger, matches against the GL control account, drafts the correcting journal where required, routes it for approval, and logs the evidence. The analyst arrives on day one of close with a clean data foundation, ready to ask better questions.

The shift this produces is structural. When data gathering is no longer manual, the analyst's time reallocates to interpretation. The bottom of the value pyramid disappears. The middle gets compressed. The top gets bigger.

Intelligent variance analysis

Agents monitor revenue, cost, and cash against forecasts continuously, not just at close. Anomalies surface as they happen rather than when the management pack is built. Commentary drafts itself, anchored in the underlying transactions, before the variance meeting starts. The CFO walks into the review with the first cut already written, the supporting evidence already attached, and the proposed response options already drafted.

This changes the rhythm of finance from periodic to continuous. The week-three flux conversation becomes a week-one conversation. The board meeting becomes a debate about decisions, not a recap of facts.

Dynamic scenario planning

Agents model what-if scenarios on demand. They take a verbal question ("What happens to our coverage ratio if rates compress by 25 basis points and we lose 5 per cent of our top-quartile customers?") and translate it into a structured run against the planning model. They return the result with a chart, a narrative explanation, the assumptions used, and the confidence band. They distribute the answer in natural language, in the channel where the decision will be made.

Scenario planning stops being a quarterly exercise. It becomes a working tool that anyone with the right access can use, with the governance and traceability that the audit committee requires.

The marriage metaphor is a useful frame, but it is not a strategy. The harder work for a CFO is recognising that agentic AI is not a productivity feature bolted onto the existing operating model. It is a new operating model. The technology stack, the team shape, the controls, and the metrics all have to change together. The teams that try to retrofit agents onto last decade's process map will see the modest productivity gains the technology can deliver in isolation, but they will miss the order-of-magnitude improvements available to teams that redesign the function around what agents make possible.

There are four operating-model challenges every CFO will need to answer. None of them is optional. All four interact with one another, and the technology choices made today will either enable or constrain the answer to each.

Challenge 1. The massive redistribution of work between humans and agents

The agentic era marks a radical redivision of labour between people and virtual workers. It is already happening. Across roughly one-third of all occupations, AI systems perform more than a quarter of the tasks. The numbers in finance are higher than average, because finance is data-heavy, rule-heavy, and process-heavy, which is exactly the territory agents do well in.

For FP&A specifically, this means several things at once. The proportion of analyst time spent on data preparation, consolidation, reconciliation, and basic commentary drops sharply. The remaining human work concentrates at the decision points: approve, reject, escalate, redesign. New roles emerge that did not exist five years ago: agent product owner, prompt designer, scenario architect, controls steward for autonomous systems. Performance management shifts from measuring output volume (how many reports, how many models, how many reconciliations) to measuring decision quality (how many of our scenarios turned out to be useful, how many of our flux explanations were right the first time, how many of our recommendations were acted on).

Recruitment changes. The graduate hire of tomorrow needs less Excel mastery and more business acumen, more agent fluency, more capacity to challenge the output of an autonomous system. Training needs to shift accordingly. Compensation models need rethinking. When an agent does 60 per cent of an analyst's old role at 5 per cent of the marginal cost, what does the analyst now own that justifies their pay grade? The answer is not "less work for the same money." It is "harder work, at a higher level, with a much larger span."

The technology implication is that the operating system needs an orchestration layer that governs, monitors, and optimises autonomous decisions in real time. It needs to track which work is being done by which actor (human or agent). It needs to allocate new work to the right actor based on complexity, risk, and capacity. It needs to surface performance attribution so that finance leaders can manage virtual workers with the same rigour they manage human ones.

For the CFO, the practical question is: what proportion of the function should be agent-driven in twelve months, twenty-four months, and thirty-six months, and what is the org-design implication at each milestone? Answering that question well is the difference between gradual margin improvement and structural transformation.

Challenge 2. Agents moving into core, high-value workflows and processes

AI agents are extending automation into judgement-based processes such as forecasting, decision making, capital allocation, and risk management. This is where the value lives, and it is also where the danger lives. A recommendation engine that misallocates A$50m of capital is materially worse than a chatbot that gets a customer query slightly wrong.

The technology implication is that the operating system must support safe autonomy. Every agent action must be auditable, explainable, and resilient. Human-in-the-loop capabilities are non-negotiable for high-stakes decisions. Embedded oversight capabilities are required, including simulation environments where new agent behaviour can be tested before deployment, rollback mechanisms when something goes wrong, and continuous monitoring for compliance, accuracy, and bias. Reliability and assurance become as central to the operating system as speed and scale.

For FP&A this manifests in concrete ways. When an agent recommends a A$50m capital reallocation between business units, the CFO needs to be able to ask three questions and get clear answers. First, what assumptions did the agent use, and which of those assumptions came from primary disclosure versus inference? Second, what is the confidence band, and what would need to be true for the agent to change its recommendation? Third, who approved the recommendation, when, and under which controls? If the operating system cannot answer those three questions on demand, the recommendation cannot be acted on safely. If the operating system answers them by default, the recommendation can be acted on quickly.

There is an organisational point here as well. As agents take on more judgement, the human role shifts from doing the work to validating the work. This is psychologically harder than it sounds. The analyst who used to feel ownership through building a forecast now feels ownership through challenging one. That requires a different skill set, a different management approach, and a different culture. The CFO who recognises this early and invests in the cultural transition gets ahead. The CFO who does not see it ends up with a frustrated team and a half-deployed agent stack.

Challenge 3. Continuous adaptation and real-time flexible optimisation

Because agents can learn, adapt, and self-optimise, change becomes continuous. The agentic system fine-tunes operations to maximise outcomes. This challenges traditional ideas of stability and control in a finance function that has historically been built on the principle that this quarter's process should look broadly like last quarter's process.

The technology implication is that operating systems need to be modular, composable, and self-adaptive. New tools, data sources, and agent capabilities must be integrated without disrupting the core. Simulation environments are required so potential changes can be tested before they go live. Real-time observability is needed to track performance, safety, and alignment with business goals.

For finance, the planning model itself starts to behave like software. Continuous integration of new drivers. Continuous deployment of new scenarios. Version control over agent logic. Rollback when something breaks. The vocabulary of software engineering enters the FP&A vocabulary, because the FP&A function is now being run as a piece of software with agents as its operators.

The cultural shift here is significant. A planning model that updates itself as actuals land does not feel like a planning model. It feels like a living artefact. The forecast that re-runs itself when a material variance appears does not feel like a forecast. It feels like a prediction engine. The traditional rituals of finance (the monthly close, the quarterly forecast, the annual plan) do not disappear, but they stop being the centre of gravity. The centre of gravity moves to the continuous stream of decisions that agents are now able to support.

For the CFO, the practical question is: which of our current planning rituals are still useful, which are habit, and what new rhythms should we introduce now that the cost of replanning has collapsed? Many teams discover that the monthly close, in its current form, is a remnant of an era when consolidation was hard. When consolidation is no longer hard, the close becomes a verification step rather than a production step. That changes everything downstream.

Challenge 4. The blurring of boundaries

Value creation increasingly happens across networks of interconnected agents. Procurement agents negotiate directly with vendor agents. Customer service agents interact with brand agents to deliver personalised experiences. The edges of the enterprise dissolve into a wider digital ecosystem.

The technology implication is that the operating system needs to be open, interoperable, and secure. Agents must be able to operate across systems, vendors, and clouds. Shared protocols, APIs, and trust frameworks are required to manage identity, permissions, and data exchanges. The Model Context Protocol is the architectural standard emerging to enable this. It gives agents a defined way to navigate a server's data, retrieve structured information, execute operations, and return results that downstream applications can trust.

For FP&A, the blurring of boundaries shows up in several places. The variance analysis agent calls a peer benchmarking agent, for example, to pull data from IBISWorld. The peer benchmarking agent can also call a macro forecast agent that pulls from the Reserve Bank of Australia. The macro forecast agent calls a regulatory watchlist agent that pulls from APRA. Each agent does one thing well. The orchestrator coordinates them. The audit trail captures every call. The finance team gets a unified answer, but the unified answer is composed of work done by many specialised agents operating across many systems.

This is genuinely new. The traditional enterprise architecture was built on the assumption that the enterprise's data lived inside the enterprise and was processed by the enterprise's people. The new architecture assumes the enterprise's decisions are informed by data and processed by agents that span well beyond the enterprise boundary. The legal, regulatory, and commercial implications of that shift are still being worked out. The finance teams that engage early, set the right standards, and build the right trust frameworks will be the ones that benefit. The teams that wait will inherit somebody else's standards.

For the CFO, the practical question is: which agent interactions across the enterprise boundary deliver the most value, and how do we manage the identity, permission, and audit requirements that follow? Answering this well is the difference between a finance function that operates as an island and one that operates as a node in a much larger value network.