Human-in-the-loop: a practical AI control framework for finance leaders

What if the biggest risk in AI adoption isn't the technology? It's treating AI like a tool instead of a team member.

With 92% of CFOs already using AI in their finance operations, the challenge is no longer whether to adopt AI but how to integrate it effectively. No finance leader would onboard a new team member without clear roles, responsibilities, and oversight, yet AI is often implemented without these critical elements.

AI delivers value when managed like a skilled professional: working within a team with well-defined responsibilities and appropriate supervision. A human-in-the-loop (HITL) framework gives finance teams a practical method for deciding what to automate, what to augment, and when to reserve decisions for experienced professionals.

What human-in-the-loop means for finance

Human-in-the-loop provides a simple framework to align AI adoption with the core realities of treasury: cash visibility, liquidity risk, payments control, compliance, and executive decision-making. Instead of asking "Where can we apply AI?", the method asks a better question: What level of human involvement should each process require to be fast, safe, and explainable?

To answer the question, teams must assess processes along two critical dimensions:

• Frequency: How often does the process run: daily, weekly, monthly, or ad hoc?

• Strategic importance: If the process fails, what is the impact on liquidity, risk, compliance, reputation, or executive decisions?

Framing AI implementation along these dimensions moves the discussion from technological experimentation to strategic operational design. The conversation shifts from “Can we do this?” to “Should we do this, and how?” Leaders can deploy AI where the technology will create leverage, while ensuring humans remain in charge where accountability and business judgment are paramount.

An intelligence of imitation, not creation

Contrary to popular belief, AI does not "invent" in the way humans do. Generative AI produces outputs by learning patterns from vast datasets. Those datasets are ultimately created, selected, and labeled through human activity and business systems. The reliance on human-generated data is precisely why AI cannot be treated as a standalone actor in finance.

If AI systems increasingly train on content produced by other AI systems, rather than on grounded, real-world business data, performance can degrade over time. Some researchers describe the phenomenon as the "Habsburg effect," a metaphor referencing the risks of "inbreeding" within a closed system.

When a model feeds on its own outputs, the model can drift, amplify bias, and gradually lose connection to reality. In treasury terms, that might mean AI-generated forecasts that look statistically valid but miss the operational drivers your team knows matter. The lesson is clear: AI cannot function on its own. AI requires a human environment, proper supervision, and strong business grounding.

In finance terms, the AI risk is not abstract. Confident-sounding forecasts can become less anchored to operational reality, anomaly detection may start "learning the wrong normal," and reporting narratives can converge on generic, unhelpful explanations. The remedy is not to avoid AI (though some teams may be tempted). Embedding AI in a team dynamic with supervision, feedback, and strong business grounding delivers sustainable value.

Machine learning vs. generative AI: understanding the two engines of AI in finance

Finance teams today work with two distinct categories of AI, and confusing them leads to misplaced expectations. Machine learning (ML) and generative AI are built differently, trained differently, and suited to very different tasks. Knowing the distinction helps finance leaders choose the right tool and set the right governance expectations for each.

Machine learning models are trained on structured, numerical data to recognize patterns and make predictions. A fraud detection model, for instance, learns from thousands of historical transactions to score new ones by risk. A cash forecasting model learns the relationship between past cash flows, seasonality, and business drivers to project future balances. These models are precise, auditable, and reliable within the narrow domain they were designed for. Their outputs are deterministic: given the same inputs, they produce the same outputs. For finance, that predictability is a strength. ML-based systems excel at anomaly detection, reconciliation matching, budget code allocation, and structured forecasting.

Generative AI models operate on a fundamentally different principle. Tools like GPT from OpenAI are trained on vast amounts of text to understand and generate language. Rather than predicting a number, they predict the next most likely word or token, given the context. This approach makes LLMs exceptionally capable of tasks that require language fluency: summarizing a liquidity report, drafting a board commentary, answering a question about a covenant, or explaining why a cash variance occurred. Unlike ML models, LLMs are generative and probabilistic, meaning the same prompt can produce slightly different responses. Their value lies not in numerical precision, but in the ability to translate data into meaning, synthesize unstructured information, and accelerate human workflows that rely on communication and interpretation.

The most powerful AI deployments in finance combine both. ML models generate the numbers; LLMs translate those numbers into actionable narratives. A cash forecasting workflow, for example, might use ML to compute the forecast and surface the key drivers, then use an LLM to draft the management commentary in the appropriate tone and format for leadership review. Neither model replaces the other, and neither replaces the finance professional who validates the output and owns the decision.

The four operating modes for AI in finance

Understanding which type of AI is at work in each process is fundamental to applying the right level of human oversight. That's where the four operating modes come in.

Like any finance professional, AI excels only when working as part of a team. Once frequency and strategic importance are assessed, most finance processes fall naturally into one of four operating modes. The modes give teams a shared language for defining roles, responsibilities, and controls.

1. Human-owned (full control): Humans execute and decide. AI may provide information (summaries, search results, comparisons) but does not generate the final output. The mode is best suited for highly strategic decisions, novel situations, and any task where context and responsibility cannot be delegated. Examples include M&A financing strategies or negotiating new credit facilities.

2. Human-led with AI assistance: AI drafts, recommends, highlights drivers, or flags inconsistencies. Humans validate, adjust, and approve the final output. The human-led mode is often the highest-return starting point for finance teams because it improves speed without compromising sign-off or accountability. Think of the approach as giving your team a highly capable analyst to prepare the first draft.

3. AI-executed under human supervision: AI runs the workflow and produces outputs, while humans supervise outcomes through thresholds, exception handling, and review queues. People do not review everything; people review what matters. The focus shifts from manual execution to oversight and quality control.

4. AI-autonomous, but fully auditable: AI executes end-to-end within strict guardrails and comprehensive logging. "Autonomous" here never means "invisible." The term means the process is stable and controlled enough to run without constant human attention, while still leaving a complete, transparent trail of inputs, outputs, and decisions for audit and continuous improvement.

A common mistake is treating all processes as Mode 4 (autonomous) too quickly. When payment approvals run without exception handling, a single edge case can create downstream problems that take weeks to resolve. Speed without guardrails creates liability, not efficiency.

For CFOs and finance leaders, the central idea is clear: the goal is not automation for its own sake. The true objective is repeatable performance within a control environment that can be defended, both internally and externally. Without clarity on ownership, control points, and evidence, teams either generate outputs that are fast but difficult to trust or create outputs that are trustworthy but too slow to scale.

What success looks like: treasury use cases

Human-in-the-loop success is best demonstrated through practical use cases. The following examples show how the right operating modes and AI governance can drive speed, accuracy, and accountability.

• Cash forecasting (human-led with AI assistance): AI-powered cash forecasting can propose forecast adjustments and surface drivers: changes in collections timing, payroll spikes, or unusual variances. The treasury team reviews what is material, validates assumptions with business context, and owns the narrative presented to leadership. In practice, many teams see compounding returns here: every cycle becomes faster because the team starts from a better first draft and focuses time on what actually changed.

• Payments and anomaly detection (AI-executed under human supervision): AI can monitor payment behavior and flag anomalies: unusual beneficiaries, new banking details, out-of-pattern amounts, or rule violations. The team reviews only the exceptions, clears valid items, and escalates suspicious ones. Over time, thresholds and controls evolve based on validated outcomes, improving detection quality while keeping segregation of duties intact.

• Liquidity reporting (human-led with AI assistance): AI-enabled liquidity reporting can draft an initial explanation for week-over-week cash movement and highlight the likely drivers across accounts or entities. Finance validates accuracy, ensures consistency with actuals, and adjusts the messaging for stakeholders. This use case is especially helpful when leaders need both numbers and narrative quickly.

• Routine reconciliations (AI-autonomous, but auditable): For stable, rules-based reconciliations, AI can match and close items automatically, provided every decision is traceable and exception handling is defined. Autonomy is earned through controls: tolerance rules, change management, and periodic review. The benefit extends beyond time saved: the real gain is consistency and the ability to concentrate expert attention on exceptions.

AI governance that enables speed

Effective AI adoption succeeds when governance is designed to reduce (not add) friction. For finance teams, the practical governance checklist is:

• Data clarity: Establish common definitions (e.g., "available cash"), clear data lineage, and quality rules.

• Access and confidentiality: Enforce least-privilege access, segregation of duties, and data retention policies.

• Auditability: Maintain logs of inputs, outputs, approvals, and model or version changes.

• Value tracking: Tie AI usage to measurable outcomes like cycle time, accuracy, or exception rates.

• Escalation paths: Define what happens when AI confidence is low or a result is material and requires human judgment.

When AI governance is designed well, it enables leaders to scale AI usage safely and with confidence. Governance without enforcement is just documentation. It’s not enough to have controls on paper; they must produce auditable evidence.

Scaling AI with confidence

AI does not replace finance judgment. AI extends finance judgment. Like any new team member, AI performs best with a clear role, a manager, and an operating model that turns capability into consistent, defensible results. The human-in-the-loop framework offers a structured, reliable approach to scaling AI adoption with clarity and control, rather than just enthusiasm.

Of the 92% of CFOs using AI in their finance operations, those who succeed will be those who treat AI like a team member from day one.