AI Governance Is Not Just About the Model. It's About the Business Process.

Generative AI is changing how companies work. Finance teams are using it to analyze performance, research contracts, explain variances, prepare forecasts and answer questions that previously required hours of work.

The attraction is obvious. AI can process enormous amounts of information and produce useful analysis in seconds.

But there is a problem.

The faster AI becomes, the faster it can make mistakes. And the more connected it becomes, the more consequential those mistakes can be.

An inaccurate answer in a chatbot might be inconvenient. An inaccurate answer used to update a forecast, approve a transaction or communicate with shareholders can create a much larger problem.

That distinction is at the heart of a new publication from the Committee of Sponsoring Organizations of the Treadway Commission (COSO), Achieving Effective Internal Control Over Generative AI (GenAI), published in 2026.

COSO's central message is important: Organizations do not need to abandon established internal control principles to use AI. They need to apply those principles to the new ways AI operates.

AI introduces a different kind of control problem

Traditional business software generally follows defined rules. Given the same inputs and configuration, a conventional calculation should produce the same result.

Generative AI is different.

It can produce convincing answers that are wrong. It can misinterpret a document, overlook an important exception or draw a conclusion from incomplete information. Its behavior can also change when models, prompts, connected data or configurations change.

And unlike a conventional reporting application, an AI agent may be able to initiate actions across several business systems.

This creates a challenge for management: How do you benefit from AI's speed and flexibility while maintaining confidence in the information, decisions and actions it produces?

COSO approaches this question by examining eight capabilities commonly found in generative AI systems, from extracting information and transforming data to forecasting, monitoring, coordinating workflows and interacting with people.

That is a useful distinction. The controls you need depend on what AI is doing, not simply which AI product you purchased.

Consider a real business failure

In November 2022, a customer used Air Canada's website chatbot to ask about bereavement fares after a death in the family. The chatbot explained that the customer could purchase tickets at the regular price and apply for a discounted bereavement fare afterward.

There was one problem. That was not Air Canada's policy.

The airline's published policy did not allow the customer to claim the discount retroactively after travel. In fact, the chatbot linked to a policy page that contradicted its own answer.

The customer relied on the chatbot, purchased the tickets, and subsequently requested the discount. Air Canada refused.

The dispute eventually reached British Columbia's Civil Resolution Tribunal. In February 2024, the tribunal found Air Canada liable for negligent misrepresentation and awarded the customer C$812.02, including damages, interest and fees.

The amount was relatively small. The governance failure was not.

The company had provided information through an automated system that contradicted its authoritative business policy. The customer acted on that information, and the company was held responsible.

This is precisely the kind of problem COSO's 2026 guidance on generative AI is designed to help organizations address.

Consider the control questions:

• Information quality: Was the response grounded in the current, authoritative policy?
• Control activities: What validation prevented the chatbot from communicating an incorrect entitlement?
• Accountability: Who was responsible for ensuring that the automated response was reliable?
• Monitoring: How would the company detect incorrect answers before they affected additional customers?

The tribunal decision does not establish that Air Canada's chatbot used a generative AI model. Nevertheless, the failure illustrates a risk that COSO identifies as particularly important for generative AI: plausible, incorrect information being accepted and acted upon without sufficient verification.

For a finance organization, the same control weakness could appear in an AI-generated interpretation of a customer contract, a shareholder communication, a cash forecast or a recommendation to change financial records.

Five established principles, applied to a new technology

COSO's GenAI guidance applies the five components of its established Internal Control Framework.

Control Environment. Who owns the AI-enabled process? Who is authorized to configure it, use it and approve its results? Management must establish clear responsibilities and expectations, rather than assuming the technology provider owns the risk.

Risk Assessment. What could go wrong? An incorrect summary and an unauthorized financial transaction have very different consequences. AI risks must be assessed in the context of the business activity, including inaccurate information, prompt manipulation, inappropriate disclosure and changes in model behavior.

Control Activities. What prevents, detects or corrects problems? Depending on the risk, controls may include source verification, reconciliation, restricted access, approval requirements, segregation of duties and human review.

Information and Communication. Can the people relying on an AI result understand where the information came from, which assumptions were used, what limitations exist and whether the result has been verified?

Monitoring Activities. How does the company know its controls are still working? An AI process that performs well today may deteriorate after a source-system change, model update or configuration adjustment.

The framework is not asking companies to slow every AI interaction with layers of approval. It is asking them to establish controls proportionate to the risks.

A particularly important distinction: AI assistance versus AI reliance

COSO's guidance draws attention to a distinction that finance leaders should understand.

There is a difference between using AI to assist with a control and relying on an AI output as evidence that the control operated effectively.

Suppose AI suggests matches between invoices and purchase orders.

If a qualified employee independently verifies the matches before approval, the AI is assisting the process.

If management approves the transactions based solely on the AI-generated results, the organization is relying on the AI system to a greater degree.

The second situation requires stronger evidence that the AI and its supporting controls are reliable.

This is why simply adding a human approval button is not necessarily enough. A reviewer must have the information, competence and opportunity to challenge the result.

Human oversight only creates value when the human can exercise meaningful judgment.

The answer is not to keep AI disconnected

One response to AI risk is to restrict it to standalone chatbots, disconnected from company systems.

That can reduce certain access and execution risks, but it also limits what AI can accomplish.

Without business context, users must manually assemble information. They may inadvertently omit important records, work from outdated figures or introduce errors while copying data between systems.

Disconnected AI does not eliminate control problems. It moves many of them into manual processes.

The better objective is to connect AI to the information and workflows it needs, while establishing appropriate controls over each step.

That includes validating information when it enters the system, maintaining business relationships and source references, using reliable calculations where deterministic methods are appropriate, and preserving approval authority for consequential decisions.

COSO specifically encourages organizations to consider whether generative AI is even the right technology for a particular task. Some activities are better handled by conventional rules and calculations.

We agree.

How Farsight approaches the problem

Farsight DS is built around the idea that AI should work within the company's existing business environment, rather than operating as an isolated source of answers.

Farsight DS connects financial, operational and business information from the systems companies already use. It combines that information with decision-support capabilities such as forecasting, contract intelligence, risk analysis and business workflows.

This creates an opportunity to apply controls where they matter most.

When an AI-assisted analysis depends on a contract, the relevant terms should be traceable to the underlying document. When an answer involves financial results, the figures should be grounded in appropriate source data and calculations. When a recommendation requires action, that action should respect the company's permissions and decision authority.

The objective is not to make AI infallible.

It is to make AI-supported work more reliable, reviewable and accountable.

Management remains responsible for evaluating risks, configuring controls, verifying their operation and determining where human judgment is required.

Governance must continue after implementation

AI governance is not a policy written once and filed away.

Models change. Source data changes. Business processes change. New capabilities are introduced, sometimes faster than existing governance processes can accommodate.

Businesses need a recurring cycle: establish governance, inventory AI uses, assess risks, design controls, implement them, and monitor and adapt.

Better decisions require both value and control

The conversation about AI often treats value and governance as opposing forces.

Move faster, or maintain control.

Automate more, or preserve accountability.

Connect more systems, or reduce risk.

That is the wrong way to frame the problem.

Properly designed controls can make broader AI adoption possible. Reliable information, clear authority, traceable analysis and effective monitoring help organizations use AI for more consequential work with greater confidence.

COSO's GenAI guidance provides a practical foundation for that approach.

For finance leaders, the opportunity is not simply to give employees a faster way to ask questions, it is to build business processes in which AI helps people understand what is happening, evaluate what might happen next, and take appropriate action, without losing the controls that make those decisions trustworthy.

AI can accelerate the work. Management must still be able to trust the process.

Reader worksheet · free, no signup

GenAI use-case control assessment worksheet

For

CFOs, controllers, risk owners and process owners.

Inputs

One AI use case: its owner, data sources, output, authorized actions, business impact and current controls.

Output

A documented preliminary control-gap assessment and action plan.

Complete one worksheet per use case. For each row, record a status (Yes, Partial, No or Not applicable), the evidence or gap, an owner and a remediation due date. Start by recording: use case, accountable owner, data sources, output, authorized actions, business impact, current controls.

Status: Yes / Partial / No / Not applicable

#	Control question	Status	Evidence / gap	Owner / due
1	Is the use case inventoried with a named accountable owner?
2	Are purpose, intended users and permitted actions defined?
3	Are approved sources current, complete and traceable?
4	Are extraction, mapping and calculation results validated?
5	Are confidentiality, tenant, role and disclosure boundaries enforced?
6	Have hallucination, prompt injection, drift, fraud and vendor risks been assessed?
7	Are high-impact outputs independently reviewed, with documented approval?
8	Are AI-generated recommendations separated from authority to execute actions?
9	Are model, prompt, connector and threshold changes controlled and retested?
10	Are quality measures, exceptions, escalation and remediation monitored?

Completed illustrative example. Use case: AI drafts shareholder responses from financial records. Accountable owner: Shareholder relations lead. Data sources: Approved financial statements and authorized shareholder documents. Output: A draft response for review, not an external send. Authorized actions: Research and draft only; no autonomous disclosure.

• Are confidentiality, tenant, role and disclosure boundaries enforced? Status: Partial. Gap: Entitlement-based disclosure: recipient-specific rights are not yet confirmed against governing documents. Action: Require an authorized reviewer to verify entitlement and approve disclosure before any external send. Owner: Shareholder relations lead. Due: Before production use.
• Are model, prompt, connector and threshold changes controlled and retested? Status: No. Gap: No regression testing when the model changes. Action: Establish a controlled test set of representative and adversarial questions, and retest on model, prompt and retrieval changes. Owner: AI process owner. Due: Before production use.

Other rows remain unassessed, not presumed compliant.

Download the worksheet (Excel)

A completed worksheet is not a certification or a formal COSO effectiveness opinion.

Sources

• Committee of Sponsoring Organizations of the Treadway Commission (COSO), Achieving Effective Internal Control Over Generative AI (GenAI), 2026.
• Moffatt v. Air Canada, 2024 BCCRT 149 (British Columbia Civil Resolution Tribunal).