This is the sixth article in our six-part series on the core pillars of Enterprise Agentic AI. The journey so far has taken us through Safety and Trust (responsible agent design), Control (centralised oversight), Quality (performance monitoring and continuous improvement), High Performing Hybrid Teams (Agentic Org Design) and Infrastructure and Security (hosting and protecting the digital workforce). This final article, exploring everything in regard to commercial control of agents – both financial and risk and liability, is written collaboratively by Rob Price, co-founder of Futuria, together with Stuart Fotheringham , founder & CEO of DealAnalysis.
Agentic AI has the potential to transform business operations and unlock extraordinary value, but only if its costs, risks, and returns are actively managed. This final pillar focuses on the commercial lens: the financial and environmental sustainability of deploying, operating, and scaling agentic systems.
As financial institutions like JPMorgan Chase have demonstrated, scaling agentic AI is not just a technical challenge but a governance imperative. Their 450+ use cases and material productivity gains (reported in the billions) highlight both the upside, and the governance demands. Meanwhile, most enterprises remain unprepared for the complex web of regulatory, legal, and commercial risks autonomous agents create.
Financial perspectives vary significantly across roles. CFOs and Finance Directors focus on strategic cost architecture and enterprise-wide risk. Finance Controllers and IT Finance teams prioritise operational monitoring and system integration. Budget holders and department managers need practical tools for day-to-day cost control and performance measurement.
The commercial dimensions of agentic AI present unique challenges that go far beyond traditional software procurement. This pillar covers six critical areas:
- first, establishing robust financial operations and cost control mechanisms
- second, developing new approaches to pricing and business case validation
- third, managing vendor relationships and concentration risks
- fourth, addressing insurance and liability considerations
- fifth, evolving performance metrics for human-agent collaboration
- and finally, ensuring environmental sustainability throughout the agent lifecycle.
Together, these elements form the foundation for commercially sustainable agentic operations.
FinOps for Agentic AI: Cost Monitoring and Control
Just as cloud adoption demanded new FinOps capabilities, so too does Agentic AI. Enterprises must implement fine-grained monitoring of AI costs, including:
- Token usage across LLM/SLM calls (by task, by agent, by timeframe)
- Compute costs (inference and training, centralised or distributed)
- API usage (and the impact of additional volume on the underlying applications and databases) and third-party model licensing fees
- Idle time vs productive time for agents and agent teams.
In regulated industries, supervisory guidance (e.g. Basel Committee, 2022) now emphasises that institutions retain full accountability for AI-driven financial decisions. This means FinOps reporting must integrate into enterprise-wide risk frameworks, ensuring transparency “commensurate with the risk of the banking activity.”
This visibility enables continuous optimisation, cost forecasting, and sustainability impact measurement. It also supports decisions about switching between model providers, adapting to usage trends, and identifying underused or inefficient agents.
It’s important to look at the commerciality of Agentic AI through different finance viewpoints:
- CFO/FD Perspective: This comprehensive cost visibility enables strategic decisions about model provider diversification, long-term budget planning, and enterprise-wide optimisation strategies
- Finance Controller Perspective: Focus on implementing the monitoring infrastructure, establishing automated reporting cadences, and integrating these metrics with existing financial management systems
- Budget Manager Perspective: Use these granular insights to set realistic departmental agent budgets, identify cost optimisation opportunities, and justify agent-related expenditure.
The thing is, it isn’t that simple, is it? Most organisations will have a range of established software platforms, Microsoft, SAP, Oracle, ServiceNow, Salesforce and many more. Each of them is building agents or enabling employees in organisations to build agents. FinOps needs the holistic view across all of that. It relates back to my point in Pillar 2 of the repository, because we will need to take a view on the value of agents and agent teams. Not only do I want to know that my agents and agent teams do a great job, and reliably so, but I want to know that they are doing it with value for their operating cost. It also isn’t just about the agents ‘operating cost’ either is it, because, as we hinted above, if I create an agent that uses tools (MCP or otherwise) that create massive volume increases on existing systems – then the use of those agents will cause a consequential uplift in the costs across the infrastructure estate. This can affect licensing costs, application latency, data platform throughput, and user experience. So we need to be measuring that too!
Commercial governance must extend to forecasting and managing indirect usage patterns to avoid degraded performance, unexpected charges, or operational bottlenecks elsewhere.
Pricing Strategy and Cost Architecture
There isn’t an emerging standard in the market yet for the way in which the supply chain price these solutions. Name a unit, and someone will have priced it that way – price per agent, price per workflow, price per connector, priced per token, even priced by some random made-up new unit of work. There are challenges with many of these approaches.
“The lack of pricing standardisation creates strategic procurement challenges and makes multi-year budgeting difficult. We recommend focusing on negotiating flexible terms that allow for agent and model evolution while maintaining cost predictability.”
Variable Cost Management: The tension between predictable budgeting and inherently variable agentic workloads presents a fundamental challenge. Unlike traditional software licensing, agentic systems have multiple cost variables—token consumption, API calls, compute cycles, storage—that fluctuate based on agent behaviour, not just usage volume. Organisations need frameworks for setting cost boundaries while allowing agents autonomy to deliver value, or for suppliers to take on the responsibility for wrapping certain variable costs into fixed units. Consider implementing spending alerts, monthly caps, and approval workflows for high-cost agent activities to maintain financial control.
Agent Collaboration Economics: The “more agents, better outcomes” paradox creates complex pricing dynamics where increased agent collaboration improves results but can exponentially increase costs. Multi-agent orchestration, inter-agent communication overhead, and the value premium of collaborative intelligence all require new pricing models that balance cost control with outcome optimisation. How do we optimise towards that balance lever of the ‘right quality’ for the ‘right price’?
Agent Digital Wallets: An emerging approach involves giving agents self-managing budgets where they can make autonomous spending decisions within defined parameters. This includes implementing fraud prevention mechanisms, spend approval workflows, and comprehensive audit trails for agent-initiated expenses. Digital wallets enable agents to optimise their own resource usage while maintaining organisational cost control – though of course this doesn’t guarantee a right or good answer, just one that has been achieved within the budget. One could of course argue that is much like human project delivery!
Agent digital wallets will require robust approval workflows, spending limit controls, and integration with existing expense management systems. They will also critically need clear audit trails and exception reporting for compliance purposes.
Business Case and ROI Validation
Beyond POC Metrics: Many early implementations show impressive demonstrations but struggle with production-scale ROI. The recently published MIT report highlighted this point with organisations struggling to scale POCs. Similarly, McKinsey research identifies the need for board-level AI risk reporting with specific KPIs—model accuracy, drift detection, bias scores, explainability, and audit trail completeness. These must be tied directly to ROI validation if investments are to withstand scrutiny. We’d argue that’s because the majority aren’t thinking about Operationalising Agentic AI in the way we have described in these Pillar articles. Organisations must focus on measuring true business impact beyond technical metrics—emphasising outcome-based KPIs, productivity gains, and revenue attribution rather than just efficiency measures. Whilst ‘Digital Workers’ aren’t human replacements, we do strongly believe it is increasingly important to consider them more like a workforce, than as a system. They need to evolve, learn and optimise towards improved performance, and of course at an improving price point.
ROI validation should, therefore, move beyond proof-of-concept metrics to enterprise-scale value, focusing on outcome attribution and long-term functional productivity gains rather than short-term efficiency.
Boards are beginning to apply stress-testing methods (e.g. Monte Carlo simulations, extreme-event scenarios (‘black swans’)) to agentic workloads to better understand long-tail financial risks. These techniques, borrowed from financial risk management, provide quantitative evidence for business case validation.
Scale Economics: The cost curve paradox where initial deployments are expensive per unit of work, but value often only emerges at scale. A framework for modelling the investment valley between proof-of-concept and profitable operation is essential for securing continued investment and managing stakeholder expectations. We have increasingly referred to our most impactful multi-agent teams as Foundational, in the sense that their capacity for complex work and functional productivity is such that they make the business case work on their own, vs. other teams that may do helpful work, but wouldn’t make sense in their own right. A simple example might be our Due Diligence Foundational Multi-Agent team, that would connect to a data room and massively reduce the manual effort to produce a risk analysis, vs. for example, our Policy Updater (Bronze level) Multi-Agent team that does a useful task – updates policy in line with latest regulation/legislation – but you’d probably only use once a year per policy.
The “investment valley” framework is critical for board-level discussions and securing continued funding. Use the Foundational vs. Gold / Silver / Bronze classification to prioritise investment and set realistic expectations for different agent categories. Understanding the distinction between the functional productivity level of agent teams directly impacts resource allocation decisions and helps justify departmental agent investments.
Model Migration Burden: The hidden costs of model updates and changes when agents are tightly coupled to specific LLMs represent a significant operational reality. Organisations (either directly or through their supply chain) must plan for model versioning, regression testing, and behavioural drift management as ongoing operational expenses, not one-time implementation costs. We are already seeing early signs of this with postponed obsolescence dates of models, which we suspect relates to the complexity of upgrade and the need to refine the agent dynamics. Model migration costs should be treated as recurring operational expenses, not capital investments. Budget for regular testing, retraining, and potential performance degradation during transitions.
Vendor Concentration and Interoperability Risk
Model Portfolio Strategy: Rather than standardising on a single LLM provider, maintain a diversified portfolio with clear switching cost analysis. This includes:
- Abstraction layers that allow model swapping without, or at least minimising, agent retraining
- Regular benchmarking across providers for cost/performance optimisation
- Maintaining relationships with 2-3 providers to ensure competitive leverage.
We also recommend the ability to use small language models (SLMs) in some agents where appropriate – align the complexity and power (and therefore cost) of the model to the level of work being done by the local agent.
Multi-vendor strategies are increasingly recommended by regulators as safeguards against systemic concentration risk. The Financial Stability Board identifies four channels of systemic vulnerability: procyclical amplification, third-party dependency, training data correlation, and hallucination under novel conditions. Enterprises need explicit strategies to mitigate these.
Knowledge Asset Portability: To enable a more seamless evolution through a dynamic evolving space, organisations should consider architecting, or acquiring, their agent knowledge systems to be provider-agnostic:
- Standardised knowledge representation formats
- Agent learning and fine-tuning data (if you want to fine-tune, we prefer not to) maintained in portable formats
- Business logic and workflow definitions separated from model-specific implementations.
Exit Strategy Planning: Every agent deployment should include a clear decommissioning and migration plan, including:
- Data extraction and knowledge preservation procedures
- Rollback capabilities during model transitions
- Cost/benefit analysis thresholds for switching vs. maintaining existing systems.
Enterprise Risk Management and Insurance Frameworks
Agentic AI introduces five categories of enterprise risk requiring board-level attention and specific mitigation strategies. For CFOs, these risks must be integrated into existing enterprise risk frameworks and monitored at board level, with each category requiring specific mitigation strategies and ongoing governance oversight.
Secure multi-agent framework architectures can address many of these risks proactively through design rather than reactive controls. Built-in explainability, comprehensive audit trails, controlled autonomy boundaries, and distributed decision-making reduce single points of failure while maintaining transparency for compliance and insurance purposes. Nonetheless, it is important to be aware of the risk profile.
- Operational Risk: Agent failures causing business process disruption, including cascade effects where agent errors propagate across interconnected systems. Unlike traditional system failures, agent errors may not be immediately apparent, complicating detection and containment
- Reputational Risk: Autonomous agent decisions affecting brand and customer trust. A single poorly performing agent in customer-facing roles could generate thousands of inappropriate responses before detection, amplifying reputational damage far beyond traditional software failures
- Financial Risk: Unpredictable cost escalation from variable usage patterns, vendor concentration exposure, and hidden indirect costs affecting enterprise infrastructure. This includes the risk of budget overruns from agents that operate autonomously within defined parameters but exceed expected resource consumption
- Compliance Risk: Regulatory obligations for AI governance, explainability, and accountability under emerging frameworks like the EU AI Act. Traditional software compliance approaches don’t address the unique challenges of autonomous decision-making systems
- Strategic Risk: Competitive disadvantage from poor agentic AI adoption, over-reliance on single vendor approaches, or failure to scale beyond proof-of-concept implementations.
Agent Error Insurance: Traditional professional indemnity and errors & omissions policies aren’t designed for AI agents making autonomous decisions. New insurance products are emerging, or might need to emerge, that cover:
- Agent-generated content that causes reputational damage or legal liability
- Financial losses from agent trading/purchasing decisions within approved parameters
- Regulatory violations when agents misinterpret compliance rules
- Data breaches caused by agent behaviour rather than system vulnerabilities.
CFOs should engage with insurance brokers early to understand emerging products and coverage gaps, considering the cost of additional coverage against the potential liability exposure from agent decisions. Architecture choices that provide built-in risk mitigation may reduce insurance premiums and coverage requirements. Of course, we would also point back to the discussion in Pillar 3 on Quality and Performance, on the controls that should be put in place to minimise any likelihood of these events.
Market evidence is starting to emerge. Lloyd’s, in partnership with Armilla, has launched AI-specific coverage products protecting against financial losses from underperforming systems and third-party harm. Deloitte projects $4.7 billion in global AI insurance premiums by 2032, paralleling the trajectory of cyber insurance. Case studies illustrate the value: Aviva’s 80+ AI models cut claims assessment time by 23 days, saving £60m in 2024, while Ping An’s AI-driven fraud detection saved over $1bn. These examples highlight both AI’s upside and the growing necessity of coverage frameworks.
Liability Attribution Framework: When an agent makes a costly mistake, determining responsibility becomes complex. Is it the organisation that deployed it, the vendor that provided the model, or the team that configured its parameters? Clear contractual frameworks and insurance coverage need to address this “liability gap” where traditional accountability models break down, though it is a complex area that will no doubt evolve.
Secure multi-agent architectures with comprehensive decision logging and explainable AI components provide the evidence chains necessary for liability attribution, reducing the “black box” problem that complicates responsibility determination.
At this stage, requiring that providers of agents and multi-agent systems are able to provide fully auditable, explainable decisions on the work of their agentic solutions at least provides the future evidence chain that you may need.
Legal precedents are beginning to emerge. The Mobley v. Workday case established that AI vendors can be held liable for discriminatory decisions, while Air Canada was forced to honour discounts offered by its chatbot. Tesla Autopilot cases further illustrate the complexity of causation in AI-related financial losses. These precedents signal a future where liability will be tested in unpredictable and high-profile ways across industries.
Finally, cross-border data and IP concerns are increasingly material. For instance, AI-generated content may not qualify for copyright protection in some jurisdictions (e.g., US), while others (UK, China) allow partial recognition. This introduces potential commercial risks when deploying agentic systems across markets. Organisations must treat IP and licensing as integral to financial risk governance, not afterthoughts.
Agent Audit Trails for Claims: Insurance providers increasingly require detailed logging of agent decision-making processes, including reasoning chains, data sources consulted, and confidence levels. This could create additional operational costs but is essential for claims validation and risk management. Finance Controllers should budget for enhanced logging capabilities and ensure integration with existing compliance reporting systems, choosing architectures that provide audit trails as a built-in feature rather than an expensive add-on.
It is worth noting that we, as Futuria, have baked in that explainability and logging into our architecture design from the start. Whilst it is not always needed or consulted in the delivery of agentic work, it is this point about a persistent audit chain being available, combined with our focus on digital responsibility, that led us to include it. This architectural approach to risk mitigation – building in explainability, audit trails, and controlled autonomy boundaries – addresses many enterprise risk categories proactively rather than reactively.
Performance Metrics Evolution
Traditional productivity metrics, such as “Revenue Per Employee” become misleading when agents augment human work. We need to consider the alternatives:
Revenue Per Human-Agent Team: Measuring the combined output of human-agent collaborations rather than treating them as separate units. This avoids the trap of simply maximising agent count while recognising that the best results often come from thoughtful human-agent partnership.
Value Density Metrics: Revenue per “cognitive unit” – where both human hours and agent processing time are normalised into comparable units based on their contribution to outcomes. This prevents the race to simply deploy more agents without considering their actual value contribution.
Capability Expansion Rate: How quickly teams can take on new types of work or serve new markets when augmented by agents, rather than just doing existing work faster.
Quality-Adjusted Productivity: Revenue per employee/agent adjusted for outcome quality, error rates, and rework requirements. Pure speed metrics can incentivise quantity over quality and create dangerous misalignment between measurement and value.
Some financial institutions are experimenting with “AI-adjusted ROE” (return on equity) that weights productivity gains against new categories of risk capital allocation. This signals a shift where performance metrics are not just about value creation but also about resilience under regulatory stress tests.
These evolved metrics are essential for board-level reporting and investor communications. Focus on value density and capability expansion as strategic indicators of agentic transformation success.
Carbon Impact: Tracking and Minimising Footprint
Whilst we recognise that there are differing political views in different geographies, and indeed within different market sectors, on climate change and sustainability, then it is also clear that there are many organisations committed to measure and reduce their impact on the planet. Some may use the UN Global Compact and SDGs to consider topics such as Energy and Water, and some may have started to use the Global Digital Compact, thinking specifically about the impact from technology.
In an Agentic world, we can choose to implement solutions in ways that are potentially more sustainable than the simple creation or use of large AI models. As a recent podcast guest said, ‘it is less about how powerful your models are, and more about what impact you choose to make by using them’.
Organisations adopting Agentic AI can actively measure the carbon footprint of their agentic operations—by workload, model, and location. LLM inference can be energy-intensive, so wherever possible, lighter-weight SLMs or task-specific models can be used. Right-sizing infrastructure, reducing unnecessary model calls, and shifting workloads to lower-impact platforms all contribute to reducing emissions.
Sustainability practices should be embedded into both agent design and runtime operations, as part of a broader AgenticOps discipline that balances performance, cost, and impact.
As enterprises scale their agentic estate, sustainability becomes a real concern in the management of obsolescence. Agent lifecycles should include retirement of underused or obsolete agents to avoid resource waste and reduce surface area for security or governance oversight, plus consideration given to assessment of workloads across data centres and edge compute (see Pillar 5) to balance performance with impact.
Agent obsolescence isn’t just a technical issue—it’s a governance responsibility. Enterprises must regularly assess which agents continue to provide value, which require refactoring, and which should be sunset to reduce operational complexity and risk.
Licensing Models and Commercial Risk
Understanding the full licensing picture—across open source, commercial APIs, and embedded technologies—is vital. Questions to ask include:
• Are usage tiers clear and predictable?
• Are agent-generated outputs subject to reuse restrictions?
• Are your LLM, agent, and multi-agent vendor’s roadmaps aligned to your priorities?
Commercial risk also arises when agents begin ingesting or relying on external or uncontrolled data sources—particularly if employees unintentionally expose sensitive material by feeding it into agents or model interfaces. Even well-intentioned actions (e.g., uploading a policy document to “make the agent smarter”) can inadvertently breach licence terms, contracts, or regulations.
Organisations must design controls and guardrails that govern both what data agents can access, and how employees interact with them, especially when proprietary or third-party content is involved.
Beware the Ripple Effects
Business Continuity and Recovery Planning: Agent dependencies create new categories of business risk. What’s the financial impact if your agent infrastructure becomes unavailable? Organisations need to establish Recovery Time Objectives (RTOs) for critical agents and understand when agent failure becomes business-critical failure.
The cost of agent downtime must be weighed against the investment in redundant systems, backup models, and failover capabilities. For agents handling customer interactions or critical business processes, the business continuity requirements may be as stringent as those for core enterprise systems—with corresponding infrastructure and operational costs.
The long-tail risk of cascading failures is amplified in financial services, where AI-driven decisions can propagate across markets in minutes rather than days. Regulators are actively considering circuit breaker requirements, mandating system-wide halts under extreme conditions. Enterprises should adopt similar approaches in internal agentic architectures.
Wrapping It Up: Maturity, Not Perfection
These six pillars don’t define a rigid checklist. Rather, they describe a direction of travel—a maturity path toward responsible, high-performing agentic operations.
Some organisations may start with strong infrastructure but need to build out governance. Others may prioritise cost control while figuring out human-agent team dynamics. That’s fine.
What matters is recognising that Agentic AI introduces new kinds of responsibility—and new kinds of opportunity. With clear intent, measurable progress, and adaptive governance, it’s possible to move deliberately and effectively toward an operating model that works.
For financial institutions in particular, the roadmap includes establishing board-level AI governance committees, negotiating more balanced vendor liability contracts, implementing multi-vendor strategies, stress testing agent portfolios, and exploring AI-specific insurance coverage. Those that act early will be better positioned when high-risk system obligations apply from August 2026 and as US/UK frameworks evolve.
The Six Pillars of Agentic Governance:
- Safety and Trust – Why you’re using agents and agent teams, and ensuring they are operating responsibly
- Control and Governance – Ensuring the Agentic Mesh, the Enterprise estate of agents and multi-agent teams, operates coherently at scale
- Performance and Quality – Continuously monitoring and benchmarking output, even where baselines don’t exist yet
- High Performing Human-Agent Teams – Designing the best blend of human and machine capability and interaction
- Infrastructure – Secure, up-to-date, and performant platforms with resilient AgenticOps and ModelOps
Commercial Sustainability – Managing cost, risk, legal exposure, environmental impact, and lifecycle value.
If you’re working thoughtfully across these domains, you’re not just playing with agents—you’re actively engineering the next generation of your business.
And that’s the point: the Agentic Enterprise isn’t a product or a project. It’s a posture. One that sets you up to go further, faster, and with control.
If you would like to discuss or further explore any of the points in any of the Six Pillars of Operationalising Agentic AI, then please contact me at rob.price@futuria.ai, or on linkedin.
It has been suggested that we turn this series of articles into a white paper, or even a book. If you are interested in supporting that initiative, or indeed publishing such a book, then please get in contact.
We’ve written this series of articles because we passionately believe in the ability of multi-agent systems (a subset of Agentic AI) to be able to deliver complex work within organisations that delivers massive value around operational resilience and improved performance. It is so easy to get it wrong though, but that should not deflect from the achievement of the positive goal.
Through my pioneering work in Corporate Digital Responsibility over the past decade, we are also passionate about thoughtful positive impact on people, and we aim to do that by innovating to augment and support people to be, in the words of one of our customers, incredible.
Note: This article was written with the assistance of ChatGPT and Claude for some content generation and refinement.
