A practitioner's guide to building bottoms-up TCO models that survive CFO scrutiny, including the non-obvious costs that naive models underestimate by 40–60%, and the timeline flexibility that an environment changing this fast demands.
Most organisations building TCO models for agentic AI start with a reasonable framework: upfront build costs, ongoing operational costs, three-year projection, done. The problems are what the framework omits and what it assumes will hold still.
Traditional software TCO models account for licensing, infrastructure, implementation, and maintenance. These categories exist for agentic AI, but they represent only 50–65% of true total cost. The remaining 35–50% falls into categories that software-trained finance teams do not instinctively model: coordination tax, human adoption and retraining, governance and security infrastructure, the cost of failures, and the opportunity cost of internal teams diverted from other work.
But there is a more fundamental issue. Traditional TCO models assume relative stability; the technology stack you deploy in year one is recognisably the same stack in year three. Agentic AI does not work this way. Foundation models turn over every 6–12 months. Agent marketplace offerings that do not exist today will be production-ready next quarter. Pricing for grounded search, which currently runs seven times the cost of standard retrieval, will shift as competition intensifies and architectures mature. Open-source alternatives that are experimental today will be enterprise-viable tomorrow.
Organisations that discover these costs after deployment face an unpleasant recalculation: the agent system that appeared to deliver 5:1 returns actually delivers 2:1 or 3:1 when fully loaded costs are included. The investment is still sound (2:1 returns over three years are excellent) but the credibility damage from overstated projections undermines organisational confidence in subsequent AI investments. Getting the model right the first time matters for reasons beyond arithmetic.
This article provides a complete TCO framework, including worked examples calibrated to financial services deployments. The goal is not precision; agentic AI economics involve genuine uncertainty that honest models acknowledge. It is completeness and flexibility. Every material cost category should appear in the model, even if the estimate carries a wide range. And the model itself must be structured for the pivots that this environment demands.
Enterprise software TCO is typically measured over standard periods: three years in commercial enterprise, five years in public sector (often structured as a three-year base with two option years), and occasionally seven or ten years for major infrastructure investments. Each period includes cost-of-living adjustments (typically 3–5% annually) applied to labour, licensing, and managed services.
Agentic AI does not fit neatly into any of these standard horizons, because the technology and market are evolving faster than the planning cycle. The model you deploy on a three-year horizon may be obsolete in eighteen months: not broken, but superseded by something materially cheaper or more capable. The agent marketplace that barely exists today will be a major procurement channel by year two. The governance platforms that require substantial custom investment now will be commercially available as managed services within the planning window.
The practical recommendation is a layered time horizon:
Year 1: Committed investment. This is the build year; discovery, architecture, MVP, production deployment, initial governance. Year-one costs are largely fixed once the decision to proceed is made. Model these with specificity.
Years 2–3: Flexible operations. Ongoing costs with explicit optionality. Model consumption costs should assume at least one major model transition (with associated testing and validation costs). Governance costs should assume platform maturation that may reduce custom investment. Agent architecture should assume potential migration to marketplace or SaaS alternatives for components that are currently custom-built. Apply a 3–5% annual cost-of-living adjustment to all labour-dependent costs.
Years 4–5 (where applicable): Option-year planning. For public sector and long-cycle enterprise deployments, years four and five should be modelled as option years with wider uncertainty ranges. The technology landscape in 2029 is genuinely unpredictable from 2026. Model these years at the framework level (cost categories and rough magnitudes) without false precision on specific line items. The value of option-year modelling is not accuracy but preparedness: it forces organisations to identify which decisions must be made now and which can be deferred.
Critical principle: build for flexibility, not optimisation. In an environment changing this fast, the ability to pivot; to swap models, migrate architectures, adopt marketplace agents, leverage open-source alternatives; is worth more than marginal cost optimisation on the current stack. TCO models should explicitly value flexibility. An architecture that costs 15% more but avoids vendor lock-in may deliver better five-year economics than a cheaper architecture locked into a specific provider whose pricing and capabilities are uncertain beyond the next two quarters.
Before modelling costs, organisations must determine where their agents will run. This is not purely a technology decision. For many financial services organisations, it is a decision constrained (or entirely determined) by regulation, security policy, data sovereignty requirements, or institutional risk appetite.
The model most TCO frameworks assume. Agents call commercial foundation model APIs (OpenAI, Anthropic, Google, etc.) hosted in cloud infrastructure. Costs are variable and consumption-based; you pay per token, per query, per retrieval. This is the model described in most of this article's cost layers.
Advantages: no capital infrastructure investment, immediate access to frontier model capabilities, elastic scaling, and the provider handles model updates, security patches, and infrastructure management. Disadvantages: variable costs that are difficult to forecast, dependency on provider pricing decisions, data leaving organisational boundaries, and some organisations are simply blocked from this model entirely.
A significant number of financial services organisations cannot use cloud-hosted foundation models. The reasons vary: regulators that prohibit sensitive data from leaving organisational infrastructure, security policies that classify LLM API calls as unacceptable data exposure, data sovereignty rules in specific jurisdictions, institutional risk appetite that does not tolerate dependency on external AI providers, or straightforward organisational preference for infrastructure they control.
These organisations must invest in local compute. GPU infrastructure (typically NVIDIA A100/H100 clusters or equivalent) running open-source or licensed models within their own data centres or private cloud environments. This inverts the cost structure entirely.
Fixed infrastructure investment. Local compute is a capital expenditure, not an operating expense. A production-grade GPU cluster capable of running current-generation open-source models (Llama, Mixtral, or comparable) at enterprise scale requires $500,000–$2,000,000 in initial hardware investment, plus $100,000–$300,000 annually in facilities, power, cooling, and maintenance. This is a fixed cost regardless of utilisation; the GPU cluster costs the same whether it processes ten thousand transactions or ten million.
The race to consume. Because infrastructure cost is fixed, the economic logic inverts: value depends on utilisation. An underutilised GPU cluster is an expensive paperweight. Organisations must fill available capacity with high-value agent workloads to justify the investment. This creates an internal dynamic that is healthy when it drives adoption of valuable use cases but pathological when it pressures teams to deploy agents prematurely to show utilisation.
The fight over constrained resources. Once capacity is consumed, GPU compute becomes a constrained resource that multiple teams compete for. This is where optimisation and ROI become directly relevant to allocation decisions. When a compliance team and a trading desk both need GPU hours, the organisation must allocate based on value; which agent deployment produces more ROI per GPU hour? This requires the measurement infrastructure described in the budgeting and token economics articles, applied to compute allocation rather than token spend.
Model selection is strategic. On-premises organisations cannot simply call the latest frontier API. They must choose which models to deploy on their infrastructure, which requires ongoing evaluation of open-source model capabilities, fine-tuning costs, inference optimisation, and the trade-off between model quality and compute requirements.
Open-source becomes essential, not optional. For cloud-hosted deployments, open-source models are an optimisation lever. For on-premises deployments, open-source is the foundation. The organisation's entire AI capability depends on the quality, reliability, and continued development of open-source models.
For small and mid-size organisations, and for specific use cases even in large enterprises, the most practical deployment model is purchasing agent capabilities as SaaS applications. KYC verification, compliance monitoring, document processing, customer service, fraud detection: vendors increasingly offer these as turnkey agentic AI services with per-seat, per-transaction, or subscription pricing.
TCO for SaaS agents is structurally simpler but not necessarily cheaper. Subscription costs for agentic AI SaaS typically run $50,000–$500,000 annually per application. That range can exceed the cost of a custom-built agent; but the comparison is misleading because SaaS pricing includes infrastructure, model consumption, maintenance, and governance that a custom build must budget separately. The honest comparison is SaaS subscription versus fully-loaded TCO of a custom alternative, not SaaS subscription versus build cost alone.
Integration is the hidden SaaS cost. SaaS agents still require integration with organisational systems: data feeds, identity management, workflow orchestration, reporting infrastructure. Integration costs run $15,000–$75,000 depending on the number and complexity of connection points.
Governance is shared, not eliminated. The vendor governs the agent's internal behaviour. The buyer retains responsibility for governing data quality into the agent, decision quality out of the agent, regulatory compliance of the agent's outputs, and audit trail completeness for regulatory examination. These responsibilities translate to real cost, typically $15,000–$40,000 annually for compliance-adjacent SaaS agents in financial services.
The critical trade-off is control versus speed. SaaS agents deploy fast and require minimal internal AI expertise. But the buyer cannot inspect the model, modify the agent's reasoning, fine-tune behaviour for edge cases, or ensure the agent meets explainability requirements that the vendor does not prioritise. For commodity processes where regulatory exposure is low and the vendor's standard offering is sufficient, SaaS is often the best economic choice. For processes that are core to competitive differentiation, involve sensitive data, or face stringent explainability requirements, the loss of control may be unacceptable regardless of the cost advantage.
SaaS is also the natural starting point for organisational learning. Organisations deploying their first agent capabilities often do better starting with SaaS; learning what agentic AI can and cannot do, developing internal governance competence, and identifying which processes warrant custom investment, before committing to the complexity and cost of custom builds.
Major cloud and AI providers (Microsoft, Google, AWS, OpenAI, Anthropic, and others) aggressively pursue multi-year committed spend agreements; typically $5 million to $100 million or more in annual commitments, drawn down against AI services consumption over three to five years. These agreements offer substantial upfront discounts (20–40% below list pricing) in exchange for guaranteed annual minimums.
The economic trap is structural. Committed spend agreements convert variable consumption costs into fixed obligations. The discount is real, but it creates three problems that most organisations underestimate at signing:
The consumption ramp never matches the commitment curve. Enterprise AI adoption takes longer than anyone projects. The business cases that justified the commitment assumed deployment timelines that slip, agent performance that ramps slower than modelled, and organisational adoption that encounters resistance. Year one of a $10 million annual commitment might see $3 million in actual consumption. The remaining $7 million is paid regardless; sunk cost that produces no value.
The race to consume distorts deployment decisions. When an organisation is burning $500,000 per month in unused commitment, pressure builds to deploy anything that consumes credits; regardless of whether the use case has been properly validated, the governance is in place, or the ROI justifies the deployment. Organisations report deploying low-value or premature agent applications specifically to draw down commitment, producing deployments that consume resources without generating proportional value.
Lock-in is explicit and contractual. Unlike organic cloud consumption where switching providers means rebuilding integrations, committed spend agreements create contractual lock-in with financial penalties for early termination. The flexibility premium; the ability to pivot to better models, cheaper providers, or open-source alternatives as the market evolves; is explicitly sacrificed for the volume discount.
TCO modelling for committed spend: Organisations with existing committed spend agreements should model AI agent TCO differently. The marginal cost of model consumption may be near-zero (drawing against an already-paid commitment), which makes the ROI maths look excellent; but only if the commitment was going to be fully consumed regardless. If agent deployments are being accelerated to draw down unused commitment, the honest TCO model should attribute a share of the excess commitment cost to those deployments rather than treating consumption as free.
Organisations evaluating new committed spend agreements should model the consumption ramp realistically (not optimistically), calculate the break-even utilisation rate, and compare the net economics against pay-as-you-go pricing with the flexibility to switch providers. In most 2026 scenarios, the break-even calculation favours committed spend only for organisations with mature, high-volume AI operations.
Most large financial services organisations will operate hybrid deployments; some combination of pay-as-you-go cloud APIs, committed spend drawdowns, and on-premises infrastructure. The specific mix depends on regulatory constraints, existing contractual obligations, data sovereignty requirements, and organisational AI maturity. Hybrid deployments combine the complexity of all constituent models and add the coordination cost of managing workload routing between environments.
The nine cost layers in this framework apply to all deployment models, but their relative magnitudes shift dramatically:
| Cost Layer | SaaS | Cloud (Pay-as-you-go) | Cloud (Committed) | On-Premises | Hybrid |
|---|---|---|---|---|---|
| Infrastructure | None (vendor-managed) | Low (provider-managed) | Low (provider-managed) | Very High (capital + ops) | High (multi-environment) |
| Model Consumption | Bundled in subscription | Variable, dominant cost | Fixed obligation (discounted) | Fixed (amortised over utilisation) | Mixed |
| Build & Integration | Low build, moderate integration | Moderate | Moderate (provider-constrained) | Higher (model deployment, fine-tuning) | Highest (multi-environment) |
| Governance & Security | Shared (vendor + buyer) | Moderate | Moderate | Higher (full stack responsibility) | Highest |
| Human Costs | Lowest (vendor handles operations) | Moderate | Moderate | Higher (specialised GPU/ML ops staff) | Highest |
| Flexibility / Lock-in | Vendor lock-in (data + workflow) | Provider lock-in risk | Contractual lock-in (highest) | Hardware lock-in risk | Varies by mix |
| Failure Costs | Vendor-absorbed (partially) | Variable (consumption-based) | Masked by commitment (still real) | Fixed (wasted capacity) | Mixed |
| Unused Capacity Risk | Low (subscription-based) | None (pay per use) | High (commitment vs. consumption gap) | High (utilisation-dependent) | High |
| Internal Expertise Required | Lowest | Moderate | Moderate | Highest | High |
| Control / Customisability | Lowest | High | High (within provider) | Highest | Varies by mix |
The worked examples in this article use cloud-hosted economics, as they illustrate the most cost layers. Organisations evaluating SaaS agents should model subscription + integration + shared governance as primary cost drivers. Organisations planning on-premises or hybrid deployments should adjust Layer 3 (infrastructure) from thousands to hundreds of thousands, shift Layer 4 (model consumption) from variable operating cost to amortised capital cost, and add specialised staffing to Layer 6 (human costs) and Layer 9 (opportunity cost).
Traditional software procurement follows established patterns: build custom, buy commercial, or license SaaS. Agentic AI is developing a more complex and more fluid acquisition landscape that TCO models must accommodate.
Custom-built agents. Designed and developed for the organisation's specific processes, data, and requirements. Highest upfront cost, greatest architectural control, most flexibility to modify and optimise. Appropriate for core competitive processes where the agent's behaviour is a strategic differentiator. The TCO framework in this article focuses primarily on custom-built agents, as they involve the most cost layers.
Marketplace-licensed agents. Agent marketplaces are emerging rapidly. Marketplace agents are pre-built for common use cases (document processing, customer service, data extraction) and licensed on usage-based or subscription terms. Lower upfront cost, faster deployment, but limited customisation and potential lock-in to the marketplace provider's ecosystem. TCO for marketplace agents concentrates in licensing fees, integration costs, and the governance overhead of operating an agent you did not build and cannot fully inspect.
Managed-asset agents. Some agents warrant treatment as managed assets; deployed, monitored, and maintained by a third party with the organisation retaining ownership of the data, models, and decision logic. This model suits organisations that want agent capabilities without building internal AI operations teams. TCO shifts from build and governance costs to managed service fees.
SaaS-embedded agents. Increasingly, agentic AI capabilities are embedded in existing SaaS applications. KYC platforms, compliance monitoring tools, customer service platforms, and document management systems are incorporating agent functionality as features rather than standalone systems. TCO for SaaS-embedded agents may be as simple as an incremental licensing fee; but organisations must account for governance and integration costs that the vendor does not cover.
Agents with proprietary data and APIs. The most valuable agents will increasingly be those bundled with proprietary data assets: regulatory databases, market intelligence feeds, entity resolution graphs, sanctions lists, and purpose-built APIs. These agents command premium pricing because the data is the differentiator, not the model. TCO must account for data licensing costs that may escalate independently of compute and model costs.
The portfolio reality: Most mature organisations will operate a mix of all five models simultaneously. Custom agents for core processes, marketplace agents for commodity tasks, SaaS-embedded agents for platform-integrated workflows, managed agents for capabilities outside internal expertise, and data-bundled agents for intelligence-intensive operations. TCO modelling must accommodate this heterogeneity.
A complete agentic AI TCO model includes nine distinct cost layers. Most organisations model layers one through three. The remaining six account for the gap between projected and actual costs.
Before building anything, someone must define what the agent does, map its capabilities to actual business processes, assess data readiness, and design the system architecture. This work is often treated as “pre-project” and excluded from TCO calculations. It should not be, because the quality of discovery directly affects every downstream cost.
The cost range is wide because scope varies enormously. A well-bounded agent handling invoice classification against a clean data source is a $5,000–$15,000 discovery effort. An agent orchestrating across multiple legacy systems with inconsistent data models, compliance requirements, and integration points is $25,000–$50,000.
Discovery costs correlate inversely with downstream costs. Organisations that spend $5,000 on discovery for a complex deployment will spend $50,000 more in rework, re-architecture, and remediation than organisations that spend $30,000 getting the architecture right before building. This is the well-documented cost-of-change curve applied to a new domain, and it is routinely ignored in agentic AI budgeting because organisations underestimate architectural complexity.
A critical discovery decision in the current environment: build versus acquire. Discovery should explicitly evaluate whether the target use case is better served by a custom agent, a marketplace agent, a SaaS-embedded capability, or a managed service. This evaluation should include not just current cost comparison but flexibility assessment; how easily can the organisation pivot from one acquisition model to another as the market matures?
What good discovery produces: A cognitive task specification defining exactly what the agent does and does not do. A data readiness assessment identifying gaps, quality issues, and integration requirements. An architecture document specifying model selection, orchestration patterns, context management strategy, and governance integration points. An acquisition strategy recommending build, buy, or hybrid with explicit rationale and exit criteria. Without these artefacts, everything downstream is improvisation. (The Agentic AI Sprint Factory Sprint 0 phase delivers exactly these artefacts in two weeks.)
The core development cost. This includes proof of concept or MVP development, data preparation and engineering, integration with existing systems, and security and compliance certification.
Proof of Concept / MVP ($15,000–$60,000). A measurement-oriented build delivering auditable results in 6–12 weeks. The purpose is not to deploy a production system but to validate assumptions: Does the agent perform at expected accuracy? Does the data support the use case? Are integration points feasible? Do the unit economics hold?
The temptation to skip MVP and proceed directly to production build is strong and should be resisted. MVP builds cost 10–20% of full production builds but prevent 40–60% of capital waste by identifying fundamental feasibility issues before production investment. An MVP that reveals inadequate data quality, insurmountable integration complexity, or unacceptable accuracy saves the full production budget.
Data Preparation and Engineering ($10,000–$70,000). Builds the semantic infrastructure agents require: vector stores, knowledge graphs, feature pipelines, data quality validation, lineage tracking. The cost depends on the state of existing data. Organisations with mature data governance spend at the low end. Organisations whose data exists in inconsistent formats across siloed systems spend at the high end.
Integration and API Orchestration ($5,000–$20,000 per system). Connects the agent to the systems it needs to interact with: core banking platforms, CRM systems, document management, regulatory databases, market data feeds. Cost scales linearly with the number of integration points.
Security and Compliance Certification ($7,000–$50,000). Establishes data lineage, output explainability, and decision auditability required by regulatory frameworks: Basel III, GDPR, AML/KYC, SOX, and emerging AI-specific regulations. This is not optional in financial services and should not be modelled as such.
For cloud-hosted deployments, infrastructure costs are modest: $200–$2,000 monthly for compute, storage, networking, and any specialised instances. These costs are the most predictable component of cloud-based agentic AI TCO, and typically the smallest; which is why organisations that over-focus on infrastructure budgeting are optimising the wrong line item.
For on-premises deployments, infrastructure is the dominant cost layer and fundamentally changes the economic model. GPU clusters, storage, networking, power, cooling, facilities, and specialised operations staff transform infrastructure from a minor operating expense into a major capital investment. Organisations required to run local compute should model infrastructure as the primary cost driver and evaluate all other layers relative to it.
LLM API fees for the foundation models powering agent reasoning. This is the layer where agentic AI economics diverge most sharply from traditional software.
Model consumption costs are variable, driven by task complexity rather than transaction volume. A “happy path” invoice classification costs fractions of a cent in tokens. The same agent encountering a malformed document, ambiguous vendor name, or mismatched line items can consume an order of magnitude more tokens on a single exception. Exceptions are not rare in production environments; they are why agents exist in the first place.
A critical cost multiplier that most models miss: grounded search; retrieval-augmented generation using enterprise knowledge bases, regulatory databases, or document stores; currently runs approximately seven times the cost of standard model queries. An agent that performs five grounded retrievals per transaction has a fundamentally different consumption profile than one performing simple classification. Organisations modelling token costs based on simple query pricing will underestimate consumption by 3–5x for agents with significant retrieval requirements.
Model consumption costs are also dynamic. Foundation model providers adjust pricing regularly. New model tiers emerge with different cost-performance profiles. Input and output tokens carry different prices. Cached tokens cost less than fresh tokens. TCO models must account for this dynamism; static cost projections become obsolete within two quarters.
This is the cost layer that generates the worst surprises in production, and it has no meaningful parallel in traditional software economics. Agent failures do not produce clean error messages and graceful fallbacks. They produce runaway consumption, silent degradation, and cascading downstream damage.
Retry loops. These failures are a specific manifestation of the compound error problem, where error rates multiply across workflow steps rather than adding. An agent encountering an ambiguous input, a transient API failure, or an edge case outside its training may enter a retry loop, repeatedly attempting the same operation with the same failing approach. Unlike traditional software retries, each agent retry consumes substantial tokens (the full context must be reprocessed each time). A retry loop processing a complex document can burn through hundreds of dollars in tokens before anyone notices.
Runaway jobs. Agents operating without hard consumption ceilings can run indefinitely on difficult inputs. A compliance agent attempting to reconcile contradictory regulatory guidance might generate increasingly elaborate reasoning chains, each one longer and more expensive than the last, without converging on an answer. These jobs run until a human notices or a budget ceiling triggers; and in organisations without budget ceilings, they simply run.
Cartesian product explosions. Multi-agent systems and agents with access to multiple data sources can generate combinatorial query patterns; a query that should match three records instead matches three million because of a missing filter, a malformed join condition, or a misunderstood schema. The agent dutifully processes the results, consuming tokens proportional to the data volume returned. These failures can produce single-incident costs exceeding the agent's entire monthly budget.
Silent quality degradation. Perhaps the most expensive failure mode because its costs are indirect. An agent whose accuracy degrades from 95% to 85% over three months does not generate alerts if monitoring only tracks consumption and uptime. The cost appears downstream: incorrectly processed transactions requiring rework, compliance flags missed, customer interactions handled poorly.
Quantifying failure costs: This layer is the hardest to forecast because failure frequency and severity are highly variable. As a baseline, budget 5–15% of expected model consumption costs as a failure reserve in year one, declining to 3–8% in subsequent years as governance matures. For organisations operating without hard consumption ceilings, retry limits, and anomaly detection, budget significantly higher.
Traditional TCO models treat the human side of technology deployment as a one-time training cost. For agentic AI, human costs are ongoing, substantial, and structurally different from anything organisations have modelled before, because the technology is not a tool that people learn to use; it is a coworker whose capabilities, limitations, and behaviours change continuously.
Adoption and change management. Deploying an agent into an existing workflow is a change management challenge, not just a technology rollout. Staff who have performed a task for years must trust an AI system to handle part of their work, learn when and how to intervene, understand how to evaluate agent outputs, and adapt their own workflows around agent capabilities. Organisations that underinvest in adoption see low utilisation rates, higher exception escalation rates, and slower ramp-to-value. Budget $15,000–$50,000 for initial adoption programmes per agent deployment.
Continuous retraining of humans. Agents change more frequently than traditional software. Model updates alter agent behaviour. Retraining modifies capabilities. New features expand scope. Governance updates change escalation thresholds. Each change requires the humans working alongside the agent to update their understanding of what the agent can and cannot do. Budget $10,000–$40,000 annually for ongoing human retraining per agent system.
The digital assistant transition. Beyond individual agent deployments, organisations are moving toward digital assistant models; integrated AI interfaces that augment daily work across multiple functions. This transition requires broader workforce development: not just “how to use this agent” but “how to work effectively with AI assistants as a category.” Budget this centrally rather than per-agent, at $50,000–$200,000 annually for a mid-sized financial services organisation, declining over time as organisational AI maturity increases.
The cost of orchestrating human-agent-system interactions in operational contexts. This layer represents the ongoing overhead of making agents work within existing organisational structures.
Exception handling labour. Agents encounter situations outside their training or capability. These exceptions escalate to humans who must understand the agent's reasoning, assess the situation, make a decision, and in some cases retrain the agent to handle similar situations in the future. In early deployments, exception rates run 15–25% of total volume. Over time, well-governed agents reduce this to 5–10% as common exceptions are incorporated into training.
Supervision and quality assurance. Production agents require ongoing review of output quality. For compliance-adjacent agents in financial services, budget $2,000–$5,000 monthly for quality assurance labour.
Escalation protocol management. The rules governing when and how agents escalate to humans require ongoing calibration. Too aggressive escalation defeats the automation purpose. Too permissive escalation creates risk. Finding and maintaining the right threshold is an ongoing management task.
Quantifying coordination tax: As a heuristic, budget coordination tax at 15–30% of the gross efficiency gain the agent produces. If an agent saves $500,000 annually in labour costs, expect $75,000–$150,000 in coordination tax. The net efficiency gain is $350,000–$425,000; still excellent, but materially different from the $500,000 gross figure that appears in naive models.
Governance and security for agentic AI are not extensions of existing IT governance. They are substantially new capabilities that most organisations must build or acquire; and the platforms supporting them require significant improvement from their current state.
Agent-specific monitoring and observability ($5,000–$25,000/year). Traditional APM tools designed for deterministic software are insufficient for probabilistic AI systems. Agent monitoring requires understanding the governance gap; the structural vulnerability created when agents operate across systems that each enforce only their own policies. Agent monitoring requires tracking consumption patterns, output quality metrics, behavioural drift indicators, failure mode detection, and decision audit trails.
Security infrastructure ($10,000–$40,000/year). Agents introduce security surface area that traditional security tools do not cover. The Agentic AI Risk Catalog identifies 10 cybersecurity-specific risks, including prompt injection, supply chain compromise, and lateral movement, each with distinct economic exposure profiles. Current security platforms are adapting but not yet adequate; organisations should budget for security tooling that will need upgrading within the planning window as threats evolve and tools mature.
Continuous retraining of agents ($5,000–$30,000/year). Agents must be retrained as data patterns change, business rules evolve, and new edge cases emerge. Financial services environments change frequently: regulatory updates, new product offerings, market condition shifts, and customer behaviour evolution all create retraining requirements.
Incident response ($2,000–$15,000/year). When agents fail in production (and they will) someone must diagnose the failure, assess impact, implement remediation, and update governance protocols to prevent recurrence.
Audit and compliance reporting ($5,000–$15,000/year). Producing documentation for internal audit, regulatory examination, and board reporting. The cost of producing audit-ready documentation should be modelled explicitly and will increase as regulatory scrutiny intensifies.
Drift detection and remediation ($3,000–$10,000/year). Systematic comparison of current agent behaviour against baselines to detect behavioural drift; the gradual shift in agent outputs over time.
The critical planning assumption: governance and security platforms are immature today and will require significant investment to reach production-grade capability. Budget for platform improvements and migrations within the TCO window. Organisations that assume current tooling is sufficient will face unbudgeted platform costs within 12–18 months.
The hardest costs to quantify and therefore the most frequently omitted. This layer combines the traditional opportunity cost of internal resources with the increasingly critical cost of avoiding vendor and architectural lock-in.
Internal resource opportunity cost. Data engineering time, IT team time, business SME time, and management attention diverted to AI programmes from other priorities. Estimate internal FTE allocation (typically 0.5–2.0 FTE for a single agent system) and value at fully loaded cost: $75,000–$400,000 annually.
Lock-in avoidance investment. In an environment where the best model, the best framework, and the best platform may all be different twelve months from now, the ability to pivot is not a luxury; it is a strategic requirement. Lock-in avoidance has real costs: abstraction layers that enable model portability, architecture patterns that decouple business logic from specific providers, evaluation frameworks that allow rapid benchmarking of new alternatives, and the discipline to avoid commitment-based pricing that sacrifices flexibility for short-term savings.
Open-source leverage. Open-source models and frameworks are maturing rapidly. Organisations that invest in the capability to evaluate and deploy open-source alternatives; local inference infrastructure, model evaluation pipelines, fine-tuning capacity; gain strategic flexibility that proprietary-only architectures lack. This capability requires upfront investment ($20,000–$50,000 for initial infrastructure and skills development) but pays dividends when commercial model pricing shifts, when data sovereignty requirements emerge, or when the performance-cost frontier of open-source models crosses the threshold for specific use cases.
The TCO implication: organisations should model a “flexibility premium”; the additional cost of maintaining architectural portability and avoiding lock-in; and weigh it against the discount available from deeper vendor commitment. In most 2026 scenarios, the flexibility premium is the better investment. The time for deep optimisation will come when the market stabilises and the organisation has significant deployed footprint to optimise.
To illustrate how these layers combine, consider a mid-sized bank deploying an agentic AI system to automate customer due diligence for KYC onboarding.
Context: The bank's KYC team processes 50,000 new customer applications annually. Current cost per application: $120 (blended labour, systems, and overhead). Total annual cost: $6 million. Average processing time: 5 business days. Error rate requiring rework: 12%.
Acquisition decision: After discovery, the bank evaluates marketplace KYC agents (several now available as SaaS offerings) against a custom build. The marketplace options handle 60% of standard applications but lack integration with the bank's proprietary risk models and internal sanctions screening. Decision: custom build for the core agent with marketplace integration for entity resolution and document extraction components.
Expected agent performance: Handle 70% of applications autonomously (standard risk profiles with complete documentation). Reduce cost per automated application to $35. Reduce processing time to same-day for automated applications. Reduce error rate to 4%.
| Cost Layer | Item | Estimate |
|---|---|---|
| Layer 1 | Discovery and architecture (including acquisition assessment) | $35,000 |
| Layer 2 | MVP build | $45,000 |
| Layer 2 | Production build (post-MVP validation) | $120,000 |
| Layer 2 | Data preparation | $40,000 |
| Layer 2 | Integration (3 systems × $15K) | $45,000 |
| Layer 2 | Security and compliance certification | $35,000 |
| Layer 3 | Infrastructure (12 months) | $18,000 |
| Layer 4 | Model consumption (12 months, incl. grounded search premium) | $48,000 |
| Layer 5 | Failure reserve (10% of consumption + incident costs) | $15,000 |
| Layer 6 | Human costs; adoption programme + ongoing retraining | $55,000 |
| Layer 7 | Coordination tax (12 months) | $45,000 |
| Layer 8 | Governance and security infrastructure (12 months) | $55,000 |
| Layer 9 | Opportunity cost (1.5 FTE) + lock-in avoidance | $245,000 |
| Year 1 Total | $866,000 | |
| Cost Layer | Item | Estimate |
|---|---|---|
| Layer 2 | Maintenance (20% of build cost) | $57,000 |
| Layer 3 | Infrastructure | $18,000 |
| Layer 4 | Model consumption (assumes one model transition) | $42,000 |
| Layer 4 | Model transition costs (testing, validation) | $20,000 |
| Layer 5 | Failure reserve (reduced as governance matures) | $10,000 |
| Layer 6 | Human costs; ongoing retraining + digital assistant transition (share) | $35,000 |
| Layer 7 | Coordination tax (reduced as agent matures) | $37,000 |
| Layer 8 | Governance and security infrastructure | $57,000 |
| Layer 9 | Opportunity cost (reduced to 0.75 FTE) | $120,000 |
| Year 2 Total | $396,000 | |
| Cost Layer | Item | Estimate |
|---|---|---|
| Layer 2 | Maintenance | $52,000 |
| Layer 3 | Infrastructure | $18,000 |
| Layer 4 | Model consumption (optimised routing, possible open-source components) | $32,000 |
| Layer 5 | Failure reserve (steady state) | $8,000 |
| Layer 6 | Human costs; ongoing retraining | $25,000 |
| Layer 7 | Coordination tax (steady state) | $32,000 |
| Layer 8 | Governance and security infrastructure | $50,000 |
| Layer 9 | Opportunity cost (reduced to 0.5 FTE) | $80,000 |
| Year 3 Total | $297,000 | |
Direct savings: 35,000 automated applications × ($120 − $35) cost reduction = $2,975,000 annual savings at steady state. Three-year cumulative (ramping from 50% automation in year 1 to 70% in year 3): approximately $7.4 million.
Indirect value: Reduced processing time improves customer experience and conversion. At 5% improvement in application completion rate (fewer abandons due to slow processing), the bank gains 2,500 additional customers annually. At average first-year customer value of $400, this represents $1 million in annual revenue enhancement.
Error reduction value: Reducing rework from 12% to 4% on 50,000 applications saves approximately 4,000 rework cycles annually. At $85 per rework cycle, this saves $340,000 annually.
Three-year cumulative value: approximately $11.5 million.
Three-year ROI: 7.4:1 ($11.5M value / $1.559M cost)
A typical three-layer TCO model (build + infrastructure + model consumption) would have estimated:
The naive model understates true TCO by 63%. Reported ROI would have been 20:1; impressive on paper, but indefensible when actual costs surface. The fully-loaded 7.4:1 ratio is still excellent and, critically, it is defensible. It also accounts for the flexibility investment that positions the bank to capitalise on market improvements in years two and three.
For organisations requiring five-year projections, model years four and five with wider ranges and explicit assumptions about market maturation:
| Assumption | Year 4–5 Estimate | Rationale |
|---|---|---|
| Model consumption | $20,000–$35,000/year | Open-source maturation, continued pricing decline |
| Governance platforms | $35,000–$50,000/year | Commercial platform maturation reduces custom investment |
| Human costs | $15,000–$25,000/year | Organisational AI maturity reduces retraining burden |
| Coordination tax | $25,000–$35,000/year | Steady state with incremental improvement |
| Maintenance + infrastructure | $60,000–$75,000/year | Stable with COLA |
| Annual total (range) | $155,000–$220,000 |
Five-year cumulative TCO range: $1,870,000–$2,000,000. Five-year cumulative value (at steady-state rates with COLA): approximately $20 million. Five-year ROI range: 10:1–10.7:1.
Pilot deployments operate under favourable conditions: curated data, limited scope, dedicated attention from the strongest technical staff, simple integration requirements. Production deployments operate under real conditions: messy data, full scope, shared attention, complex integration. Costs per transaction in production routinely run 2–3x pilot costs. The correction is simple: never extrapolate pilot unit economics to production without a complexity adjustment factor. A factor of 2.0–2.5x is appropriate for most financial services deployments.
Agents do not perform at steady-state capability on day one. There is a ramp period; typically 3–6 months for moderately complex agents; during which the agent operates at reduced accuracy, higher exception rates, and increased coordination tax. TCO models must account for the ramp period with reduced value capture assumptions. Modelling full value capture from month one overstates year-one returns and creates unrealistic expectations.
TCO models often assume that headcount reductions are immediate and permanent. In practice, organisations rarely reduce headcount proportionally to automation gains. Staff are redeployed rather than eliminated. Unless the organisation has committed to specific headcount actions, model the value as throughput capacity gain rather than headcount savings. The financial impact may be equivalent, but only if the organisation has demand to absorb the additional capacity.
Foundation models improve rapidly. The model powering today's agent will likely be superseded within 12–18 months by a model that is either cheaper, better, or both. Transitioning to a new model is not free: it requires testing, validation, potential retraining of dependent components, and revalidation of output quality. Budget $10,000–$30,000 per major model transition, with one to two transitions expected per three-year period.
Organisations that price model consumption based on standard query costs and then deploy agents with grounded search discover that retrieval-augmented queries cost approximately seven times standard queries. A TCO model that assumes standard pricing for an agent performing five grounded retrievals per transaction underestimates consumption costs by 3–5x. Identify which agents require grounded search during discovery and price consumption accordingly.
“Our agents won't fail” is not a planning assumption. It is wishful thinking. Every production agent system will experience retry loops, runaway jobs, quality degradation events, and consumption anomalies. The question is not whether failures occur but whether the TCO model accounts for their cost and whether governance infrastructure limits their impact. Organisations operating without hard consumption ceilings, retry limits, and anomaly detection are accepting unlimited failure cost exposure; the equivalent of operating without insurance.
Every line item in an agentic AI TCO model carries genuine uncertainty. Acknowledging this uncertainty; presenting low, expected, and high estimates for each cost layer; produces a model that is more credible and more useful than false precision. Present the TCO as a range: best case (favourable assumptions across all layers), expected case (most likely outcome for each layer), and stress case (unfavourable assumptions across all layers). For the KYC example above, the three-year range spans approximately $1,200,000 to $2,000,000.
Not all costs in the TCO model represent cash expenditure. Opportunity costs and redeployed labour are real economic costs but are not typically reflected in P&L impact the same way as vendor invoices and cloud bills. Presenting the model with a clear separation between cash costs (what leaves the bank account) and economic costs (total resource value consumed, including redeployed internal effort) allows stakeholders to evaluate both perspectives.
Governance is not an optional add-on. It is a structural requirement for sustainable returns. Models that present governance as a “phase 2” consideration or a “nice to have” misrepresent the economics. Governed and ungoverned deployments have fundamentally different return profiles over time. Ungoverned deployments that appear cheaper in year one become dramatically more expensive by year three as degradation, drift, compliance exposure, and uncontrolled failure costs accumulate.
In the current environment, the most expensive mistake is not overspending on the current architecture; it is locking into an architecture that cannot adapt as the market evolves. Structure TCO models to value flexibility explicitly: what does model portability cost? What does maintaining open-source optionality cost? What does avoiding long-term vendor commitments cost? These are real costs, and they are almost always worth paying until the market stabilises and the organisation has sufficient deployed footprint to optimise against known patterns.
All labour-dependent costs; coordination tax, human retraining, governance staffing, exception handling, quality assurance; should include annual cost-of-living adjustments of 3–5%. This is standard practice in enterprise and public sector TCO models and should not be omitted for AI deployments simply because the technology costs may be declining. People costs and technology costs move in opposite directions; the model must capture both trends.
Agentic AI economics are dynamic. Model pricing changes, agent performance improves (or degrades), coordination tax evolves, governance requirements shift, and new acquisition options emerge. A TCO model built at project inception and never updated is unreliable by the end of year one. Establish a quarterly cadence for updating actuals against projections, revising forward estimates, and adjusting investment decisions based on observed rather than projected economics.
TCO models serve two purposes: they inform investment decisions and they establish credibility with stakeholders. The second purpose matters as much as the first.
When presenting agentic AI TCO to executive stakeholders, lead with the fully-loaded model. Organisations that present the naive model first and then reveal additional costs later; or discover them in post-deployment analysis; lose credibility not just for the current initiative but for subsequent AI investments. Presenting a complete, honest TCO model that still shows compelling returns builds the institutional confidence necessary for sustained AI investment at scale.
Four principles for the stakeholder conversation:
Show the maths. For every cost estimate, show the reasoning. “Model consumption is estimated at $48,000 annually based on 50,000 applications, 70% automation rate, average token consumption of X per application at current model pricing of Y per token, with a 7x multiplier for grounded search queries and Z% exception surcharge.” This level of detail survives challenge and demonstrates analytical rigour.
Acknowledge uncertainty explicitly. “Our three-year TCO estimate ranges from $1.2M to $2.0M. The primary sources of uncertainty are model consumption costs, which depend on exception rates and grounded search volume we will only observe in production, failure costs, and coordination tax. We will narrow these ranges after the MVP phase.”
Connect costs to governance. Every governance cost in the model should be connected to a specific risk it mitigates. “The $55,000 annual governance and security infrastructure cost includes monitoring that detects accuracy degradation within 48 hours, consumption ceilings that prevent runaway jobs from exceeding $500 per incident, and retraining schedules that maintain performance as regulatory requirements change. Organisations without this infrastructure experience 30–50% ROI degradation over a three-year period.”
Frame flexibility as investment, not waste. “We have budgeted $20,000 in year one for open-source evaluation capability and architectural portability. This allows us to transition to a model that is 40% cheaper if one emerges in year two; which, given the pace of development, is probable rather than possible. The alternative is committing to a three-year vendor contract that saves 15% now but forecloses adaptation.”
The TCO framework in this article is designed to be applied, not just read. Three services connect directly to the cost layers and modelling challenges described above:
“We need to validate the economics before committing to a full build.” The Agentic AI Sprint Factory is explicitly designed as a TCO de-risking mechanism. Sprint 0 produces the discovery artefacts (Layer 1), the MVP validates unit economics (Layer 2), and the DMAIC Baseline Report and Coordination Tax Impact Assessment provide the measured inputs for Layers 5–7; replacing estimates with observed data before the organisation commits to production investment.
“We need help building the governance infrastructure that the TCO model says we can't skip.” The AI Governance Framework Design engagement produces the policy suite, risk committee charter, materiality assessment methodology, and lifecycle control framework that constitute Layer 8. For organisations operating in Singapore's regulatory environment, the MAS AIRG Readiness Assessment maps current state against requirements and produces a prioritised compliance roadmap; connecting governance investment directly to regulatory exposure.
“We need ongoing advisory to keep the model current and the governance operational.” The Fractional AI Governance Advisor provides the quarterly model updates, regulatory readiness support, and continuous governance review that this article argues are essential for keeping TCO projections aligned with reality.
This article is part of a seven-article series on the economics of agentic AI in financial services.
Start with measurement, not deployment. The Sprint Factory delivers auditable cost baselines and governance infrastructure in 6 to 12 weeks, so your board invests with confidence in the full picture.
Schedule a Briefing Explore the Sprint Factory