Total cost of ownership of an AI project

The total cost of ownership of an AI project includes diagnosis, data preparation, development, integration, model consumption, infrastructure, security, governance, adoption, operation, and retirement. Budgeting only the pilot distorts the ROI: the financial decision must model at least three use cases and the entire system lifecycle.

Taxonomic definitions

1. Total cost of ownership of AI — AI TCO

The total cost of ownership of artificial intelligence represents all the financial resources necessary to design, implement, operate, control, scale, and retire an AI system. AI solution for a certain period of time.

It does not equal the cost of hiring a license, developing a model, or running a proof of concept. It includes:

Initial investment.
Recurring costs.
Costs of integration.
Variable consumption.
Human operation.
Security and compliance.
Managing change.
Contingencies.
Migration or withdrawal of the solution.

2. FinOps for AI

It is the discipline that connects technical decisions about models, infrastructure, and consumption with financial responsibility. In AI, spending can depend on input and output tokens, API requests, reserved capacity, processing time, storage, tool calls, and the volume of inferences.

Its goal is not to reduce costs indiscriminately, but to optimize the relationship between:

Cost.
quality of response.
speed.
risk.
result of business.

3. MLOps and LLMOps

MLOps It groups the necessary practices for deploying, monitoring, and maintaining machine learning models. LLMOps Extend these practices to language model-based solutions, including prompts, vector databases, knowledge retrieval, response evaluation, token consumption, security, traceability, and version control.

Both disciplines turn a prototype into a commercially viable business service.

Comparative table: pilot budget vs. business TCO

Variable	Budget limited to the pilot	Budget based on TCO	Financial risk of omitting it
Horizon	4 to 12 weeks	24 to 36 months with periodic check-ups	Approving an initiative that loses profitability as it scales up
Data	Controlled display	Cleaning, integration, permissions, updates, and continuous quality	Later costs of correction and incorrect answers
Models	An API or model	Routing, alternative models, contingency and migration	Supplier dependency and increased unit cost
Infrastructure	Test environment	Production, storage, networks, backups and high availability	Interruptions and unplanned expenses
Integrations	A source of information	ERP, CRM, ITSM, channels, identity and transactional systems	Manual labor that eliminates the expected savings
Security	Basic controls	Identity, encryption, filtering, evidence, records, and incident response	Information and sanctions exhibition
Operation	Project team	Support, observability, evaluation, updating, and incident management	Silent degradation of the service
Adoption	Initial training	Process redesign, responsible parties, communication, and usage measurement	Low utilization and return below the expected level
Scalability	Controlled users	Real volume, competition, peaks and expansion by areas	Non-linear growth in spending
Retiro	Not contemplated	Data export, replacement, filing and closing of contracts	Technological lock and exit costs
Main metric	Technical precision	Cost per result, EBITDA, ROI, payback and residual risk	Decisions focused on technology, not on value

Why the model price does not represent the cost of the project

Stanford's AI Index 2025 reported that the proportion of organizations that used AI in at least one function increased from 55% in 2023 to 78% in 2024; the use of generative AI increased from 33% to 71%. However, only 1% of the executives surveyed described their generative AI deployments as mature.

The difference between adopting AI and operating it profitably usually lies in the costs that appear after the pilot:

Business integrations.
Continued preparation of information.
Security checks.
Human supervision.
Evaluation of results.
Consumption management.
Support and updates.
Government and compliance.

The cost of inferring a system with performance comparable to GPT-3.5 fell by more than 280 times between November 2022 and October 2024. In parallel, the cost of hardware decreased by approximately 30% annually and its energy efficiency increased by 40% annually.

This does not mean that the business TCO decreases in the same proportion. As the unit price of the model decreases, companies expand their users, processes, data sources, agents, and automations. The spending shifts towards integration, operation, evaluation, governance, and scale.

How to calculate the TCO of an AI project

1. Define the economic outcome before choosing the technology

The budget must start with an operational baseline. Without it, the planned savings become difficult to audit.

Minimum information required

Monthly volume of operations.
Average time per activity.
Direct labor cost.
Cost of errors, rework and escalations.
Income affected by the process.
Current level of service.
Seasonal patterns and peak demand.
Number of users.
Systems that must be integrated.
Regulatory and contractual risks.
Cost of maintaining the process unchanged.

Basic financial formulas

TCO for the period. TCO = initial investment + recurring costs + expansion costs + risk and compliance costs + exit costs
Accumulated ROI. ROI = ((accumulated economic benefit − TCO) / TCO) × 100
Recovery period. Payback = initial investment / monthly net profit
Annual impact on EBITDA. EBITDA impact = recurring operating savings + incremental margin − IA OPEX − incremental human cost

Benefits must be calculated based on verifiable results, not on “potentially freed” hours. An hour saved only generates financial impact when:

Reduce overtime hours.
Avoid hiring new employees.
Increases production capacity.
Reduce losses or errors.
It accelerates revenue.
It allows resources to be reallocated to measurable activities.

2. Budgeting the entire cycle of the solution

2.1 Diagnosis and business case

Before building it, it must be determined whether the process justifies AI, traditional automation, or a combination of both.

The budget must cover:

Mapping of the current process.
Identification of decisions and exceptions.
Baseline of costs and performance.
Data feasibility analysis.
Risk assessment.
Architecture design.
Definition of indicators.
Prioritization of use cases.
Financial model by scenarios.

This work avoids using an AI model to solve problems that could be automated with rules that are less costly and risky.

2.2 Data preparation and governance

Data often generates costs before and after deployment. A once-loaded document base does not equate to a maintainable business knowledge system.

They must include:

Inventory of sources.
Extraction and transformation.
Normalization and deduplication.
Classification of information.
Access management.
Elimination or masking of sensitive data.
Labeled.
Quality validation.
Synchronization.
Change history.
Retention and disposal.
Those responsible for the control of information.

The cost of correcting the data source must also be budgeted. Cleaning up faulty information within the AI application can work in the pilot, but it creates operational debt as the volume increases.

2.3 Architecture, models and inference

The cost of the model depends on the selected architecture. Not all cases require the largest model or the same quality for each interaction.

Variables that must be modeled

Number of transactions.
Average entry tokens.
Average output tokens.
Length of the context.
Number of documentary recoveries.
Use of embeddings.
Calls to external tools.
Reattempts.
Caché.
Competition.
Required latency.
Backup models.
Development, testing and production environments.
Reserved capacity or consumption on demand.

Formula for monthly consumption

Monthly cost of AI = model consumption + embeddings + storage + search + tool calls + infrastructure + observability + retries

The most useful financial unit is not “monthly platform cost”, but:

Cost per resolved conversation.
Cost per processed document.
Cost per ticket avoided.
Cost per accepted prediction.
Cost per assisted sale.
Cost per completed transaction.
Cost per productive hour recovered.

The FinOps Foundation points out that AI management must cover the entire cycle: training or tuning, inference, orchestration, tools, and operation. It also recommends linking spending with signals that demonstrate effective use and business value.

2.4 Integration with business processes

A correct response that cannot execute an action has a limited value. Business projects often require integration with:

ERP.
CRM.
ITSM.
Trading platforms.
Identity systems.
Databases.
Internal portals.
Mail and collaboration.
Channels of attention.
Payment systems.
Legacy applications.

The budget should consider:

APIs and connectors.
Middleware.
Customized development.
End-to-end testing.
Credential management.
Permit control.
Error handling.
Immutability.
Reconciliation of transactions.
Continuity in the face of failures.
Future changes in connected systems.

Integration often becomes a recurring cost: every modification of a source system may require adjustments, testing, and documentation.

2.5 Security, governance and compliance

The AI government should not be treated as a legal review at the end of the project. NIST structures the management of AI risk around four functions: governing, mapping, measuring and managing. Its guide also includes data quality, validation, testing, supervision, and legal reviews throughout the life cycle.

The budget must include:

Classification of the level of risk.
Privacy assessment.
Access control.
Encryption.
Records and traceability.
Security tests.
Protection against prompt injection.
Content filters.
Validation of sources.
Assessment of bias.
Human supervision.
Incident management.
Technical documentation.
Legal reviews.
Audits.
Third-party management.

For companies with operations or users in the European Union, the financial risk can be substantial. The AI Act’s sanctioning regime provides for fines of up to 35 million euros for certain prohibited practices or 7% of the annual global turnover, subject to the application and circumstances provided for in the regulation.

The cost of compliance must be compared with:

Potential fines.
Litigations.
Suspension of the system.
Loss of contracts.
Compensation.
Remedy.
Reputational damage.
Increase in insurance premiums.

2.6 Management of change and adoption

A project that works technically, but does not change the process, maintains two costs: the new system and the previous operation.

The budget must include:

Function redesign.
Training by profile.
Operating manuals.
Internal communication.
Management of objections.
New controls.
Escalation schemes.
Adoption metrics.
Those responsible for the process.
Support during the transition.

Relevant metrics include:

Active users versus enabled users.
Frequency of use.
Percentage of accepted responses.
Percentage of processes completed.
Human intervention.
Proven time savings.
Mistakes after adoption.
Return by area.

3. Include the costs of continuous operation

An AI solution is not finished when it is published. It must be monitored because things can change:

The data.
The process.
The behavior of the users.
The provider's models.
The integrations.
The legal requirements.
The pattern of consumption.
The quality of the responses.

Recurring costs that must be budgeted for

Consumption of models.
Infrastructure.
Licenses.
Storage.
Observability.
Support.
Automated evaluation.
Human review.
Update of knowledge.
Prompt adjustment.
Re-training.
Correction of integrations.
Security tests.
Incident management.
Audits.
Administration of suppliers.

Operational reserve

The budget must include a reserve for:

Volume increases.
Changes in prices.
Replacement of models.
New regulatory obligations.
Security incidents.
Data of lower quality than expected.
Additional integrations.
Greater human oversight.

The reserve should not be presented as a free margin. It should be related to defined risks, responsible parties, and release criteria.

4. Incorporate the cost of scaling up

The pilot usually operates with selected users, controlled data, and direct support from the responsible team. Production introduces different conditions:

Peaks of demand.
Longer consultations.
Users with unexpected behaviors.
Inconsistent documents.
More languages.
Operational exceptions.
Reattempts.
Simultaneous integrations.
Need for continuous availability.

The cost can increase non-linearly when a response activates several agents, queries, validations, and tools. The FinOps Foundation warns that generative AI workloads can present unpredictable scaling dynamics, where the cost of inference multiplies as usage grows.

Therefore, the budget must calculate three scenarios:

Set	Volume assumption	Allegation of adoption	Financial use
Conservator	Lower than expected	Slow adoption	It assesses the risk of not achieving profits
Base	Planned volume	Planned adoption	It supports the operational budget
Expansion	Accelerated growth	New areas and processes	It anticipates capacity and pressure on margins

Each scenario must show:

Accumulated TCO.
Unit cost.
Accumulated profit.
ROI.
Revenge.
Cash flow.
Impact on EBITDA.
Residual risk.

5. Use a complete budget distribution

The following distribution is a Scanda-oriented template for a 36-month TCO. It does not constitute a tariff or a universal benchmark; it must be adjusted to the use case, architecture, regulation, and data maturity.

Component	Informative participation of the TCO	What it covers
Diagnosis and business case	7%	Process, baseline, feasibility, architecture, and metrics
Data and knowledge	18%	Preparation, quality, permits, integration and updating
Development and integrations	25%	Applications, flows, APIs, testing and automation
Models and infrastructure	12%	Inference, storage, search and environments
Security, government and compliance	10%	Controls, validation, documentation and audit
Adoption and operational redesign	8%	Training, communication, processes, and transition support
Operation and observability	15%	Support, evaluation, maintenance and optimization
Contingency, migration and retirement	5%	Risks, supplier substitution and closure
Total	100%	Complete cycle of the system

An organization with fragmented data may require a higher proportion for preparation and integration. A regulated system can focus more budget on validation, documentation, human oversight, and compliance.

6. Controlling the budget with unit economics

The investment committee should not only review accumulated spending. It should assess how much it costs to produce an outcome that the business accepts.

Cost indicators

Cost per inference.
Cost per active user.
Cost per completed process.
Cost per accepted response.
Cost per successful automation.
Cost of human review.
Budget variation versus actual spending.
Consumption by area.
Consumption by model.
Consumption by provider.

Indicators of value

Productive hours recovered.
Reduction of errors.
Capacity increase.
Reduction of cycle time.
Decrease in escalations.
Incremental income.
Avoided losses.
Compliance with service levels.
Incremental margin.
Impact on EBITDA.

Stanford reported that the areas where organizations most frequently reported cost reductions from AI were service operations, with 49%; supply chain and inventory management, with 43%; and software engineering, with 41%. In revenue generation, marketing and sales stood out, with 71%; supply chain, with 63%; and service operations, with 57%.

These percentages indicate where impact is being reported, but they do not replace the business case of each company. The return must be validated against the internal baseline.

7. Identify hidden costs

The costs most likely to fall outside the initial budget are:

Human supervision

Validation of responses.
Approvals.
Handling exceptions.
Correction of results.
Escalation to specialists.

Low quality

Reattempts.
Retraining.
Inutil answers.
Incorrect decisions.
Loss of user confidence.

Technological department

Migration of models.
API changes.
Data export.
Replacement of components.
Contract renegotiation.

Growth in the context

Adding documents, history, and sources can increase token consumption, latency, and transaction costs.

Agents and automation

An agent can execute multiple steps before producing a result:

Consult information.
Call for tools.
Validate data.
Requesting approvals.
Correcting mistakes.
Retry operations.

Each step adds consumption, infrastructure, observability, and risk.

Withdrawal of the system

The end of the cycle requires:

Exporting information.
Revoke credentials.
Close integrations.
Keep evidence.
Delete data.
Replacing processes.
Finalize contracts.

8. Establish financial approval criteria

A project should not proceed to production without documentation:

TCO at 24 or 36 months.
Conservative scenarios, base and expansion.
Main economic unit.
Owner of the benefit.
Owner of the cost.
Operating budget.
Quantified risks.
Controls required.
ROI threshold.
Maximum recovery period.
Suspension conditions.
Exit strategy.

Warning signs

The benefit depends solely on “productivity” without any financial conversion.
There is no baseline.
The budget ends at the end of the pilot.
The data is considered free of charge.
The cost of human review has not been calculated.
There is no person responsible for the operation.
Consumption cannot be attributed by area.
No usage limits have been defined.
There is no security budget.
The replacement of the provider has not been evaluated.

Predictive conclusion

The unit cost of models will continue to decline, and smaller models will solve an increasing proportion of tasks. However, the TCO will not disappear: it will shift towards data, integration, agents, security, observability, governance, and operation. This prediction stems from the accelerated reduction in the cost of inference and the expansion of the financial scope that FinOps already attributes to AI workloads.

Companies with better results will not necessarily be the ones that contract the model with the lowest price, but those that:

Assign costs per process.
Midan accepted the result.
Use models according to complexity.
Control consumption from the design stage.
Incorporate government from the beginning.
Update the business case with real data.
They can replace components without having to rebuild the entire solution.

The next budgetary pressure will come from AI systems capable of executing actions. A chatbot generates responses; an agent can generate responses, query systems, make decisions, and trigger transactions. That increased autonomy requires more controls and raises the potential cost of every error.

FAQ: Total cost of ownership of AI projects

1. What horizon should a Financial Director use to calculate the TCO of AI? For an entrepreneurial initiative, Scanda recommends modeling a base horizon of 36 months and a detailed view of the first 12 months. The calculation should be updated quarterly with consumption, adoption, unit costs, and actual profits. Contracts, depreciation, and treatment of CAPEX or OPEX must align with the company’s accounting policy.

2. How can a CIO prevent AI consumption from exceeding the budget? It must establish limits per user, process, area, and environment; measure tokens, calls, retries, and tools; apply caching; use different models based on complexity; and create deviation alerts. The central indicator should be the cost per accepted result, not just the cost per token.

3. How to decide between developing, purchasing, or integrating an AI solution? The comparison must consider TCO, time to generate value, internal capacity, security, customization, vendor dependency, and exit costs. Buying can reduce initial time; developing can offer greater control; integrating components often balances speed and flexibility. The decision should be based on the process and not on a technological preference.

Budget the entire operation, not just the pilot

MaIA by Scanda helps organizations turn an artificial intelligence initiative into a controllable business case: from readiness and data assessment to architecture, automation, governance, adoption, and financial measurement. The approach is not about deploying isolated bots, but about organizing an AI and automation strategy connected to critical processes, efficiency, productivity, and growth capacity.

Request an AI Readiness and TCO assessment with Scanda. Identify the visible and hidden costs of your initiative, build investment scenarios, and determine which use cases can generate returns without transferring technical risks to EBITDA.

Total cost of ownership of an AI project: what to include in the budget