NEODY Super Series · Chapter 01

How much does AI cost?

The subscription starts the conversation. The real budget emerges when models, data, integration and supervision need to work together.

Start reading
Executive summary

The real cost combines implementation, subscriptions or usage, infrastructure, integration, supervision and maintenance. Compare proposals by dividing that total by tasks completed to the required standard. A NEODY simulation shows how R$6,800 in monthly operating expenses can lead to R$92,000 in first-year spending, with two implementation months and ten operating months. These are teaching assumptions, not a market quote.

One company wants to automate order processing. Another wants to answer customers. A third plans to use agents to prepare sales proposals. All three might choose the same artificial intelligence model and end up with very different budgets.

The difference lies in the work required to turn a response into a reliable operation. Reading a message is one step. Identifying the customer, checking inventory, applying the correct commercial terms and recording the order require data, rules and integration.

The most useful unit of comparison is the cost of a task completed to the required standard. This is the opening thesis of the NEODY Super Series Where Brazil can gain from AI: investigate what it costs to operate and which outcomes justify that cost.

How much does it cost to implement AI in a company?

There is no single price that represents AI projects. The budget depends on the process, volume, data quality, systems involved and degree of autonomy. An individual subscription and an agent integrated with an ERP system serve different needs.

NEODY proposes separating three accounts: the investment to get started, recurring operating expenses and the effort needed to maintain quality. This is an editorial assessment method, rather than a supplier price survey.

Layer Costs included What the proposal should specify
Implementation Assessment, configuration, data, integration and testing Deliverables, responsibilities and acceptance criteria
Operation Licences, usage, infrastructure and support Included volume, overages and limits
Quality Supervision, correction, evaluation and maintenance Who reviews, how much and who fixes issues
Continuity Export, replacement and termination Data access and an exit plan

IBM’s definition of total cost of ownership considers expenses throughout a product or service’s lifecycle. Applied to a business decision, it requires looking beyond the initial payment.

A manager should request two figures: expected spending over the first 12 months and cost per accepted unit of work. The first reveals the budget requirement. The second helps determine whether the project improves the economics of the process.

Subscription, usage or reserved capacity?

The charging model changes who bears the risk of volume. Per-user subscriptions follow the number of authorised users and the plan’s terms. Usage-based APIs follow consumption. With reserved capacity, some expenses may remain even when utilisation falls.

Pricing pages from AWS and Google Cloud show that generative services can distinguish input, output and processing options. Rates depend on the model, region, configuration and commercial terms. This analysis does not present those tables as final quotes for Brazilian companies.

Tokens are processing units used for text and other inputs. A request may contain instructions, conversation history and documents as well as the visible question. Technical cost depends on what was actually processed.

An agent may also need several calls to finish one task. It retrieves information, interprets a system response, performs an action and checks the outcome. Repeated steps and corrections can increase spending without increasing useful deliveries.

Asking only for the price per call is therefore insufficient. Estimate how many calls, lookups and attempts make up an accepted delivery. The proposal should also explain how the customer monitors usage and stops an execution that exceeds its limit.

For a budget in Brazilian reais, record the billing currency, exchange-rate assumption, payment costs and charges applicable to the specific agreement. Currency conversion alone does not necessarily represent the final invoice.

Integration can change the budget

Consider a sales agent checking availability. If inventory already has structured access, up-to-date documentation and consistent rules, some work can be reused. If every branch records products differently, integration begins with organising the process.

This effort does not disappear when the model improves. A system can interpret a description competently while receiving an outdated price. Its response will be fluent, but the operation will still be wrong.

When comparing suppliers, investigate where discount rules, customer records, credit limits and document versions reside. Ask how changes reach the agent and how it confirms that it is consulting the current version.

There is also internal work. Sales must explain exceptions; operations must validate the workflow; technology must authorise access; and someone must accept test results. Even when these hours do not appear on an external invoice, they consume company capacity.

A reasonable budget makes these dependencies explicit. “Integration included” needs a list of covered systems, required access, exceptions and maintenance responsibilities. A demo integration may fail when it receives incomplete data or a different production configuration.

A simulation: R$6,800 per month does not describe the whole first year

The following figures are fictional teaching assumptions. They do not represent an interviewed company, market price research or a commercial offer. Their purpose is to show how to build the calculation with transparent assumptions.

Imagine a document process receiving 3,000 requests per month. The company budgets R$24,000 for implementation, followed by these operating expenses:

Recurring item Monthly assumption
Licences and model usage R$1,200
Infrastructure, lookups and records R$800
Support and maintenance R$1,500
Human review: 40 hours at R$60 R$2,400
Corrections and evaluations: 15 hours at R$60 R$900
Recurring total R$6,800

Assume two months of implementation and ten months of operation in the first year. Spending for that period is R$24,000 + 10 × R$6,800 = R$92,000. With R$24,000 for implementation and nine billed operating months, it is R$85,200. The calendar changes the result; proposals must cover the same period.

A full year of stabilised operation costs 12 × R$6,800 = R$81,600, without charging an already-paid implementation again. Implementation plus 12 operating months costs R$105,600. Each number answers a different question.

Now assume only 2,400 of the 3,000 requests are completed within the deadline and quality standard. Recurring cost per accepted delivery is R$6,800 ÷ 2,400 = R$2.83. Dividing by all 3,000 inputs would produce R$2.27 and obscure part of the operational failure.

To compare alternatives over 12 months, a manager may allocate R$2,000 monthly for implementation. Average cost becomes R$8,800 ÷ 2,400 = R$3.67. This management allocation does not replace accounting treatment or change when cash is paid.

Choosing this denominator follows the FinOps Foundation’s guidance to connect resource costs to business outcomes rather than limiting analysis to technical consumption.

Break-even depends on quality

Continue the simulation. Assume a contracted alternative costs R$5 per accepted document and can meet the same standard. For 2,400 documents, avoidable spending would be R$12,000. Compared with recurring AI costs of R$6,800, the difference is R$5,200 monthly.

In this simplified scenario, recovering R$24,000 in implementation requires R$24,000 ÷ R$5,200, or approximately 4.6 months of stabilised operation. Implementation and ramp-up would extend payback measured from project inception. The calculation assumes the previous expense can genuinely be eliminated without additional termination costs.

If only 1,200 documents are accepted, the alternative costs R$6,000 while AI operation remains at R$6,800, assuming unchanged fixed expenses. The monthly gain disappears. A tool that looks inexpensive per use can become expensive when its completion rate is low.

Employee hours saved cannot automatically be treated as money saved. Payroll may remain unchanged. The company gains available capacity, whose value depends on reallocation, additional output or expenses actually avoided.

AI pays for itself when the company captures enough value to cover implementation and operation. A good demonstration does not yet answer that calculation.

For more on this distinction, see NEODY’s analysis of AI payback (Portuguese).

Human supervision needs a volume and an owner

“Human review” is an incomplete budget line without frequency, duration and criteria. Reviewing every document, checking a sample or intervening only in exceptions creates different costs and risks.

A team may save drafting time and spend more on checking. It may also improve quality without reducing duration. These effects must remain separate so that productivity and quality are not confused.

Generative AI at Work, published by NBER, found productivity gains in a specific customer-support setting. It provides evidence for that application without guaranteeing the same effect across processes.

A METR experiment involving 16 developers and 246 tasks, using early-2025 tools on projects familiar to participants, found a 19% increase in completion time. In its February 2026 update, METR identified selection and measurement problems in the subsequent study, making current effects difficult to estimate. The earlier finding should not be treated as a diagnosis of all 2026 tools.

The business implication is to test the actual process. Agreements should identify who handles exceptions, when review occurs and how the effort is measured. Ignoring supervision can make a proposal look competitive simply because some expenses were shifted to the customer.

Eight questions before approving a proposal

  1. What is an accepted delivery? Define content, quality and deadline.
  2. What is the baseline? Record current cost, volume, time and errors.
  3. What does implementation include? Identify integration, data, training and testing.
  4. What happens as usage grows? Understand overages, limits and pricing tiers.
  5. How much work remains with the team? Measure review, exceptions and corrections.
  6. How will usage and quality be monitored? Require process-level information and alerts.
  7. Which expense or business outcome can be attributed to the project? Separate revenue from margin and capacity from savings.
  8. How does the company leave the agreement? Check export, continuity and replacement.

The FinOps Foundation’s AI guidance emphasises tracking costs, setting usage limits and connecting spending to outcomes. Managers need to understand expenses during operation, with time to correct deviations.

A pilot should use real inputs and common exceptions. Selecting only well-organised documents can artificially reduce complexity. Decide in advance which failures stop expansion and which can be corrected without affecting customers.

At the end of the test, show the distribution of results. A favourable average can hide tasks that took much longer, cases returned to staff or users who abandoned the tool. Expansion requires understanding those differences.

What this calculation helps you decide

It helps select the first process, compare proposals over the same period and size supervision. It also shows when a simpler solution may meet the need at lower cost. Rules-based automation, existing software or better records may solve part of the problem.

The method does not establish a market price, forecast results for every company or prove that the fictional scenario will recur. The Brazilian cases in chapter 02 explain why application, lead time, productivity and financial outcomes must be investigated separately.

Before approving a project, the manager should be able to complete one sentence: we will invest this amount to deliver this work, at this volume, to this standard, within this period. If the sentence still depends on a promise, so does the budget.

Sources and documentation

  1. AWS: Amazon Bedrock pricing
  2. Google Cloud: generative model and service pricing
  3. FinOps Foundation: unit economics and business outcomes
  4. FinOps Foundation: financial management of AI
  5. IBM: total cost of ownership
  6. NBER: Generative AI at Work
  7. METR: experienced-developer experiment, 2025
  8. METR: experiment update and measurement limitations, February 2026

Sources accessed on 1 October 2026.

About the author

Luiz Claudio Rangel holds a master’s degree in Creative Economy Management, Strategy and Innovation from ESPM. He has more than 15 years of experience in strategy, marketing, sales and customer experience. As a Sebrae consultant, he has diagnosed and supported more than 300 companies. At NEODY, he writes about artificial intelligence, the digital economy, strategy and decision-making.