An AI budget for 2027 should not be built as a single line item for “neural networks,” but as a set of categories tied to specific processes: data audit and preparation, development and integration, model usage (tokens or in-house infrastructure), support, employee training, security, and contingency reserves. It’s better to allocate the money in tranches: the first for the audit and pilot, and the rest only after the pilot delivers results against a pre-agreed metric. This kind of budget is easier to defend to the CFO and harder to waste.
This article was prepared by AIrassvet using open sources verified on October 5, 2026. It is intended for business owners, CEOs, CFOs, and leaders of IT and digital teams who, during budget season, are deciding how much to allocate to AI next year and for what. Market figures are cited with attribution; all calculation examples are illustrations based on hypothetical volumes, not forecasts for your company. AIrassvet is not affiliated with the vendors or research firms mentioned; the section at the end describes the publisher’s services.
Contents
- Why you need to plan your AI budget now
- What makes up an AI budget
- Step 1. Start with a list of processes, not a total amount
- Step 2. Calculate model costs
- Step 3. Break the budget into tranches
- Step 4. Build in reserves and an exit plan
- Step 5. Prepare answers for the CFO
- Common AI budgeting mistakes
- Example AI budget structure for mid-sized businesses
- Frequently asked questions
- Primary sources and verification method
Why you need to plan your AI budget now
October through December is when most companies build next year’s budget. If AI doesn’t make it into the plan now, projects in 2027 will have to be funded from reserves or delayed until the budget is revised.
The market is moving toward higher spending, but with one important caveat:
- Forrester in its 2027 budget planning guides surveyed more than 2,600 business and IT leaders: more than 80% expect budgets to grow over the next 12 months, and up to a quarter of companies plan to increase spending by at least 10%. The analysts’ key takeaway, as paraphrased by K2 NeuroTech, is that simply increasing AI spending does not create long-term value without upgrading data, the operating model, and project governance (spbit.ru, August 7, 2026).
- In Russia, an online survey by VK Tech and MARC Agency of 154 large companies found that 74% plan to invest in AI solutions over the next two years, 62% plan to invest in AI agents, and 45% say data security is their biggest challenge (Vedomosti, April 3, 2026).
- AI is taking money away from adjacent budget lines. In a U.S. study by IANS of 500 security leaders, 69% said AI is a top priority for new security spending, while 55% of companies said their overall security budget is flat or shrinking (CISOClub, September 16, 2026).
Planning takeaway: the question is no longer whether to allocate money to AI, but how to distribute it so you have something to show a year from now. The study samples are different (global, Russian large enterprise, and U.S. CISOs), so their figures should not be added together or applied directly to small businesses.
What makes up an AI budget
The most common mistake is budgeting only for development or only for a model subscription. In a real project, expenses are spread across several categories, and some of them recur every month.
| Category | What’s included | Expense type | What people often forget |
|---|---|---|---|
| Process audit and selection | Process analysis, data assessment, impact calculation | One-time | A baseline “before” measurement — without it, there’s nothing to compare against |
| Data preparation | Cleaning, labeling, exports, access rights | One-time and partially recurring | The work of internal data owners |
| Development and integration | MVP, integration with CRM, 1C, email, and messaging apps | One-time | Testing on real data and revisions after the pilot |
| Model usage | API tokens or in-house GPU servers | Recurring, volume-dependent | Increased usage after a successful pilot |
| Support and maintenance | Monitoring, fixes, model and prompt updates | Recurring | Model version changes by the provider |
| Employee training | Training on how to use AI, policies and procedures | One-time and recurring | Employee time spent on training |
| Security and compliance | Permissions, logs, anonymization, review of 152-FZ requirements | One-time and recurring | Audit of AI agent permissions |
| Reserve | Unexpected changes, price increases, provider switch | Contingency reserve | Exit plan if the provider shuts down the service |
We covered in detail what goes into the cost of the projects themselves in the article “The Cost of Implementing AI in Business in 2026”, and recurring support costs in the piece on AI agent support and SLAs.
Step 1. Start with a list of processes, not a total amount
A top-down budget (“let’s allocate this much to AI”) almost always either goes unspent or gets spent on experiments that produce no results. It’s more reliable to work bottom-up:
- Make a list of candidates — processes with lots of repetitive work involving text, documents, calls, or data: request processing, source documents, customer responses, reporting, quality control.
- For each process, record the baseline: how many transactions per month, how much time per transaction, how many errors, how many people are involved.
- Assess data readiness: where the data is stored, in what format, whether API access is available, whether there is personal data.
- Select two or three processes with the highest volume and a clear metric. Put the rest on a reserve list for the second half of the year.
If you haven’t done this analysis yet, it should be the first line item in the budget. How to do it quickly is described in the article on auditing business processes before implementing AI.
Step 2. Calculate model costs
Model usage costs are the only line item that grows with the project’s success: the more processes you move to AI, the more tokens you use. That’s why you need to calculate it based on volume, not pull it out of thin air.
Calculation example (illustration using hypothetical volumes). Let’s take an AI assistant that handles 1,000 inquiries per business day, 22 business days per month. Each inquiry uses about 3,000 input tokens (question, instructions, knowledge base excerpts) and 500 output tokens. Prices are based on OpenAI’s pricing for GPT-6.1 Sol as of October 5, 2026: $2 per 1 million input tokens and $10 per 1 million output tokens (OpenAI documentation).
| Metric | Calculation | Result |
|---|---|---|
| Input tokens per month | 1,000 × 3,000 × 22 | 66 million |
| Output tokens per month | 1,000 × 500 × 22 | 11 million |
| Input token cost | 66 × $2 | $132 |
| Output token cost | 11 × $10 | $110 |
| Total per month | 132 + 110 | $242 |
What’s important to understand about this calculation:
- The volumes in the example are hypothetical. Use your own numbers: the number of transactions, context length, and response length can be measured in the pilot.
- Caching changes the numbers. According to the same OpenAI page, cached input tokens cost $0.10 per 1 million — if the instructions and knowledge base are repeated, input cost can be significantly lower.
- Agents consume more than chatbots. An agent that performs multiple steps and calls tools may use several times more tokens for one task than a simple response. Budget based on pilot results, not on the assumption of “one request — one answer.”
- Convert to rubles using the budget exchange rate and keep in mind that foreign services are available to Russian companies with limitations. Alternatives include Russian models or your own server; a comparison of cost models is provided in the article “Cloud or On-Prem LLM: TCO”.
We covered how to keep model costs under control as volume grows in the article on stable AI costs as token usage grows.
Step 3. Break the budget into tranches
Phased funding is the main tool for protecting an AI budget. Instead of approving the full annual amount up front, approve the framework and the conditions for moving between stages.
| Tranche | What is funded | Condition for moving to the next stage |
|---|---|---|
| Tranche 1: audit and pilot | Process analysis, data preparation, MVP for one process | The pilot achieved the agreed metric on real data |
| Tranche 2: production launch | Integration, security, employee training, scaling across the full process | Metrics are stable in operation, and model costs are within plan |
| Tranche 3: rollout | Expansion to adjacent processes, shared platform | The first process has delivered confirmed results over a quarter |
This approach removes the CFO’s biggest fear: spending an annual budget on a project that never works. If the pilot does not hit the metric, you lose only the first tranche and gain insight into why it didn’t work.
Step 4. Build in a reserve and an exit plan
The AI market changes faster than the budget year lasts: providers release new model versions with different pricing and retire older ones, while Russian companies’ access to foreign services can change. That’s why your budget should include:
- a reserve for price and volume changes — in model usage and support costs;
- a reserve for switching providers — if a model is discontinued or access is restricted, the project may need to be migrated;
- funding for an exit plan — an architecture that lets you replace the model without rewriting the entire solution. How to do this is covered in the article “AI Service Exit Plan”.
The size of the reserve depends on how tightly the project is tied to a single provider. We are not giving a universal percentage: it should come from the risk assessment for the specific project.
Step 5. Prepare answers for the CFO
An AI budget is protected not by a presentation about trends, but by answers to specific questions. Prepare them in advance.
| CFO question | What to Prepare |
|---|---|
| What Do We Get for This Money? | A List of Processes with Current Metrics and a Target KPI |
| How Will We Know It Worked? | The KPI, measurement method, and pilot duration |
| What Happens If It Doesn’t Work? | Tranche 1 as the maximum loss, stop criteria |
| How Much Will It Cost Each Month After Launch? | A calculation of model and support costs based on volume |
| Why Not Wait a Year? | The cost of the current manual process and competitor actions |
| What Are the Data Risks? | Where the data is processed and how 152-FZ requirements are met |
| Can We Get a Subsidy or Grant? | A review of available support programs |
A detailed explanation of effect metrics and ROI calculations is provided in the article “The Impact of AI Agents: Metrics and ROI”, and government support programs are covered in the overview of AI project grants in 2026.
Common AI Budgeting Mistakes
- A budget with no processes. The money is allocated but not tied to tasks, so it goes to subscriptions and experiments with no results.
- Development only, with no operations. The pilot is built, but there’s no money for tokens, support, or improvements, so the solution stops after a few months.
- No baseline. A year later, it’s impossible to prove the impact because no one measured the starting point.
- The full amount upfront. Without tranches and stop criteria, a failed project consumes the entire budget.
- Security “later.” Access rights, logs, and personal data processing are added after an incident instead of during the architecture stage—and they cost more.
- Forgotten employee time. Process owners and data owners spend time on scoping, testing, and training—that’s also a project resource.
An Example Budget Structure for a Mid-Sized Business
Below is a template you can copy into a budget spreadsheet. Amounts are intentionally omitted: they depend on the processes, volumes, and chosen architecture.
- Process and Data Audit — Q1, one-time.
- Pilot for the First Process — Q1–Q2, one-time: data preparation, MVP, testing.
- Model Costs for the Pilot — starting from the pilot launch, monthly, calculated based on volume.
- Production Launch of the First Process — Q2–Q3, if the pilot metric is achieved.
- Model Costs in Production — monthly after launch, recalculated based on actual volume.
- Support — monthly after launch.
- Employee Training — by launch time and for new employees.
- Security and Compliance — during the architecture stage and annual review.
- Second Process — Q3–Q4, if the first process shows confirmed results.
- Reserve — for the full year.
Frequently Asked Questions
How Much Should We Budget for AI in 2027?
There is no universal number. Start with a list of two or three processes with a clear metric, estimate one-time costs for audit, data, and development for each, monthly costs for models and support, and then add a reserve. The budget should follow the processes, not the other way around.
What Expense Categories Should Be Included in an AI Budget?
Audit and process selection, data preparation, development and integration, model usage (tokens or your own servers), support, employee training, security and compliance, plus a reserve for price changes and provider switches.
How Do You Calculate Token Costs?
Multiply the number of operations per month by the average number of input and output tokens per operation, then by the provider’s price per million tokens. It’s better to measure average volumes during the pilot: agents with multiple steps consume significantly more tokens than simple chatbot replies.
How Do You Convince the CFO to Fund AI?
Show specific processes with current metrics, the target KPI, a calculation of monthly costs, and phased funding: the first tranche for audit and the pilot, with the next tranches only after the metric is achieved. That way, the maximum loss is limited to the first tranche.
What’s Better: Cloud Models or an On-Premises Server?
It depends on volume, data requirements, and service availability. For small or uneven workloads, a cloud API is usually simpler. For large, stable volumes and strict data requirements, it makes sense to evaluate an on-premises deployment. You should compare total cost of ownership, not just token price.
Do You Need a Reserve in an AI Budget?
Yes. Provider prices and models change more often than once a year, and access to certain services may be restricted. A reserve is needed for volume growth, price changes, and moving the solution to another model. Its size should be based on a risk assessment for the specific project.
Primary Sources and Verification Method
Verified on October 5, 2026:
- spbit.ru, August 7, 2026 — a summary of the findings from the Forrester Budget Planning Guides 2027 (survey of more than 2,600 executives) by K2 NeuroTech.
- Vedomosti, April 3, 2026 — an online survey by VK Tech and MARC of 154 large Russian companies.
- CISOClub, September 16, 2026 — a summary of the IANS 2026 Security Budget Benchmark Report (500 U.S. security leaders).
- OpenAI documentation for GPT-6.1 Sol — prices per 1 million tokens: input $2, cached input $0.10, output $10.
The token cost calculation is based on the listed prices and hypothetical volumes and is for illustration only. The structure of the expense categories, tranches, CFO questions, and common mistakes is an editorial analysis by AIrassvet. The studies were conducted on different samples and are not a forecast for any specific company.
How AIrassvet Helps You Plan an AI Budget
If you’re building next year’s budget and want to include AI in a way that lets you justify and control the spend, you can break the task down step by step. On the AIrassvet website, which also uses the name “Vladimir Lomtev’s AI Studio,” the applicable areas are described as follows:
- Process audit: selecting candidate processes, capturing current metrics, and assessing data readiness.
- Data preparation and MVP development for one process as the first budget tranche.
- Integration of AI agents with CRM, 1C, email, and messaging apps when moving toward production deployment.
- Deploying language models within the company’s own environment if the calculation shows that in-house infrastructure is a better fit than a cloud API.
The first step is to choose one process, describe its current metrics, data sources, constraints, and the criterion by which you will consider the pilot successful.
Bottom line: budget from processes and in tranches
The AI budget for 2027 should be built from the bottom up — from specific processes with a measured baseline and a clear metric, not from a lump sum “for neural networks.” It should include not only development, but also recurring costs for models and support, training, security, and contingency reserves.
The main way to protect the budget is through tranches: money for production deployment is allocated only after the pilot has delivered results. This way, the company limits risk without pushing AI off for another year.