
A bank plans portfolios, products, branches and cost centres rather than a company. The budget model has to bring the loan and deposit portfolios, treasury, transfer income, allocations and provisions together into a forecast balance sheet that actually balances. On top of a model like that, AI stops being a demonstration and starts answering the finance team questions.
In short
- A bank budget model: portfolios, treasury, transfer pricing, allocations, forecast balance sheet.
- Top down and bottom up planning within one budget procedure.
- Monthly load of actuals from the core system and plan versus actual from the same data.
- An AI layer in five steps: from a question in plain language to scenarios and recommendations.
- Everything inside the bank, open models, no licence fees for the AI layer.
- Not tied to one vendor: IBM Planning Analytics, another platform, or RBC Group own development.
Sound familiar?
- The budget lives in dozens of Excel files whose links only one person remembers.
- Nobody is sure which version of the budget is current or which units have finished their part.
- Profitability by product, branch and cost centre, net of treasury compensation, is invisible.
- Transfer income and expense are calculated by hand, or not at all.
- Allocating indirect costs takes weeks and is redone from scratch every year.
- Planning runs on last year plus a percentage.
- The budget defence turns into an argument about figures rather than about decisions.
How the budget process runs

Strategic targets come down across the branch network, products, months and currencies, checked against the strategic plan. Operating budgets come up the other way: business lines and cost units fill in their own forms locally.
Consolidation calculates transfer pricing, allocations and provisions and produces the forecast balance sheet and the profit and loss budget. From then on actuals are loaded monthly from the core system, and the same reports show budget performance with export to Excel.
What the model covers

The full set of blocks we have built on bank projects:
- Loan portfolio: disbursements, repayments, rates, overdue
- Deposit portfolio and cost of funding
- Guarantees and letters of credit
- Non performing portfolio
- Treasury and funds transfer pricing
- Provisions and risk
- Fee income and expense
- Operating income and expense
- Administrative expenses
- Personnel costs
- Fixed assets, intangibles, depreciation
- Capital expenditure and new investment projects
- VAT
- Allocation of indirect costs
- Forecast balance sheet
- Profit and loss budget
- Segment reporting
- Budget approval workflow
Transfer pricing and allocations
These are the two mechanisms a bank builds a separate budget model for. Funds transfer pricing shows what a business line earns after settling with treasury for funding. Allocations spread indirect costs on defined keys, the same way every period.
Without them, product and branch profitability remains a matter of belief. With them it is calculated from the same postings as the reporting, and it survives a conversation with the board.
The AI layer: five steps

AI in budgeting does not switch on at once. Each step rests on the one before it: you cannot explain a variance without access to the data, look for anomalies without knowing the norm, forecast without history or recommend without a forecast.
- Access. A question in plain language instead of hunting for the right dashboard, answered from the model with a reference to the cells. Russian and Kazakh.
- Explanation. Variance against plan broken down by volume, rate, mix and currency, with a written comment on every material deviation.
- Control. Submissions from units and loaded actuals checked against history, against comparable units and against seasonality, before the budget is approved.
- Forecast. Machine learning models with external drivers on top of the platform built in forecast, with an accuracy report for every series.
- Decision. Scenario modelling, client clustering, recommendations on which drivers to move to reach the target.
The AI runs inside the bank
Built in agents of planning platforms live in the cloud. That does not suit a bank that keeps data inside its own perimeter, so we deploy the AI layer on the bank own infrastructure.
- Open models. Apache 2.0 and MIT licences, with no licence fees for the AI layer.
- The bank picks the model. Candidates are compared on reference questions from the finance team, in Russian and Kazakh, with a documented score. The decision is taken together with information security.
- Query log inside the perimeter. Every call to the model is recorded within the bank.
- User level rights. The agent reads the model in read only mode and sees exactly what the employee asking the question is allowed to see.
- Protection against prompt injection. A separate filter on the input, so the model cannot be steered off task or made to reveal more than it should.
What we build it on
The main platform is IBM Planning Analytics: that is where most of our bank implementations are, it has built in forecasting and the familiar way of working through Excel.
The methodology and the structure of the model do not depend on the tool. The same budget scope is built on another planning platform, or on our own development if the bank would rather avoid vendor licences. The choice of tool is discussed before the project rather than imposed by it.
CoInPro: RBC Group own development
CoInPro keeps contracts, indicators and projects inside one budget. The payment calendar, the profit and loss budget, the balance sheet and the cash flow are all built from a single journal of entries, so a project estimate and the contract signed against it are never counted twice.
Unit forms stay in Excel: the workbook is loaded as it is, and mapping rules turn every cell into postings with their dimensions. The product is at its first working version and we are taking it into pilot projects.
How we start: a pilot
We do not ask a bank to sign a large project blind. We start with a pilot of three to four months on real bank data.
- One working model, usually the loan portfolio, on the bank own data.
- Solution architecture and environment requirements.
- A requirements register for the full model, from which scope and timeline are calculated.
- A demonstration of the AI on the same data: questions, variance analysis, anomaly detection.
- A comparison of language models and a choice made together with information security.
After the pilot the bank sees the solution on its own figures and decides whether to go further. Part of the pilot cost is credited against the implementation.
What we need from the bank
- Access to management and statutory reporting and to extracts from the core system.
- Methodology: the current budget procedure, allocation keys, transfer pricing rules. If they do not exist, we develop them together.
- An environment for the model and for the AI layer inside the perimeter.
- A working group: finance, IT and information security.
RBC Group experience
We have worked with banks in Kazakhstan, Uzbekistan and the wider region since 1991: budgeting, management reporting, data warehouses and regulatory data marts. Related solutions are described on the Solutions for banks, data warehouse and activity based costing pages.
Questions and answers
How is bank budgeting different from budgeting in an ordinary company?
The unit of planning is not an expense line but a portfolio, a product, a branch and a cost centre. On top of that come funds transfer pricing between treasury and the business lines, allocation of indirect costs on defined keys, and a forecast balance sheet that has to balance. An ordinary budget model has none of these mechanisms.
Does it have to be IBM Planning Analytics?
No. We run these projects on IBM Planning Analytics because that is where most of our bank implementations are, but the methodology and the model do not depend on the tool. The same solution is built on another planning platform or on RBC Group own development.
Where does the AI run and where does the data go?
The whole layer is deployed inside the bank. The language model, semantic search, forecasting models and agents run on the bank servers, and every query is logged inside the perimeter. Bank data is never sent outside.
Which models does the AI layer use?
Open models under Apache 2.0 and MIT licences. There are no licence fees for the AI layer. The bank chooses the specific model together with its information security team, based on a measured comparison on its own questions, including questions in Kazakh.
Why not the vendor built in agent?
Built in agents of planning platforms run in their cloud versions. A bank that keeps data inside its own perimeter cannot use a cloud agent, and ML forecasting, anomaly detection and clustering on warehouse data are not part of it anyway.
Can we start small?
Yes, and we recommend exactly that. A pilot of three to four months: one model on real bank data, the architecture, a requirements register and a demonstration of the AI on the same data. After the pilot both the scope and the effect are visible.
How long does a full implementation take?
The bank budget model takes seven months or more to go live, plus support through the first budget campaign in the system. The AI layer is delivered in releases, each starting once the data it works on is ready.
Let us look at your bank budget model
We will show what the forecast balance sheet, transfer pricing and variance analysis look like on your bank structure, and propose the scope of a pilot.
Request a meetingmarketing@rbc-grp.solutions, +7 727 258 1712
