A data warehouse for AI: why corporate agents hallucinate without a DWH

An AI agent in finance or banking performs exactly as well as the data it is given is clean and connected. Without a corporate data warehouse (DWH) that holds a single version of the truth, history and access rights, the agent answers with confidence but not necessarily correctly. That is why the first money spent on AI is better invested in foundations – a warehouse, data marts and planning on top of them – than in yet another bot.

Numbers that sober you up

At conferences AI is sold as a magic wand. In earnings reports the wand still works more like a pointer.

  • Gartner: companies will abandon 60% of AI projects that are not supported by AI-ready data. 63% of organisations either lack the data management practices they need or are unsure whether they have them (Gartner).
  • McKinsey, State of AI 2026: 37% of respondents see any effect of AI on EBIT at all. The share of “high performers”, where AI delivers 5% of EBIT or more, is still around 6% (FM Magazine).
  • PwC, CEO survey in Kazakhstan (2026): only 12% of Kazakh executives have seen tangible additional revenue from AI, against roughly a third worldwide. 76% see no financial effect whatsoever, against 56% globally (PwC Kazakhstan).

The conclusion is simple. The money from AI goes not to those with the smarter model, but to those whose data is in order and wired into their processes. PwC says it outright: isolated AI projects rarely produce a measurable result.

Why AI without a warehouse starts making things up

Picture a new CFO who is handed five Excel files on day one, each with a different revenue figure for the same quarter. A human being will at least ask which one is real. An AI agent will not. It will take the first number it finds and explain elegantly why it went up.

Corporate AI “hallucinations” usually have three roots.

  1. Several versions of the truth. Revenue is counted one way in the ERP, another in the CRM and a third in the management report. Headcount in the HR system does not reconcile with payroll in accounting. The agent does not know who is right, and it should not be the one deciding that for you.
  2. No history and no context. A forecast needs three to five years of comparable data. In real life the chart of accounts has been rebuilt twice, branches have been merged, and every warehouse maintains its own item catalogue. Gartner names it as a key requirement that data be representative of the task, errors and outliers included.
  3. No access rules and no data lineage. If you cannot show where a number came from and who is entitled to see it, the agent will either lie in a way nobody can disprove, or show the CEO’s salary to a procurement manager. The second is worse.

A data warehouse solves all three at once. It is not “another database” but the company’s agreement about what counts as the truth, written down in code.

What an AI-ready warehouse looks like

A good DWH is like a restaurant kitchen: raw produce arrives from different suppliers, gets washed, sorted into stations, and only then handed to the chef. The AI agent is the chef. Sending it straight to a storeroom of unwashed potatoes is not a good idea.

A working architecture for a holding company or a bank looks like this, bottom up:

  1. Sources. ERP (JD Edwards, SunSystems, legacy accounting systems), HCM and payroll, CRM, the core banking system, branch spreadsheets.
  2. Ingestion and cleansing. Regular pipelines, quality checks, an error log.
  3. DWH core. Shared master data (customers, items, org structure, accounts), change history, records from different systems linked together.
  4. Data marts. Management, regulatory, HR, credit – each built for a specific business question.
  5. Metadata layer. A business glossary (“what we call EBITDA”), lineage (where a number came from), roles and rights. This is the layer that turns an ordinary warehouse into an AI-ready one.
  6. Planning and BI. IBM Planning Analytics (TM1) for budgeting, forecasting and consolidation; dashboards for management.
  7. AI layer. Agents and assistants that read the marts and planning models with the rights of the specific user.

Note that AI sits on the seventh floor. Start building the house there and it will be very modern and very short-lived.

Where the AI actually is

The good news: vendors are now building agents on top of data and planning models rather than in thin air. Here is what exists today, not in a 2030 slide deck.

An agent inside planning

In June 2026 IBM released the Planning Analytics Agent. It runs in Planning Analytics Workspace 2.1.21 and 3.1.9 and in the Excel add-in. The function names speak for themselves: Explain Cell (“why does this cell show this number”), Forecast Insights, Chart Insights. For custom scenarios there are more than 30 MCP tools for analysis, modelling and reporting, which can be connected to watsonx Orchestrate and other MCP-compatible agents.

In practice a CFO asks in plain words: “Why did the Shymkent branch miss its margin in the third quarter?” The agent breaks the variance down by driver inside the model itself – but only if the actuals reached that model from the DWH clean and on time.

An agent inside the warehouse

On 21 October 2026 IBM made agentic data integration generally available in watsonx.data integration. A user describes the data flow they need in words, the agent proposes sources and a pipeline design, a human approves it. More important still: the agent works with the rights of whoever invoked it, and its actions are logged together with lineage. For a bank under supervision that is not a detail, it is a condition of entry.

What you gain and what you risk

Scenario Data it runs on Effect Main risk
Explaining plan/actual variances Planning Analytics cubes + actuals from the DWH Analysts spend hours on conclusions, not on reconciliation The agent explains a load error as a “market trend”
Statistical forecasting 3-5 years of comparable history A baseline forecast in minutes; people argue about the adjustments Gaps in history after reorganisations
Data quality monitoring Pipelines and load logs The problem is visible before the board sees it False alarms that people stop reacting to
Board questions in plain language Data marts + business glossary An answer without raising an IT ticket Confused definitions, data leaking past access rights

One more market signal: according to McKinsey, the share of large companies (above $1bn in revenue) scaling agents in at least one function has grown from 27% to 40%. For smaller companies it is stuck at 22%. The gap is built on data infrastructure, not on models.

Kazakhstan: the demand is there, the foundation is not

2026 has been declared the Year of Digitalisation and Artificial Intelligence in Kazakhstan. The Law on Artificial Intelligence No. 230-VIII was adopted on 17 November 2025 and is already in force. According to figures presented by the financial regulator ARDFM, 75% of the country’s banks already use AI: scoring, anti-fraud, marketing, customer service (DKNews).

And now the main point. The brakes that financial institutions named were data quality, the absence of common standards and a shortage of specialists. Not models and not budgets – precisely what a warehouse addresses. In the real economy the picture is similar: per PwC, only 19% of Kazakh CEOs report substantial use of AI in finance, legal, tax and HR.

Typical mistakes

  • “Bot first, we’ll sort the rest out later.” A demo on three pretty questions goes down well, and on the fourth real one the agent finds three different revenue figures. The board’s trust is lost in a single meeting.
  • Treating the DWH as an IT project. If IT alone writes the metric definitions, finance will not accept them. The owner of the data is the business, not the server room.
  • “Everything at once.” Trying to pull every system in the company into the warehouse before the first result. A year later the project is eating the budget and the board is eating the project manager.
  • Planning divorced from actuals. The budget model is fed by manual exports, and the forecast goes stale before it is approved.
  • Access rights “later”. An agent on a service account sees everything. That is a direct route to an incident.

Timelines and money: orders of magnitude

The estimates below are a basis for a conversation, not a price list: everything depends on the number of sources, licences and the state of the data.

Stage Scope Duration
Data assessment Source map, 20-30 key metrics, quality review 3-6 weeks
First data mart + planning 2-3 sources, management mart, budget model in Planning Analytics 3-5 months
AI pilot on top of it One scenario: variance explanation or forecasting 6-10 weeks
Scaling Remaining domains, regulatory marts, agents for different roles 9-18 months, in waves

On money the rule is simple: an assessment and a pilot usually land in the tens of thousands of dollars, a full warehouse for a holding company or a bank in the hundreds of thousands. The most expensive item is not the project – it is another year of decisions based on numbers nobody trusts.

What an owner should do: 5 steps

  1. Ask one question about money. Not “let’s implement AI”, but “close the month and refresh the forecast in 5 days instead of 15” or “see margin by branch every week”.
  2. Name the owners of your metrics. Revenue, margin, headcount and payroll each need one person with a name, not “a department”.
  3. Commission a data assessment before choosing an AI tool. Which sources, how many years of history, where the discrepancies are, who is entitled to see what.
  4. Build a mart for a question, not a warehouse for everything. First mart plus planning model – and only then an agent on top.
  5. Measure the agent like an employee. Accuracy of answers, hours saved, number of errors caught by a human. No metric, no scaling.

In closing

AI in finance and banking is not about whose chatbot is smarter. It is about whose numbers add up the first time. A data warehouse and planning on IBM Planning Analytics give the agent what it does not have by birth: memory, rules and a conscience in the form of lineage. Whoever lays the foundation in 2026-2027 will get a return from agents while competitors are still reconciling spreadsheets.

If you want a sober assessment of whether your data is ready for AI, and which mart to start with, the RBC Group Solutions team works with DWH, IBM Planning Analytics and the AI layer on top of them across Kazakhstan, Uzbekistan and Ukraine. Let’s start with a conversation about your main question about money.

Frequently asked questions

Do you need a data warehouse to deploy AI in a company?

For a toy chatbot, no. For AI that answers questions about money, people and risk, yes: it needs shared definitions, history and access rights. Gartner forecasts that 60% of AI projects will be abandoned without AI-ready data.

What is AI-ready data?

Data that fully describes a specific task, carries metadata and is continuously monitored for quality. It is an ongoing practice, not a one-off clean-up.

What can AI do in IBM Planning Analytics?

The Planning Analytics Agent answers questions in plain language, explains cell values and helps with forecasts and charts. More than 30 MCP tools let you connect planning models to other agents.

How long does the first result take?

As a guide: a data assessment takes 3-6 weeks, a first mart with a planning model 3-5 months, an AI pilot on top of it 6-10 weeks.

What are the risks of AI agents in finance?

Confident wrong answers on dirty data, and access that bypasses the user’s rights. The cure is lineage, a role model and a human who approves decisions.

Is business in Kazakhstan ready for AI?

The demand is there: 75% of banks already use AI. But only 12% of CEOs see additional revenue from it, and the main brake is data quality.

Sources

  1. Gartner: Lack of AI-Ready Data Puts AI Projects at Risk – February 2025.
  2. FM Magazine: a review of McKinsey State of AI 2026 – September 2026.
  3. PwC Kazakhstan: 29th Annual CEO Survey – 2026.
  4. DKNews: 75% of banks in Kazakhstan already use AI – August 2026.
  5. IBM Community: IBM Planning Analytics Agent – June 2026.
  6. IBM: agentic data integration in watsonx.data integration – October 2026.



Leave a Reply