# Where Did This Number Come From? How Growth Companies Build Data Lineage to Trace AI Agent Decisions

> Data lineage shows where the number an AI agent used actually came from. How growth companies build traceability before agents go into production.

- Published: 2026-09-17
- Author: Master Mind
- Canonical: https://aimaster.fi/en/artikkelit/mista-tama-luku-tulee-nain-kasvuyritys-rakentaa-data-lineagen-ai-agentin-paatost

A CEO asks one question in the meeting: where did this number come from? When the answer is an AI agent, that question often has no answer. The agent combined data from three systems, calculated a subtotal, and produced a result — but the path back to the original data got lost along the way.

Data lineage means tracing data's path from source to outcome: which system the information came from, what transformations it went through, and which decision it ended up feeding. When an AI agent makes a business decision or calculates a number, data lineage shows what that decision was based on — not guessed after the fact, but recorded.

## Why is fragmented data without traceability a risk in agent projects?

A growth company connects its CRM, ERP, invoicing system, and spreadsheets into one data foundation for an AI agent. Every system changes the data along the way: renaming columns, rounding numbers, filtering rows. Once the agent produces a result, nobody knows anymore which transformation affected what — and tracing an error turns into guesswork.

This isn't just a technical problem. When a customer, auditor, or board member asks what a proposed price or risk rating is based on, a vague answer erodes trust in the whole system faster than a single mistake would.

## What exactly does data lineage mean?

Data lineage is a documented path describing data's journey from a source system to its final use: which table the information came from, what transformation processed it, and when and by whom it was approved. Unlike a [semantic layer](https://aimaster.fi/en/artikkelit/yksi-luku-monta-totuutta-nain-kasvuyritys-rakentaa-semanttisen-mittarikerroksen-), which unifies calculation rules, data lineage answers a historical question: where did this specific number come from and what happened to it on the way.

In practice, data lineage is built at three levels: table level (which source table the data comes from), column level (which source column maps to which output), and event level (which agent run used which data and when). For most growth companies, table- and column-level tracing is enough — a full event log is only needed once an agent makes regulated or financially significant decisions.

## How does a growth company build data lineage in practice?

The first step isn't buying a tool — it's mapping: which systems does the AI agent pull data from, and which transformations affect that data before the agent ever sees it. This mapping happens as part of the [Master Layer](https://aimaster.fi/en/tuotteet/master-layer) phase, which connects a company's existing systems securely for AI use.

- Map the sources: list every system the agent pulls data from, directly or indirectly.
- Mark transformation points: where data gets merged, filtered, or recalculated before the agent uses it.
- Capture metadata automatically: lineage information doesn't work if it has to be maintained by hand — it gets logged as part of the integration.
- Scale the depth to the need: critical decisions (pricing, credit risk, contractual obligations) need event-level logs; routine reports are fine at table level.

Once lineage is built on top of Master Layer, a [Master Mind](https://aimaster.fi/en/tuotteet/master-mind) agent can attach a reference to its origin for every number it produces. In practice, that means the agent can answer "where did this number come from" immediately — not after three days of investigation.

## What does the EU AI Act require on traceability?

Article 12 of the AI Act requires high-risk AI systems to technically allow automatic recording of events over the system's lifetime, to keep its operation traceable (EU AI Act, Article 12). The obligation applies from 2 December 2027 for Annex III high-risk systems and from 2 August 2028 for Annex I systems. It formally applies only to systems classified as high-risk, but the same principle is worth applying to any agent that affects pricing, credit decisions, or customer contracts — the regulation is coming regardless.

Data lineage isn't the log the AI Act requires, but it's a precondition for it: if you don't know where your data came from, you can't reliably record what an agent's decision was based on.

## How does data lineage differ from data observability?

Data observability monitors data health in real time — does the pipeline show anomalies or gaps right now. Data lineage answers a different question: where did this information come from and what happened to it along the way. [Data observability](https://aimaster.fi/en/artikkelit/ai-agentti-ei-valehtele-se-on-itsevarma-vaarassa-huonon-datan-takia-nain-rakenna) tells you the data is broken right now. Data lineage tells you where in the chain it broke and what else depends on it.

| Feature | Data lineage | Data observability |
| --- | --- | --- |
| Question | Where did this information come from? | Is the data healthy right now? |
| Time direction | Backward, historical path | Present and real-time monitoring |
| Use case for agents | Tracing and auditing decisions | Detecting faulty input |

## What happens when traceability is missing?

A Finnish B2B telecom operator found that its AI-assisted billing report showed two different margins for the same month to two business units. The cause wasn't bad data — it was two different transformation paths: one report calculated margin before discounts, the other after. Without documented sources, the discrepancy only surfaced after weeks of manual investigation. Data lineage would have shown the divergence immediately — both paths visible side by side.

## Should you build data lineage for all your data at once?

No. Coverage beats scope. Start with the data feeding agents that affect money, customer relationships, or regulation. The rest can wait. A sprint model makes this predictable: development moves in 3-day cycles, and the first sprint covers one critical data path end to end — not the entire company's data at once.

## What does it cost to implement data lineage?

Cost depends on how many systems are involved and how many of them already expose metadata in a structured format. Modern cloud systems often generate metadata automatically, while older ERP or customer systems need a custom wrapper. The first sprint maps the scope and delivers a concrete estimate — not a guess that grows as the project drags on.

## Summary: traceability is a precondition for trust

An AI agent that can't show where its numbers came from isn't ready to make decisions involving money or regulation. Data lineage isn't an extra layer — it's the part of Master Layer that makes an agent's answers verifiable.

## Frequently asked questions

## Frequently asked questions

### What does data lineage mean in simple terms?

Data lineage is a documented path showing which system information came from, what transformations it went through, and what outcome it fed into. It answers "where did this number come from" without after-the-fact investigation.

### How does data lineage differ from data observability?

Data observability monitors data health in real time — whether the pipeline shows anomalies right now. Data lineage answers a historical question: where did the data come from and what happened to it along the way. Both are needed, but they solve different problems.

### Does the EU AI Act require data lineage?

The AI Act doesn't explicitly require data lineage, but Article 12 requires high-risk systems to keep traceable event logs starting 2 December 2027 (Annex III) and 2 August 2028 (Annex I). Data lineage is the practical precondition for meeting that requirement.

### Where should a growth company start building data lineage?

With the data feeding agents that affect pricing, credit risk, or customer contracts. Mapping happens first as part of the Master Layer phase, and the first sprint covers one critical data path end to end.

### Does a small growth company need data lineage with just a few systems?

Yes, but at a smaller scale. Connecting even three systems for an agent is enough to create a situation where nobody knows why two reports show different numbers. Scope grows with the number of systems, but the need starts with the first integration.