> ## Content Index
> Fetch the complete content index at: https://www.slowin.ski/llms.txt
> Use this file to discover other available public pages before exploring further.

# Your AI answers today's questions with yesterday's data
- URL: https://www.slowin.ski/data-freshness-ai/
- Published: 2026-10-27T09:00:00.000Z
- Updated: 2026-10-27T09:00:00.000Z
- Description: A nightly batch was fine for a dashboard read in the morning. An AI assistant answering in the afternoon states yesterday's data in the present tense, and the fix is not streaming everything but declaring how old each answer's data is allowed to be.
- Author: Krzysztof Słowiński
- Tags: Fractional CTO, AI Readiness, Data Freshness, Data Engineering, Data Architecture, Engineering Leadership

I have seen important decisions made on data that people assumed was current and was, in fact, a day old. Nothing had failed. The nightly batch ran on time, the freshness SLA reported green, and the data was exactly as old as it had been designed to be.

Take a wind farm operator as an example. Its maintenance planners use an AI assistant that combines work orders with turbine performance data in the data warehouse, so it reads the work orders synced overnight from the maintenance system, not the live one. At 13:20 a planner asks which turbines still have open work orders for tomorrow's crew plan. The answer comes from last night's sync: turbine 7's fault, closed by a technician that morning, is still listed, and turbine 12's fault, logged at 09:47, is missing. The crew is planned for a turbine that is already fixed, and turbine 12 loses another day of generation. The alarm systems run in real time, as they should; it is the planning data around them that is a day old.

The batch did not become wrong. The reader changed. A dashboard at least shows when it was last refreshed, so a planner reading it at 08:00 has a cue; an assistant answering at 13:20 states yesterday's data in the present tense, with no "as of". **Freshness is a property of the decision being made, not of the data pipeline.**

## An SLA measured in the wrong place

Freshness SLAs were written for reporting cadence: data loaded by 07:00, monitored as job success. That measures when the pipeline finished, not how old the data is at the moment someone uses it. When an AI feature started reading the same tables, nobody renegotiated it.

The measure that matters is how old the data is: how long ago it was true in the system where it was recorded. That takes two timestamps on each record. Capture time is when the change happened in the original system, such as the technician closing turbine 7's work order; it drives the freshness requirement and the "as of" label. Load time is when the record arrived in the data warehouse; it tells the data team where a delay comes from.

Some systems keep no capture time and can only be copied as periodic snapshots. There, the snapshot time is the closest stand-in for capture time: use it for the "as of" label, but name it as the snapshot time, so nobody mistakes the moment the data was copied for the moment it was true.

## Faster pipelines answer the wrong question

When a stale answer surfaces, the usual response is to make the pipelines faster: every batch moved to hourly, or streaming across the board. Both treat freshness as a speed problem. Cost goes up, while the real question, how old the data behind this particular answer is allowed to be, stays unasked. For crew planning, a few hours may be perfectly acceptable; for a status shown to a customer, it may not be.

## Nobody owns the number

The owner of the AI feature assumes the data team handles freshness, and the data team was never told the reader changed. "How old can this data be before the answer becomes wrong?" is a product decision, not a pipeline one, so it falls between them.

The gap becomes visible when a decision is challenged later, in an audit or a post-mortem. Someone will ask what data the answer was based on and how old it was. The assistant cannot be held accountable for that answer; someone in the organisation has to be.

## The leadership mandate

The CTO's job is to make freshness a declared requirement per use case rather than a property of the pipeline. For each AI use case, someone owns one number, the maximum acceptable data age, and one decision, what happens when it is exceeded. Streaming then becomes a targeted upgrade for the few use cases whose declared freshness needs it, not a reaction to every stale answer.

## The path forward

Make every answer carry three facts:

1. The "as of" time: when the data was captured.
2. The limit: the maximum data age this use case allows.
3. The next refresh: when newer data is scheduled to arrive.

This matters more as assistants become AI agents that act on the answer, booking the crew rather than suggesting it: then no person reads the reply, and nobody is left to notice how old the data was.

When the data is older than the limit, the assistant says so ("based on data as of 02:00, older than this question allows; next refresh tonight") and hands high-stakes questions to a person. It does not trigger the pipeline itself. Where fresh data is needed, design an approved path once: a rate-limited on-demand refresh, or a live lookup of the one field that matters.

The nightly batch can stay exactly as it is. What has to change is that every answer built on it says how old it is.

[Get in touch](https://www.linkedin.com/in/krzysztofslowinski/?ref=slowin.ski)

## ****The data foundation briefing**

****Strategic insights on architecting AI-ready data foundations for engineering leaders.**

Subscribe 

Email sent! Check your inbox to complete your signup. 

No spam. Unsubscribe anytime.