Working with AI on analysis · 4/7
Being able to find the data does not mean another analyst or agent can continue the team’s past analysis.
We want an AI assistant that understands the business, answers questions, and does not get tired. But once the budget is approved and a capable model is available, what can we give it beyond database access, schemas, and a few metric documents?
Data access does not make past analysis transferable
Suppose the team analyzed returning users three months ago. The SQL and report still exist. A new version launches, and a manager asks for the analysis again.
The agent finds the old SQL, changes the dates, and produces results. Does it know why a channel was excluded last time? Was the data unreliable, or was there a special campaign? Does the old segmentation still apply?
Those answers are not necessarily visible in the code. Sometimes even the original analyst has to stop and ask what they were doing.
Data infrastructure handles access, quality, permissions, and computation. That does not automatically preserve the methods, validation, and judgments formed during analysis.
| What the team prepares | Questions it helps answer |
|---|---|
| Data access, quality, permissions, and computation | What data exists, is it reliable, and can we use it? |
| Methods, assumptions, validation, and limitations | Why analyze it this way, which judgments hold, and what must be rechecked? |
These layers complement one another. The first gives an agent data to analyze. The second gives it a basis for continuing work the team has already done.
The judgments in analysts’ heads need infrastructure too
Data infrastructure often brings data engineering to mind. If the goal is for AI to continue team analysis, analysts and data scientists also need to preserve the methods and judgments their work produces.
That includes more than metric definitions:
- Which methods suit this class of question, and which checks cannot be skipped?
- Why was an anomaly interpreted as a business change rather than dirty data?
- Which hypotheses were tested and rejected? What limitations remain?
More schema documentation cannot supply these answers. They often emerge while investigating the data and discussing it with the business.
A report saying “returning-user retention improved” transfers less than a record explaining why this observation window was chosen, which confounders were excluded, and why the improvement still cannot be attributed to the new release.
I think this analytical context belongs in the team’s infrastructure. It affects whether the next person or agent can continue the work correctly. Preserving it does not make it permanently true. Preserving the conditions is precisely what lets us decide which parts to reuse and which to validate again.
Let analysis assets form during the work
Who will organize all this?
If every analysis requires another half-day of documentation afterward, “building analysis infrastructure” will probably sink into the backlog.
This is why I put organizing knowledge during analysis into Lantide’s workflow. Working with an agent already involves rounds of discussion, execution, and confirmation. Once we align on a method, reject a hypothesis, or confirm an anomaly, the agent can organize it while the context is fresh, and a person can review it.
Automatic document writing alone is not enough. Questions, Plans, execution evidence, Reports, and confirmed knowledge must connect. Knowledge needs an appropriate category and a structured place to live.
Next time, the agent should be able to follow a report back to its question, method, evidence, and applicable conditions, then validate against current data. Finding an old report with a similar conclusion is only a start.
If every handoff still requires the original author to explain everything again, we may mainly be accumulating documents rather than useful analysis assets.
My criterion is whether the work can still be understood, validated, and continued by another person, another agent, or at another time. Does it reduce the cost of future analysis?
A practical check is to take a report from three months ago: can someone else explain why a channel was excluded, and which conditions need to be checked before rerunning it?
Further reading
Try this way of working in Lantide Data: Download Lantide Data.
Previous：Why Does AI Still Use the Wrong Definition After I Save an Analysis as a Skill? · Next：How Can a Team’s Analysis Become Experience That People and Agents Can Reuse?