Working with AI on analysis · 5/7
A team having completed an analysis does not mean another member, or the next agent, can continue it.
One value of AI in analysis that I think is easy to underestimate is the possibility of preserving analytical experience at a cost teams can sustain.
Beyond data and domain knowledge
I divide the information a team needs for analysis into three categories:
| Category | What it preserves | Event analysis example and purpose |
|---|---|---|
| Data Context | Available data, where to query it, and field and metric definitions | Event tracking, participation metrics, schemas, and query entry points tell the agent how to access the data |
| Curated Domain Context | How the business works and the methods the team has organized | Event rules, user segments, terminology, and Skills tell the agent about established methods |
| Experiential Analysis Context | Methods actually tried, evidence, tradeoffs, and limitations | A case where eligibility checks changed the participation denominator helps the next analyst understand the earlier judgment |
The first two are easier to recognize as knowledge worth organizing. The third often survives only as a deliverable. The rest is scattered across code, chat, and analysts’ memories.
What a team knows and what a team has experienced provide different kinds of help.
The team did it once. Can someone else pick it up?
Imagine an event participation analysis. The team has event tracking, a participation-rate metric, and an established method. The agent has enough to start.
Last time, however, we first suspected that the rewards were not attractive enough. After several rounds of segmentation, we found that some new users had not unlocked the event entrance. We changed the denominator to separate people who could participate but did not from those who could not participate at all.
The delivered report may contain only the revised rate and denominator definition. That is reasonable: the report supports decisions, and a manager does not need to relive every investigative turn.
For the next analyst or agent, the experience in between is useful. When participation looks low, check eligibility before digging into reward design.
But low participation next time does not prove eligibility is the cause again. We need the comparisons and evidence from the earlier case to decide what is relevant now and what needs fresh validation.
Today, access to that experience often starts with “I remember someone did this before; ask them.” New colleagues must know whom to ask. Experienced colleagues repeatedly reconstruct the story. One analysis comes with unlimited after-sales support.
Some experiences are not ready to become general rules. They can still be saved as cases, with their conditions, evidence, and unsuccessful paths. If colleagues can find and understand them, they can begin from a position the team has already investigated.
Agents need the same thing. Schema and Skills tell an agent how to query data and apply methods. Actual cases can show how the team investigated, where it remained uncertain, and what evidence led it to narrow the question. Differences in the new situation can then prompt more specific questions.
This is what I want from an agent that analyzes like a team member: refer to the team’s experience, judge how to continue, and leave new findings behind.
AI changes the cost of preserving experience
We have always known we should leave comments, organize Notebooks, and record definitions and decisions.
During an investigation, stopping to write interrupts the work. After delivery, details have faded and the next deadline is waiting. The cost of finding the right SQL, assembling results, and writing the explanation is immediate; the benefit may arrive months later for someone else.
An agent changes one condition: it is already participating in the analysis.
Method changes, query outputs, and unsupported hypotheses have often already appeared in the collaboration. At the end of each phase, I can ask the agent to organize the decisions we just confirmed and connect the evidence, then review the record.
My work already includes deciding what is worth retaining and whether it is correct. Saving that judgment in a structured form at the same moment can avoid reconstructing the entire analysis later.
This is what I keep working toward in Lantide Data: preserve deliverables alongside the process that produced them, connecting questions, Plans, evidence, and judgments so people and agents have context to continue.
Unspoken judgments still need to be made explicit. Organized content still needs review. A beautifully written guess should not become the team’s inherited experience.
AI offers a chance to make this work affordable enough to sustain. Help team members continue one another’s work, then let agents participate through the same shared context.
Further reading
Try this way of working in Lantide Data: Download Lantide Data.
Previous：Is Data Infrastructure Enough for AI to Take Over a Team’s Analysis? · Next：Context Does Not Have to Be Full Before We Compact It