part-1-four-architectures

Agents are a powerful tool for data analysis, so long as you set them up for success.

In this series I want to discuss four architectures for agentic data analysis: agentic employees, agentic data structuring, conversational analytics, and the one I think is the best case of all: pure software.

How do you decide which one fits your problem? It depends on the type of data analysis you are doing.

Not to oversimplify, but data analysis has two parts, the data and the analysis, and you need to understand both before you can pick the right architecture.

That's the question for this post: how does the type of data analysis determine which architecture to choose? The short version is that the more standardized the data and the more recurring the analysis, the easier the agent's job gets, and the better it performs.

Let's start by understanding the data.

Ad Hoc vs Standardized Data

The first question to ask of your data is whether the fields look the same every time. If they don't, the data is ad hoc: it only exists for some instances, and it looks different each time. If they do, the data is standardized: the same fields for every instance.

Take an investment team deciding which stock to buy. Some data exists in a consistent format for every company: the price history, the balance sheet, the income statement. The values differ from company to company, but the fields are predictable and you can pull them from a single source. That's standardized data.

Other data works differently. Suppose you want to know how well a company manages its inventory. There's no standard field for that. The evidence is scattered across earnings call transcripts, footnotes in filings, and industry reports, and it looks different for every company. That's ad hoc data.

Most problems have a mix of ad hoc and standardized data. The more of your data is standardized, the better.

Examples: Ad Hoc vs Standardized Data

Efficiency of Plants

Ad Hoc Data

A consultant's audit of one plant
Incident reports for the two lines that broke down
Photos of a faulty conveyor

Standardized Data

Sensor readings
Hourly output per line
Energy consumption

Investing in a Stock

Ad Hoc Data

A short-seller report on one target
Documents from a single lawsuit
Remarks the CEO made on one podcast

Standardized Data

Price history
Balance sheet
Income statement

Customer Support

Ad Hoc Data

An email thread from one angry key account
Notes from a call with a churned customer
A viral post about one outage

Standardized Data

Ticket volumes
Resolution times
CSAT scores

That covers the data. Now let's turn to the analysis you want to run on it.

One-Off vs Recurring Analysis

The question to ask of your analysis is whether you ask the same questions every time. If you don't, the analysis is one-off: a question triggered by one specific situation. If you do, the analysis is recurring: fixed questions you run every time, where only the answers change.

A recurring analysis is one that you run every time. In our financial scenario, some questions get asked of every company: are margins expanding or contracting? How does growth compare to competitors? Those are recurring analyses: the questions are fixed, only the answers change.

But some questions are one-offs, specific to a single company, say, investigating how a pending lawsuit might affect one target's flagship product. That question only exists because of something that happened to that company. That's one-off analysis.

Most problems have a mix of one-off and recurring analysis. The more of your analysis is recurring, the better.

Examples: One-Off vs Recurring Analysis

Efficiency of Plants

One-Off Analysis

Investigate why one plant's yield dropped last quarter

Recurring Analysis

Compute output-per-unit-energy for every plant, every month

Investing in a Stock

One-Off Analysis

How would the pending lawsuit affect one target's flagship product?

Recurring Analysis

Are margins expanding? How does growth compare to competitors?

Customer Support

One-Off Analysis

Figure out why our biggest account suddenly went quiet

Recurring Analysis

Weekly CSAT and resolution-time trends for every team

With both questions in hand, we can put them together.

Four Architectures

We asked two yes-or-no questions: one about the data, one about the analysis. That gives four possible scenarios, and they map neatly onto the four architectures, as shown in the four-quadrant diagram below. The horizontal axis represents the analysis and the vertical represents the data:

Each architecture is the right answer to one combination. In the rest of this series, we'll cover them one at a time, starting with the worst case (the agentic employee) and working toward the best case (pure software).

Agents for Data Analysis: The Four Architectures

Jackson Rudd

Ready to Transform Your Business with AI?