Session 1 · Before installation
What to expect from an AI agent for data analytics
Realistic expectations before the agent is installed
This session does not explain how to use the tool, that comes later. It explains what is reasonable to expect from it and what is not, because most disappointments with AI do not come from the technology. They come from measuring it against the wrong benchmark.
AI is not a faster search engine
The first question almost everyone asks an AI agent is one they already knew how to answer. And that is where the tool loses.
- For a single fact, the old way wins. "What is the price of this stock right now?" is ten times faster in a search engine. And "how much did we sell to this customer last month?": if you already know which dashboard view has it, you will get it five times faster yourself.
- And that is fine. It is not a failure of the agent. That question is simply not what it is for.
- The problem is the benchmark. If the test is "does it answer faster than what I already do?", the conclusion will always be that it does not work.
- The right benchmark is a different one: is it giving me something that nobody in the company is doing today?
A simple question to the agent can take longer than opening the dashboard yourself. That is not a failure: it is the price of an answer that has been analyzed and checked.
The value of this tool is not in quickly answering what you already know how to ask. It is in answering what nobody is asking today.
How long it takes, and why
It helps to have these ranges in mind before installation, so nothing comes as a surprise.
Why it takes that long
- It does not just fire the query. It interprets the request, drafts a work plan, checks its own plan, runs it, and confirms the number a second way before answering.
- That obsession with accuracy is deliberate, and it is ours: we would rather it take longer than hallucinate a number. If it cannot find something, it questions what it did and tries another route, and that is where a good share of the time goes.
- Hardware matters. It runs faster on a Mac. It runs faster on a recent PC. On older machines the same analysis can take twice as long. This is measurable: we run the same test on every new installation.
It is like asking a trusted colleague for an analysis: "cross this against that and tell me what you see". They do not answer you in the hallway. They go away, think it through, cross the information, and come back an hour or a day later with an answer worth reading.
When to open the dashboard yourself, and when to ask the agent
The most useful practical rule of the whole session. Follow it and you will not be disappointed.
Open the dashboard yourself
- You need one number and you know exactly where it is
- You need it right now, in the middle of a meeting
- It is a lookup you do every day and you know the path by heart
- You only want to see the data, not interpret it
Ask the agent
- You need to cross the dashboard with a spreadsheet, with emails, or with files shared by other teams
- You need interpretation: what moved, why it matters, what to look into
- You want to compare against target and project where the period will close
- It is a multi-step analysis that repeats every week or every month
- You want a deliverable: a report, a spreadsheet, a presentation
- The question starts with "why" or "what should we", not with "how much"
Two analogies: electricity and the PC
These technologies change every week, literally. Frequent updates will be normal. But the real paradigm shift is not about speed, it is about process, and it has happened before.
Electricity in the factory
When electricity reached factories, the first thing they did was remove the horse that turned the central shaft and put a motor in its place. Same layout, same logic, a bit faster. The real leap came later, when someone understood that every machine could have its own built-in motor, and the whole plant was redesigned.
The PC in the office
It was sold as "the same thing, only faster". It ended up replacing the engineer's drafting table and remaking the way entire firms worked.
The first version of any technology imitates the old process. The value shows up when the process is redesigned. Using an AI agent to do what you already do, only faster, is putting the motor where the horse used to be.
Where the value is, concretely
Two fronts, in this order of importance.
1 · New capabilities
Analysis, cross-referencing and interpretation that the organization simply was not doing. Not for lack of data, but because nobody had the time or the method. Most of the value is here.
2 · Automating the routine
Repetitive tasks that make no sense for a person to build by hand every Monday. Valuable, but it is the second front, not the first.
Two real cases, anonymized
- A wholesaler operating several legal entities. They started with: "give me sales for the last N months, turnover by product, and tell me where I am under-buying and where I am tying up working capital in slow-moving inventory". That analysis did not exist in the company before. It was built over several sessions, refining the instructions.
- A distributor in the fashion industry. They want to capture the reasoning and years of experience of the person who analyzes purchasing needs. They cross 12 months of sales, max levels, reorder points, open purchase orders and stockouts by supplier. The goal: invest their capital in the SKUs with the highest inventory turns and the shortest replenishment lead time.
The one who gets the most value is not the one with the best data: it is the one who can explain the problem, and the possible ways to solve it, most clearly. AI does not guess your business rules: it has to be trained, and it goes through several rounds of iteration and adjustment.
The best starting point: the report nobody interprets
The Monday meeting is 100% informative and zero interpretive.
The same spreadsheet gets presented every week, with targets and sales. The data is there. The reading of the data is not. And the same thing happens in the finance meeting, in credit and collections, and in logistics.
How to spot a good first use case
- It is recurring: every Monday, every month-end close
- The input already exists and someone builds it by hand
- It is purely descriptive: tables and running totals, no conclusions
- It is neither trivial nor huge: it can be broken into instructions in one session
What the agent adds on top of that same report
How a use case gets built
This is not delivered turnkey in one session. It is built through iterations and consulting, and that is how it works.
- Break the problem down into explicit instructions. Business judgment is dictated, not guessed. Every ambiguous term ("active customer", "one-time sale", "focus product", "obsolete line") has to be defined.
- Give it context and evidence, not just the data. Targets for the full year, not just the month. The history. The priorities the company has already set. The more factual data, like target versus actual, the richer the analysis.
- Structure and segmentation matter. If the information is not grouped (regions, segments, product families, owners), the report comes out unreadable. Asking for "one slide per sales rep" with forty reps is useless. Segmentation is defined before asking for the analysis.
- It is not limited to one source. The agent runs inside a folder on your computer: you can give it the dashboard, spreadsheets, access to Google Drive folders, files exported from the ERP, even your email and meeting transcripts.
- Manual today, automatic tomorrow. What gets loaded by hand today is later integrated into the dashboard. It is continuous improvement, not a one-off delivery.
If an operational process takes two or three days to update the system, the number will be wrong, with or without AI. Defining the cutoff date and knowing which field can be trusted is part of the work, and it is the company's work, not the tool's.
What we need from you
Before installing anything on your computers, we want to arrive at the first hands-on session with one of your analyses already run. We run it internally and present it to you. For that we need three things.
With that, we run the first analysis on our side and present it to you before touching your machines. So when installation and the first hands-on session arrive, the example you train with is not a demo case: it is your own report, already moving.
What comes after