The analyst's week - InsightViz Blog

Ask a BI analyst what they spend their week on and you get a version of the same list: variations of existing dashboards, the same question from a different team, reconciling two numbers that should match, and — near the end, if nothing caught fire — the analysis they were hired for.

The useful question is not whether AI helps. It is which items, because the honest answer is not all of them.

A worked example

Dana, one of two BI analysts at a 900-person distributor — a Monday-to-Friday queue

Fourteen items came in. This is what they were and what happened to each. The pattern is more interesting than the total.

The request	Time	What it really was
“Sales by region but split by segment”	35 min	A filter someone could not apply
“Same but for Q3 instead of Q4”	20 min	A date filter
“What was our margin on the Atlas range last month?”	40 min	One number
“Same chart, top 10 instead of top 20”	15 min	A limit
“Can I get this as a spreadsheet?”	20 min	An export
“Ops and finance disagree on units shipped”	3 hrs	Metric drift — real work, once
“Dashboard for the new logistics team”	1.5 days	Real, but mostly assembly
“Why did the North fall off in week 34?”	4 hrs	Analysis — the job
Six more variations of the first five	2.5 hrs	Filters, all of them

Illustrative composite. The figures are internally consistent and every step is one the product supports today.

Nine of fourteen items — a little over five hours — were filters, limits and exports. One item was the job she was hired for, and it happened on Thursday afternoon because nothing had caught fire.

The problem was never that analysts are slow. It is that the organisation has no way to ask a small question without going through one.

The queue, sorted by what it actually is

Request type	Share of Dana’s week	Who should handle it
A filter, a limit, a date range	~5 hrs	Nobody — the reader, in the moment
One number, once	~1 hr	The assistant
Two numbers disagree	3 hrs	A person, once — then never again
Build a dashboard for X	1.5 days	A draft in minutes, a person to finish it
Why did this happen?	4 hrs	An analyst. This is the job.
A new source or join	0 this week	An analyst. Always.

What the assistant takes, and what it leaves

A worked example

The same fourteen items, three months later

Same team, same requesters, same Dana. The model now has the logistics tables and the definitions were settled in month one.

What changed

Items 1–5, 9
Never arrive
Nine items — the filters, the limit, the date range, the export, the one-number question — are answered by the requesters themselves, in the dashboard or by asking. Dana does not hear about them. Five hours, gone from the queue.
Item 6
Happens once, then never
Ops and finance disagreed on units shipped because each had its own definition. Settling it took three hours in month one. It is now one metric with an owner, so the disagreement cannot recur — which is why it is absent from this week rather than reduced.
Item 7
1.5 days becomes 2 hours
She describes the logistics dashboard, gets eight laid-out panels back, and spends the two hours on the part that needed judgement — which grain, which comparison, what to leave off.
Item 8
Unchanged, and now it has room
“Why did the North fall off in week 34?” still takes four hours and still needs someone who knows that a large customer moved to fortnightly ordering. It now happens on Tuesday rather than Thursday afternoon.
New
Work that did not exist before
Two hours on the catalogue — descriptions and synonyms for fields the refusal log showed people asking about in words the model did not recognise.

Illustrative composite. The figures are internally consistent and every step is one the product supports today.

9 of 14items that stop reaching the queue

~5 hrsa week returned

1.5 days → 2 hrsfor a dashboard, because the draft is free

Unchangedthe four hours of actual analysis

Two ways this makes the job worse

Both are real, both are avoidable, and both come from getting the order wrong.

THE FAILURE MODEAn assistant on an unmodelled estate

Point it at raw tables and it answers confidently and inconsistently. Dana’s queue does not shrink — it changes shape, from “build me this” to “check whether this answer is right”, which is slower, less satisfying, and never ends.

THE DESIGNAn assistant on an agreed model

It can only name metrics and dimensions a person defined, using joins a person declared. Answers are checkable in seconds because each shows its working, so nobody re-derives them.

The second failure is subtler and nobody predicts it

If the assistant is good, people stop asking for dashboards — including for numbers they check every morning. A figure someone looks at daily belongs on a page, not behind a question that has to be re-asked and re-read each time. Watch for teams whose “dashboard” is a saved question they run manually; that is a page waiting to be built.

The work that grows instead

This is the actual change to the job, and most AI-for-BI pitches skip it. If the assistant is bounded by the model, the model is now the product, and maintaining it is the highest-leverage thing Dana does.

The model editor listing dimensions and metrics with business names, types, number formats, descriptions and synonyms — **Dana’s two hours went here.** The refusal log showed people asking about “territory” and “patch” when the field is called Region. Adding two synonyms retired a class of question that had been failing silently.

What she maintains	Why it matters	How she knows what to fix
Descriptions	The difference between the assistant picking the right metric and a plausible neighbour.	Answers that were technically right and not what was asked
Synonyms	People ask in their own words. Every synonym retires a class of failure.	The refusal log — a prioritised list written by users without being asked
Declared joins	The assistant cannot pick a wrong join, but it also cannot notice one. That error moved to her.	Reconciliation differences
Hierarchies	Region → category → product is why Marcus in the Central region can drill without asking.	Drill requests that hit a dead end

Building, when building is still the answer

For item 7, the draft is no longer the work.

The dashboard builder with a plain-English brief typed in and an eight-panel dashboard laid out in the preview beside it — **Under a minute for the draft.** Every field it proposes is checked against the model, so an invented one is dropped and named rather than becoming a panel that fails the first time the logistics team opens it. Nothing saves until she presses save.

The builder's design tab showing the field picker with dimensions and metrics on the left and a panel definition being assembled — **Both tabs write the same document.** She takes over the draft by hand for the two panels that needed judgement, and leaves the other six alone.

Exploration is still the job

Item 8 — why did the North fall off — is not going anywhere and should not. It requires knowing the business, the data, and what the answer will be used for.

The Explore workspace with the field list on the left and the chart and result table on the right — **Chart and table together, always.** The table is what an analyst checks the chart against, and any result keeps its “how was this computed” view — including which access rules applied.

flowchart TD
    Q["A request arrives"] --> T{"Can the model
already express it?"}
    T -->|"yes, small"| A["The requester answers it
filter, drill, or ask"]
    T -->|"yes, a page"| B["Draft it, then finish it
hours, not days"]
    T -->|"no"| C["An analyst:
new source, new join,
new definition"]
    C --> M["The model grows"]
    M -.->|"next time"| T
    style A fill:#e6f7f2,stroke:#9fd8c9
    style C fill:#eef1ff,stroke:#c7cdf2
    style M fill:#fff4e6,stroke:#f0cb9a

The loop that matters: every genuinely new request ends with the model being more capable, so that class of question routes itself next time. An estate where this loop does not close is one where the queue never shrinks.

Three honest caveats

The model has to exist first. On an estate with no agreed definitions, the assistant industrialises the drift. Model first, assistant second — in that order, or Dana spends her week checking answers.
Someone owns the catalogue. Descriptions and synonyms decay as the business changes. Unowned, the assistant quietly gets worse and nobody can say when it started.
Is it a better job? Less assembling, more defining. Most analysts say yes. The ones who say no usually liked the dashboard-building specifically, which is worth knowing before you roll it out rather than after.

The mechanism that bounds the assistant → · Why item 6 happened at all →

← All posts Request a demo