About

The gate

An “AI training business” is either GPU capex or a laptop and a course.

Guessing between those two produces a report that is not merely wrong but confidently wrong, and a confidently wrong number is worse than no number — it gets budgeted against. So the run refuses to start until the scope is pinned. If you decline to narrow it, the agent names two concrete scenarios, says it is researching both thinly, and proceeds on the record.

The research run, in six steps

They happen in this order, and each writes its result to the dossier rail before the next begins. That is why the rail fills while you watch, and why a run interrupted at step four still leaves you steps one through three.

1. 01

   Scope gate

   Four blanks — what, to whom, in which city, and how many by month twelve. Nothing is researched until they are filled. The run stops here and asks; it does not guess and continue.

2. 02

   Landscape sweep

   What the category looks like in the named city: who operates, at what price, under which licence. Read before anything is costed, so the model is built against a real market.

3. 03

   Cost model

   One-off capex and monthly opex, line by line, in Rupiah. PT versus PT PMA, UMP/UMK, BPJS, THR, PPN, OSS/NIB, leases paid years upfront — the places an Indonesian model diverges from a template.

4. 04

   Competitor teardown

   The operators already doing it, pulled through one shared extraction schema. A different schema per competitor produces a comparison table that cannot be compared, so there is only one.

5. 05

   Strategy

   What follows from steps three and four — where the margin is, what entry actually costs, and which of the incumbents' positions is unoccupied. Strategy is a conclusion here, not an opening move.

6. 06

   Assumption register

   Every number the run could not retrieve, listed as an assumption beside the figure it drives. The report ends with its own weakest points rather than burying them.

The marketing run, in eight steps

It can start from nothing, or from a finished research run — the desk hands over the competitor prices and the unit margin with their sources intact, so the marketing agent does not have to re-derive what has already been retrieved.

It has a hard budget of thirty retrievals per run, and it is told so. A plan finished honestly on what could be sourced beats one that spent an hour reading around the subject.

1. 01

   Brief gate

   Seven blanks, and price is the load-bearing one — it decides whether paid acquisition can ever pay back. Also the horizons you want planned to, and the budget you actually have.

2. 02

   Feedback sweep

   What customers said, clustered by cause rather than by star rating. Paste it, enter it a line at a time, or let the agent read the reviews off Tokopedia, Shopee, Google Maps or the Play Store itself.

3. 03

   Audience

   Two or three segments, ranked, each with the trigger that happens immediately before someone buys. A segment without a trigger is a demographic, and demographics do not buy anything.

4. 04

   Positioning

   The promise, the proof, and the specific complaint it is built to defeat — in the buyer’s own vocabulary, pulled from the feedback rather than upgraded into marketing language.

5. 05

   Channels

   Every rate retrieved and dated: CPMs, marketplace commissions, delivery-platform takes, real influencer fees. Each one carries whether its cost per customer sits below your unit margin.

6. 06

   Short term

   The near plan, 90 days by default, week by week, with money allocated against the budget and the metric that says stop. Kill criteria are set now, while they are cheap.

7. 07

   Long term

   Twelve months by default. Not the short plan with bigger numbers — retention, the asset that gets cheaper to run, and the expansion with the condition that unlocks it.

8. 08

   Assumption register

   Every rate it could not source, with what breaks if the guess is wrong.

If the feedback says the product is the problem, it says the product is the problem.

There are three reasons nobody is buying — the thing itself, what you said about it, or that too few people heard — and they need completely different answers. Conflating them is the usual way a marketing budget gets spent on a problem it cannot touch. The agent names which one you have, in one sentence, before it proposes anything.

How a figure is marked

Colour here is the rule set, not decoration. Every number arrives carrying how it was obtained, and neither agent will state a price it did not retrieve from a page during the run. A CPM is a price. So is an influencer’s fee and a marketplace commission — the sourcing rules do not relax because the thing being bought is attention.

Sourced

Retrieved from a vendor or operator page during this run, with the URL and retrieval date attached.

Derived

Calculated from sourced figures. The arithmetic is on the page and in the spreadsheet.

Assumed

Nothing was found. The figure is a stated assumption, and it is carried into the register at the final step.

Risk

The finding that could end the venture, or sink the plan. It is delivered above the report, before anything encouraging.

What you leave with

• From the research desk: a written report, verdict and headline risk first, and the cost model as a spreadsheet with live formulas — change an assumption and the rest moves.
• From the marketing agent: a written plan, with the channel table, the short-term schedule and the long-term arc.
• From both: the source list, every page the run read, with the date it was read.

What it is not

It is not an oracle and it is not advice. It reads public pages and does arithmetic on what it finds; prices move, and a page that was accurate on its retrieval date may not be accurate when you sign. Push back on any number in the chat — that is what the conversation is for — and read the assumption register before you commit money to anything inside it.