The AI layer
We navigate the AI so the answer is worth having
AI is easy to use and hard to use well
Anyone can ask an AI to size a market. Getting an answer you would put your name on takes knowing what to ask, what to feed it and what to throw away — and it has to be right every time, not on average.
Each press is one job, already worked out
Not one assistant asked to do everything — separate calls, each with its own written prompt and its own rules. We do the hard thinking about how to ask: written down, graded against real answers, and kept up to date as we improve it — rather than typed fresh by whoever is at the keyboard.
Applied inconsistently, it produces figures nobody can compare
Three people in one company, three prompts, three confident paragraphs. The numbers differ because the prompts differed, not because the markets did — which is worse than having no figures, because it looks like evidence.
The same call, the same way, for everybody
Everyone at your company presses the same prompt — not one tuned per person, per team or per deal. It improves over time, in versions everybody moves to together, so two analyses differ because the markets differ rather than because someone phrased it better. That is the only condition under which comparing them means anything.
It will hallucinate, and it will sound certain
A general model invents a figure as readily as it recalls one, and cites a syndicated report mill with the confidence of a statistics office. The failure looks exactly like the success.
It never touches the arithmetic
Every figure is computed by an engine you can read. Sources are screened before you see them and the report mills are named rather than quietly dropped. The AI proposes; nothing lands in your analysis until you press.
The same question, asked properly
You can do all of this in a chat window, and plenty of people do. Here is what actually differs when the answer has to survive a meeting.
| A chat window | Here | |
|---|---|---|
| The number | The model produces it. Ask twice and it moves. | An engine you can read computes it. The AI never multiplies anything. |
| The sources | Cites a report mill as confidently as a statistics office. | Screened before you see them, and the mills are named rather than dropped. |
| Between people | Everyone prompts differently, so no two answers compare. | The same call, the same way — differences come from the markets. |
| Six weeks later | A chat log, if anyone kept it. | Every kept answer sits on the analysis with its source and its date. |
| What it knows | Whatever you remembered to paste in. | Composed from the analysis, and told what you have already rejected. |
| What you send | Whatever was in the window. | Listed row by row before it goes — or run the same prompt in your own AI. |
What that looks like in the product
Each of these is its own call, with its own written prompt. They are not one assistant wearing different labels — some read the web and take the better part of a minute because the answer is only worth having if it is sourced; some reason without searching on a stronger model because the judgement is the hard part; most return in seconds because they answer one small, specific question.
- Find published references for your market
- Work out how this market can be sized
- Suggest the figures a method needs
- Draft candidate customer needs
- Review a factor's analysis against your rating
- Sharpen how you describe your product and your market
A prompt is a file, and improvements ship as a new version
Improving one means writing the next version rather than editing the last, so a result measured months ago still means something today — and an answer kept in your analysis can be read against the exact words that produced it.
One prompt per job, not one per product
A written prompt per field, per sizing method, per factor — each registered by hand, in that tool's own vocabulary, rather than one general instruction reused everywhere. Where one has not been written yet the button is simply not there, rather than falling back on something generic.
Sources are screened before you see them
A source needs a named analyst, a stated method and a date to earn the top badge. The syndicated mills that restate one another are set aside — and named, so you meet them here rather than later.
Nothing is sent until you press
Every call opens a screen listing exactly what will be sent, row by row, plus what never is. Or copy the same prompt, run it in your own AI and paste the answer back — that path gets the identical prompt, not a lesser one.
What it will not do, on purpose
Every one of these is a rule in the code, not a habit of the model.
- Produce or change a number you rely on. The engine computes; the AI proposes inputs.
- Put a figure into your analysis. Every value lands behind a press you make.
- Rate a factor. The recommendation you see is the tool's own, computed from your analysis with the facts behind it on show — the AI reviews the rating you settled on, and never offers a level itself.
- Score a need it drafted, or attach evidence it invented.
- Scale a figure to your market behind your back. A published figure that covers more or less than your market is labelled as such when it is found, and the read-across is a factor you set, beside the reason you would defend it.
- Send anything before you have seen the list of what is being sent.
You see exactly what leaves the page
Not a policy — a screen. Before any AI action runs, the tool lists precisely what would be sent and what stays behind, and hands you the prompt to run in your own AI instead. This is that screen.
Your work stays yours, and so does the choice to send it
Nothing leaves your browser until you have seen the list of what is being sent and pressed to send it. And if you would rather send us nothing at all, every prompt here is copyable — run it in your own AI and paste the answer back. That path gets the identical prompt, not a lesser one.