How Moncho has built a scalable market sizing pipeline
“How big is potato in Bangladesh?” can mean farm sales, factory sales, what households spend, or what customs records as imports. Those can all be true. They are still different questions.
Ask for a market size too early and you get a confident dollar with no shared meaning. Take potato in Bangladesh. One person may mean the value of the crop leaving farms. Another means what processors book as revenue. A third means what households spend in wet markets and shops. A fourth means import value at the border. Each reading can be useful. They answer different questions. This post explains how Moncho turns that mess into a reviewed number: lock the question, gather sources, clean them so they can be found again, pick a method the evidence can support, then publish only after a Moncho analyst decides. We call that last gate Human in the Loop (HITL).
Define, then evidence, then a reviewed size
We keep the story in order on purpose. First we say what we are measuring. Then we collect source material and clean it so the same fact can be reused for sizing and for Sherpa. Only after that do we pick a method, draft a number, and ask an analyst to publish.
That work hangs on Moncho’s market model first: how countries, landscapes, segments, needs, organizations, and products relate. Without that shared map, a size has nowhere stable to live. Learn how the market model works.
1 · Lock the question
2a · Where evidence comes from
2b · Clean so the fact can be found again
3 · Method → draft → analyst → published size
Sources are inputs. Cleaning is the process that makes those inputs findable later. A published size is a reviewed interpretation of that evidence, not a raw table dump.
“How big is the market?” is not yet measurable
Before anyone opens a spreadsheet, say what kind of economic activity the estimate represents. For potato, farm-gate value, processor revenue, household spending, and import value sit at different points on the same chain. If two people publish different dollars without naming that difference, they are usually arguing about scope, not about who is “right.”
So we pin six choices beside every serious size: where, what, who pays or produces, which money flow, when, and on what accounting basis. Without those, a formula is just algebra looking for a home.
Revenue is what a seller receives. Value added is what remains after intermediate inputs. Household spending is what buyers pay. A sizing method has to name which of those it is estimating, or the dollar will be read as something it is not.
From documents to a published size
Once the question is clear, the work becomes a pipeline you can inspect. AI agents do repeatable research labor. Tools hold catalogs, parsers, and calculators. A Moncho analyst locks scope, settles ambiguous facts, and decides what becomes official. That human gate is HITL.
For potato, decide whether you mean farm output, processing, retail spend, or imports, and for which year and place. This stops a broad label from hiding three different questions.
Agents propose which trade lines or product standards belong on which segments. Analysts confirm uncertain or high-stakes links before those mappings are trusted.
We pull from statistical offices, ministries, customs, surveys, yearbooks, and reputable global databases. The job is to extract useful tables, not to dump a whole PDF into chat.
This is not another source family. It is the process every source must pass. A bare “12,000” is useless until we know whether it is tonnes, firms, or currency, for which year and place, and where it was printed. When the automatic tag is unsure, an analyst settles that row instead of guessing.
We also keep a hard line between a sourced fact and a market-size estimate. One production volume can support several later calculations. A published size is a specific interpretation with method, assumptions, and review status.
Not every market uses the same equation. If a yearbook already prints industry sales, we cite reported sales. If we have tonnes and a sourced farm-gate price, we can build harvest value. If we only have customs, import value may be honest and still incomplete as a domestic market. The method follows the evidence.
Use reviewed facts first. Open a targeted search only for a named missing input, such as a unit price. Keep farm-gate, factory, and household readings separate instead of averaging them into one fake headline.
Check arithmetic, units, dates, coverage, and whether the result is plausible against broader official statistics. Gaps stay visible. Invented fees are refused.
A Moncho analyst reviews the amount, method, inputs, sources, and limits. Approve, ask for more evidence, or withhold. Only then does the reading become an official published size.
Most arrows move one way. The sanity step can send a draft back when an input is missing. Outside users do not decide what becomes live data.
Different evidence, different equations
Back to potato. If the yearbook already prints a sales total that matches your scope, start there. If you have harvest tonnes and a sourced farm-gate price, multiply them and label the result. If you only have customs, publish import value as import value. The method family should match what you actually have.
If the volume side is present and the price is missing, we hunt a sourced price. If the hunt fails, we hold the official dollar. We do not invent the fee.
Farm-gate output, factory sales, and household spending can describe different stages of the same market. Adding them may count the same product more than once. Keep each reading labeled, and combine only when the market definition justifies it.
Four agents do the repeatable work. Analysts still control the output.
Each card below is an agent: AI-assisted research labor with a defined job. Agents use tools. They do not replace a clear source trail or Human in the Loop.
Agent 01 · Harvest
Finds reports and pulls relevant tables from official portals, surveys, and documents.
Uses tools: source locators, PDF and table parsers.Agent 02 · Clean and tag
Normalizes units, years, definitions, and geography so each fact is searchable later.
Uses tools: validation rules, reusable evidence store.Agent 03 · Product mapping
Suggests links between trade or product classifications and Moncho segments for review.
Uses tools: product catalogs, segment lists.Agent 04 · Sizing draft
Builds a labeled working estimate from saved evidence, names missing inputs, and runs basic checks.
Uses tools: evidence retrieval, targeted search, calculators.Yearbooks, household surveys, industry and SME publications, and ministry portals.
- Example: Bangladesh Bureau of Statistics
- Example: national statistical yearbooks and household surveys
Import and export flows, read with their valuation basis and coverage limits in mind.
- Example: national customs / export promotion boards
- Example: OEC trade explorer
Useful when local evidence is thin. External or proxy inputs stay labeled as such.
- Example: FAOSTAT
- Example: World Bank Open Data
The cleaned observations that both market sizing and Sherpa read from the same place.
- Fed by Agent 02 (clean and tag).
- Not the same thing as a published market size.
A market size has to survive the next question, the next user, and the next year
Ask a generic chatbot “how big is potato in Bangladesh?” and you may get a confident dollar. Ask again tomorrow, or ask a colleague to check the work, and you often get a different number with a different story. Chat is good at improvisation. It is weak at keeping a durable, shared reading of a market.
That is why sizing runs as a pipeline. Facts are cleaned once and stored. Methods are chosen against evidence. Drafts stay labeled. A Moncho analyst publishes. Later sizing work, and Sherpa, can reuse that memory instead of rebuilding the market from scratch every time.
- Memory, not a transcript: cleaned facts and published sizes outlive any one chat session.
- Stateful across uses: the same potato reading can be inspected months later with the same scope, method, and sources.
- Shared across users: one analyst’s reviewed size becomes available to the next teammate, not locked in a private thread.
- Compounds over years: each sourced fact and each published card makes the next market cheaper to size.
- Honest limits: inventing a fee, or mixing farm-gate with household spend, is a process failure, not a clever answer.
- Quality under control: evals and human review check whether drafts still obey scope, method, and source rules when models or prompts change.
- Cost under control: we track LLM API use and token consumption across harvest, tagging, and draft steps, so spend stays visible and controllable.
When a number becomes official
A working estimate can still help an internal review. It is not the official Moncho size until an analyst has checked scope, method, inputs, sources, and limits. That gate is Human in the Loop, not a public vote.
We will not ship a count times an invented fee. Import value is not automatically the whole domestic market. When someone asks Sherpa to size a market, it should prefer a published card that matches the question. If the evidence cannot support a defensible figure, it should say so. That user journey is Part 3.
Market intelligence on Moncho is reviewed by Moncho analysts before publication.
