A data platform is the set of components that pull data out of every system a business runs (sales, accounting, inventory, support, marketing), store it in one place, clean and combine it, and make it available for reports and later for AI. For a company of 20 to 200 people the platform does not need to be large; it needs to be built in the right order, starting with the two or three reports that matter most, and operated so that the numbers stay trustworthy after the first month.
The problem a data platform actually solves
The symptom every growing business recognises: the monthly report takes a week. Someone exports from the sales software, someone else from accounting, a third person reconciles inventory, and the numbers never quite agree. Leaders ask a simple question ("which product line lost margin last quarter") and the honest answer is "give us three days".
The cause is not bad people; it is that each system holds its own version of the truth and nothing joins them. From spreadsheets to smart systems describes the first step out of this situation. A data platform is what that step grows into.
The 5 building blocks
1. Sources and connectors
Every system the business uses is a source: the sales or CRM software, accounting, warehouse, e-commerce marketplaces, support desk, advertising accounts, and yes, the spreadsheets that still hold real data. A connector pulls data from each on a schedule (nightly is enough for most businesses) without anyone exporting anything.
Rule: connect systems through their APIs or scheduled exports, never by having a person download files. The moment a human is in the loop, the pipeline breaks on the first holiday.
2. Storage: the warehouse
One database where all the data lands, kept as it arrived (raw) and then in cleaned form. For a business of this size, a managed cloud database costs from a few hundred thousand to a few million dong a month. Do not over-engineer this; the choice of database matters far less than what goes into it.
3. Transformation: the cleaning and joining layer
This is where the real work is. Customer names that differ between systems get matched. Product codes get standardised. Currencies and dates get normalised. Sales, costs and returns get joined into a single "revenue by product by month" table that everyone agrees on.
Every rule lives in code that is versioned and tested, not in someone's head or in a hidden spreadsheet column. When a number is questioned, the rule that produced it can be read.
4. Reporting
Dashboards and scheduled reports built on the cleaned tables. The first three should be the ones leadership asks for most often, built exactly to their questions. Resist building fifty dashboards; three that are trusted beat fifty that are ignored.
5. Operations
Monitoring that connectors ran, that row counts look normal, that yesterday's data arrived. Alerts when they did not. A named owner. Without this block the platform decays quietly: a connector breaks in March, nobody notices until the June report is wrong, and trust is gone.
The order to build in
For a 20 to 200 person company, the sequence that works:
- Pick 3 reports leadership actually needs. Write down the exact question each answers and who reads it.
- Connect only the sources those 3 reports need. Usually sales and accounting first. Marketplaces and marketing later.
- Build the cleaned tables for those reports, reconcile them against the old manual numbers for two months in parallel.
- Ship the 3 reports, replace the manual process, and only then add the next source.
- Set up operations from day one, not after the first failure.
This typically takes 6 to 10 weeks for the first three reports. Adding each new source afterwards takes 1 to 3 weeks.
Realistic budgets
- Build (first 3 reports, 2 to 4 sources): from around 80 to 250 million dong depending on how messy the sources are. Messiness, not volume, drives cost.
- Cloud costs: 1 to 5 million dong a month at this scale.
- Operations: monitoring, fixing broken connectors when a vendor changes an API, adding fields, answering "why does this number look off": 5 to 15 million dong a month if outsourced, or a fraction of one internal person's time. Siri9 covers this under software maintenance for the platforms it builds.
Compare this with the hidden cost of the current process: the hidden costs of manual operations puts numbers on the week-long report.
6 mistakes that waste the most money
- Starting with the technology choice. Weeks spent comparing databases while nobody has written down which three questions the platform must answer.
- Connecting everything at once. Twelve sources, no report, six months in. Connect what the first three reports need and nothing else.
- Cleaning rules in spreadsheets. A hidden column that fixes a product code is not a rule; it is a landmine.
- No parallel run. Switching off the manual report the day the dashboard launches. Run both for two months; the discrepancies are where the real cleaning rules hide.
- No owner. A platform without a named owner is abandoned within a year.
- Building for AI before the reports are right. If the monthly revenue number is not trusted, no model built on it will be either.
How the platform becomes the foundation for AI
Every AI use case a business considers (forecasting demand, scoring leads, spotting anomalies, letting an agent answer questions about the business) needs clean, joined, current data. That is exactly what blocks 1 to 3 produce. Businesses that build the platform first find their AI projects take weeks; businesses that skip it find every AI project starts with three months of data cleaning.
Two examples: using LLMs to analyse customer feedback needs tickets, reviews and surveys in one place; computer vision counting on a production line needs somewhere for the counts to land and be reconciled. Both are platform consumers.
The concepts of training data quality and drift that govern those projects are in machine learning fundamentals for business leaders.
Signs you are ready, and signs you are not
Ready: the monthly report takes more than two days; leaders regularly ask questions that take a week to answer; at least two systems hold overlapping data; someone is willing to own the numbers.
Not yet: the business runs on one system and a handful of spreadsheets that one person understands; reports take an afternoon. In that case a well-structured spreadsheet or a single business app is the right next step, not a platform. Siri9 will say so.
Frequently asked questions
Is this the same as a "data warehouse" or "BI"?
The warehouse is block 2, BI tools are block 4. A data platform is all five blocks working together, and blocks 3 and 5 are where most of the value and most of the neglect are.
Do we need a data engineer on staff?
Not for the build or the first year of operations at this scale. An outsourced team builds it and operates it; an internal owner (often the finance or operations lead) defines the questions and checks the numbers. Hire when the platform has grown past 10 sources or daily changes are needed.
How long until we see value?
The first three reports replace the manual process in 6 to 10 weeks. The week-long monthly report becomes a dashboard that is current every morning.
How do we start with Siri9?
Send us the three questions leadership most often asks and a list of the systems you use; we return a scope, a sequence and a fixed price within a few working days. Building is under business web apps, operations under software maintenance.