In August the National Data Administration announced a second batch of data annotation pilots across 32 cities. Data annotation has been formally promoted from grey-market outsourcing to infrastructure for a national AI strategy. The opportunity is real. But a veteran with more than a decade in the trade said something counter-intuitive: what I most dread hearing right now is “Boss Liu, I happen to have a project on hand — can you take it?”
First, understand this: why is the state suddenly spending money?
The one-sentence version: the state has begun systematically recruiting people at scale to teach AI to recognise things, and the effort has been written into national strategy.
- Data annotation means being AI's teacher — turning raw data into a form AI can read. Face unlock, car navigation, asking an AI to write an article: behind all of it someone labelled which part is an eye, which line is a lane marking, what counts as good and what counts as wrong.
- People used to see it as assembly-line drudgery, exhausting and undignified. But once the National Data Administration made this move, its nature changed.
- The logic underneath: compute, data and algorithms are becoming new factors of production, like water, electricity and coal. National intelligent computing capacity has reached 2.45 million PFLOPS, the Big Data Expo set the tone at 500 trillion tokens, and data is being commoditised. For AI to be genuinely smart, what matters is the quality of the data fed in — data annotation is AI's compulsory education.
Deploying pilots in 32 cities essentially turns the business of teaching AI from scattered private activity into publicly managed national infrastructure.
So the opportunity is real. Policy dividends, city pilots, tilted resources — this is an officially stamped entry-level AI post.
But that is exactly where the problem starts.
When an opportunity lands, the most dangerous move is rushing in the moment you hear about it
I spent years in operations management, so I know the value of professional data very well — this round of AI progress rests entirely on enormous volumes of labelled professional data. So when the state elevated data annotation to a strategic level, my first reaction was excitement.
But after reading that article by a veteran of more than a decade, I calmed down. He made a point I agree with completely:
The biggest risk an opportunity brings is not a shortage of chances; it is that too many people, on hearing the word opportunity, start recruiting, renting premises and buying computers before they have worked out whether the project makes money.
That veteran has done data annotation for over ten years, and what he fears most now is hearing “if there is a project, take it”. Why?
1. Winning the project may be where you start losing money
Many newcomers imagine data annotation is simple: the client hands over a project, the company hires annotators, the workers do the work, you pocket the spread.
Say a client gives you a 100-person project at RMB 300 per person per day and you pay wages of RMB 150. Run the numbers: RMB 150 net per person per day, RMB 15,000 a day across 100 people, several hundred thousand a month.
That is the arithmetic many people do before immediately looking for premises, buying computers and hiring.
But the one number you must never trust in a data annotation project is that spreadsheet.
Because the RMB 300 the client pays you is only revenue, and RMB 150 is nowhere near your full cost. You still have to ask:
- How many days of training? Does the training period produce any output?
- Who pays for team leads? Who pays for QA? How is the project manager costed?
- Recruiting costs, computers and premises, employee social insurance, tax, business development — all of it is cost
- What happens to the workstation left empty when someone quits? Does rework count as paid time?
- Most important: when does the client settle? Wages go out on time every month, but the client may only inspect and settle after three months. On a 100-person project, how much are you prepared to float?
A project quoted with a margin and a project that actually makes money are two different things.
2. Having a project is not the same as having a business
When judging a project, your first question should not be how much it pays but whether the business model holds up.
- A long-term project — how long is long-term? Three months, six months, a year? Are the 100 people already confirmed, or is it expected to grow to 100?
- Is the client the end customer, or are there two or three layers of suppliers in between? Who carries out acceptance? Who ultimately pays?
- If requirements change midway, who bears the rework cost?
Until those questions are answered, the phrase a 100-person project means nothing.
Some projects are even most profitable not at 100 people but at 30 — the team is stable, the workers are experienced, management overhead is low. By the time you really scale to 100 you have more project managers, more QA, heavier recruiting pressure, higher attrition and ever more frequent training, and you discover that revenue grew while margin thinned.
What a data annotation company should really chase is never how many projects it holds, but how many of the projects it holds actually suit it.
3. The number most likely to fool you is gross margin
Many people who have just started in data annotation love to calculate one figure: the headcount spread. The client pays RMB 300 a day, the worker gets RMB 150, so they assume RMB 150 of profit per day.
That is one of the most dangerous calculations in this industry. Real project profit has to account for at least five lines:
1. Direct labour cost: beyond annotators there are team leads, QA, project managers and trainers
2. Non-productive hours: training, meetings, waiting for data, system failures, requirement changes — the client may not pay for these, but wages still go out
3. Organisational cost: recruiting, administration, finance, premises, utilities, computers, network, depreciation
4. Delivery cost: rework rate, sampling rate, discarded data, quality remediation
5. Cost of capital: the wages you pay today may only be reimbursed by the client three months from now
So very often a project looks like it carries 30% gross margin, and the net profit that actually lands in the company account is a few points. One more round of rework, one delayed acceptance, and the profit is gone — or the project turns into a loss outright.
The most important sheet for a data annotation company is not the project schedule; it is the project profit and loss statement. Before work starts you should already know: what output volume breaks even? Below what utilisation rate do you lose money? At what rework rate must you raise the alarm? How much further can the price fall before you must walk away?
If you do not know those numbers, you are not running a project; you are gambling.
4. Finishing the project does not mean the money is earned
Data annotation has another problem that is badly underestimated: acceptance.
In many industries handing over the product essentially confirms revenue. Data annotation is different. You may label a million records, but a million records does not mean a million get settled — the client samples and finds problems, work is redone, standards shift, the algorithm team raises new requirements. In the end you may have produced a million records, 800,000 pass acceptance, and only 700,000 are finally settled, while the labour, premises and management costs of that first million are paid in full.
So an experienced data annotation project manager cannot just watch progress; they must understand quality, labour efficiency, cost and acceptance, and a little commercial negotiation besides. Whether a project ultimately makes money is decided jointly by business, production, quality, finance and the client's acceptance.
5. The longer you do this, the clearer it is that saying no is itself a skill
People new to the trade think more projects are always better. When a client asks whether you can do it, the first answer is always yes — no people, recruit; no premises, find some; no experience, learn on the job.
But companies that have been at it a while discover that not all money should be earned.
- What projects suit a team with school resources?
- What projects suit a team of experienced annotators?
- Which projects must a team with labour but no project manager never touch?
- With only RMB 200,000 of capital, how large a project can you carry at most?
- Should a team that is good at 2D annotation charge straight into embodied intelligence?
There is no single answer to these. A data annotation company's real competitiveness is not “whatever others can do, I can do”, but knowing what you have and knowing what you should not do.
So when people ask how to get into data annotation, the answer from industry veterans is completely different from ten years ago — do not look for a project first; first get three things straight:
What resources do you have? Which projects suit those resources? How do you run the numbers on a project?
Only then talk about taking work on.
For ordinary people, how do you get on board with your eyes open?
The opportunity is real, but board it clear-headed. Three tiers to consider:
Tier one: walk in with no threshold. In the pilot cities, annotation bases and parks are hiring annotators. No degree and no code required; patience and care are enough. This is the most realistic way for an ordinary person to enter the AI industry, and it suits people who want to get in early and build experience.
Tier two: climb towards specialisation and compliance. Mechanical annotation alone has a low ceiling, but if you understand an industry — healthcare, law, finance — you become valuable. Specialised data annotation needs people who genuinely know the domain, and that is precisely the moat that veterans of traditional industries are hardest to replace in. Data compliance and annotation quality review are both well-paid work requiring professional judgement.
Tier three: organise, but only after thinking it through. If you really want to be the contractor taking orders, do not rush to rent premises and hire. Learn to judge a project and to work out those five lines first. Remember the veteran's line: more important than finding a project is knowing, once you have it, whether you should really take it on.
The opportunity has arrived. But the real dividend belongs to those who see the chance and can also run the numbers.
#AI #DataAnnotation #SideIncome #Opportunity #Startup #Money