We built a freelance marketplace that takes no commission. This is what happened, including the parts that went differently than we expected. A piece that only reports the encouraging half is marketing rather than a report.
The fee problem was real, but it was only part of the marketplace problem
Marketplace commission is charged as a percentage of transaction value. Almost none of the underlying costs scale that way. Escrow, identity, discovery and reputation storage are flat or nearly flat. Only dispute resolution scales with anything, and disputes are rare.
That makes a percentage fee a position rent collected from the value two other people create. It is defended by contractual terms that stop users routing around it.
Our bet was simple: remove the commission, fund the platform through optional services, and compete on the thing incumbents cannot copy without destroying their own revenue.
The thesis about fees held. The incomplete part was what a marketplace has to do for users. Pricing matters, but pricing alone does not make a marketplace work.
The escrow economics held because the operating cost stayed low
The escrow economics were the technical risk. At fifteen percent you can afford manual reconciliation and a support team unpicking edge cases. At zero you cannot. The escrow has to be genuinely cheap to operate.
We modelled escrow as a proper state machine, a design where money can only move through defined conditions. We also used automated release conditions, daily automated reconciliation and payment rails chosen for cost rather than convenience.
That made the per-transaction operating cost small enough that optional revenue can carry it. This was the load-bearing engineering assumption, and it survived contact with users.
Disputes stayed rare because the product reduced the need for judgment
Disputes were the other technical risk. Structured agreements, milestones, an explicit revision allowance and timers with defined defaults kept human adjudication low enough that the cost is absorbable.
Designing disputes out is much cheaper than adjudicating them cheaply. That mattered more than building a low-cost dispute team.
Small jobs appeared once the percentage fee disappeared
Small jobs became viable. That was an unexpected consequence.
A twenty percent fee on a two hundred pound job is the difference between worth doing and not. On percentage platforms, the long tail of small fast work never forms. It formed on ours, and it is a larger share of volume than we projected.
The first users already understood the arithmetic
The right users self-selected. The people who joined first had done the arithmetic on what they were paying elsewhere.
They arrived with existing clients. That turned out to be the only acquisition channel we could afford.
We engineered trust first, while users were blocked by discovery
The most important thing we got wrong was solving the wrong problem first. We engineered trust: escrow, disputes, verification and reputation.
Users struggle with discovery, which is finding the right person at all. A zero commission does nothing for that.
We were slow to admit this because trust was the interesting engineering problem. Discovery is a hard, unglamorous ranking problem. The thesis was right about fees and incomplete about what a marketplace is for.
Free made people suspicious
A zero commission reads to many freelancers as a promotion that will end. It can also read as evidence that the money is being taken somewhere they cannot see.
That reaction is earned by the industry, and it cannot be argued away.
Publishing the revenue model explicitly helped. We underestimated that trust in a trust machine takes years, and no launch announcement shortcuts it.
Optional revenue takes longer because each paid service has to stand on its own
A percentage arrives automatically with every transaction. Optional services have to be chosen, which means each is a product that has to be genuinely good.
The optional services performed differently:
- Verification worked.
- Tooling worked.
- Placement is the one we are least comfortable with.
Placement is uncomfortable because paid visibility corrupts search quality if you let it. Keeping it few and clearly labelled means it earns less than it could.
Lower lifetime value changed how growth had to work
We cannot outspend anyone. A commission platform knows a user’s lifetime value is a percentage of everything they will ever earn, which justifies large acquisition spending.
Our lifetime value is a subscription and some services. Growth has to come from the product being obviously better for a specific person. That is slower and more durable, and considerably less comfortable to explain.
The commercial model depended on careful engineering around money
The commercial position rests entirely on operating cost, so it is useful to be specific about what had to be built.
Escrow ran as a nine-state machine
Escrow was built as a nine-state machine. The states were:
- Draft.
- Authorised.
- Captured.
- Held.
- Submitted.
- Releasable.
- Released.
- Settled.
- Refunded.
Disputed was a flag that suspends automatic transitions. Every transition has permitted actors and an audit record.
Three separations do most of the work:
- Authorised is separate from captured.
- Releasable is separate from released.
- Released is separate from settled.
Collapsing any of them is where platforms lose money.
Holds were calculated from transaction risk
Rather than a uniform policy, the release period is calculated per transaction. It uses payment method, both parties’ history, transaction size relative to the freelancer’s median, and whether acceptance was explicit.
Most transactions between established parties paying by transfer release in hours. That is only possible because the state machine reliably knows what condition everything is in.
Reconciliation ran every day
We ran daily automated reconciliation, checking the platform’s view against the provider’s view every day, with an alert on any drift.
Monthly manual reconciliation is how a small discrepancy becomes a large one. By the time you find it, the transactions are weeks old.
Every transition was idempotent
Idempotency throughout means repeated events do not create repeated financial effects. Every transition carries a key derived from the causing event.
We test by replaying whole days of webhook traffic shuffled and duplicated, asserting the resulting ledger is identical.
The failure this prevents is the duplicated payout, which is unrecoverable and not rare in systems not built for it.
None of that is novel engineering. It is ordinary care applied to money. It is the difference between an operating cost that a percentage must fund and one that optional services can.
The numbers showed which optional products people valued
Shape rather than absolutes, since we are not publishing full figures.
The per-transaction operating cost is dominated by payment processing, which is a pass-through we do not control. Everything we do control, escrow logic, reconciliation and storage, is a small fraction of that.
Dispute adjudication is the only line that could grow dangerously. That is why the design work to keep it rare mattered more than making it efficient.
On the revenue side, the order was:
- Organisation accounts.
- Tooling subscriptions.
- Verification.
- Placement.
That ordering was not what we projected. We expected verification to lead, because it is the most obviously valuable to an individual.
We underestimated how much a company managing thirty freelancers will pay for approvals and consolidated billing.
The lesson we would pass on is this: when you remove the automatic revenue, you find out quickly which of your optional products people actually value. It is unlikely to be the one you built first.
Clients reacted more strongly than freelancers
Clients cared more than freelancers. We built the argument for the person paying the fee.
The strongest reaction came from clients, who worked out that a fee absorbed by the freelancer is a fee reflected in their quote. Removing it made the same work visibly cheaper.
Portable reputation mattered more than we expected
We built exportable signed reputation records mostly as a principled position.
It turned out to be a decisive factor for people considering a move. They did not value it because they intended to leave again. They valued it because a platform willing to let them leave read as one confident enough not to need the lock.
We did not need to police off-platform transactions
The enforcement problem disappeared. Percentage platforms spend real energy detecting and preventing off-platform transactions.
We carry none of that cost, because nobody has a reason to leave. We had not thought about it in advance.
The switch only makes sense when the fee saving is large enough
The most common question is whether it is worth switching. The honest answer depends on a calculation most people have never done.
The calculation is:
- Work out what you paid in platform fees last year.
- Use the number, not the percentage.
- Work out what you would pay here: the tooling subscription if you want it, verification once, and nothing else.
For most people billing steadily the difference is several thousand a year. For people below a certain volume, it is small enough that switching is not worth the disruption.
We tell the second group to stay where they are. That is not modesty. A marketplace with a thin supply side does not benefit from users who arrive reluctantly and leave in a month.
Open Lance works better when you bring the work with you
We also say plainly that our discovery is weaker than the incumbents. If your work comes from being found by strangers, we are currently a worse choice. We would rather say so than have someone find out in their second month.
If your work comes from existing clients and referrals, and the platform is doing escrow, contracts and payment rather than lead generation, the arithmetic is straightforwardly in our favour.
That distinction is the single most useful thing we have learned about who this is for. It took considerably longer to work out than it should have.
The remaining risk is pressure to add a transaction fee
The honest risk is that optional revenue may never reach sufficient scale. If that happens, the pressure to introduce a small transaction fee becomes enormous, and every argument for it will sound reasonable at the time.
It is the path every predecessor took.
What we have done is structural rather than a promise. The revenue lines are products with their own roadmaps and accountability. At a difficult board meeting, the question becomes why verification is underperforming rather than why we do not just charge two percent.
Whether that holds under real pressure is not something we can prove from here, and anyone evaluating the platform should weigh it accordingly.
We would have changed the sequence
We would do three things differently:
- Build discovery first, or at least in parallel.
- Publish the economics from day one rather than in response to scepticism.
- Take the enterprise buyer more seriously earlier.
Trust engineering was necessary, but it was not the constraint on adoption. We spent our first year on the half users were not blocked by.
We added the explanation of the economics after people asked, which reads as defensive. Leading with it would have cost nothing.
Organisations engaging many freelancers have a genuine problem, real budget and no individual’s negotiating disadvantage. That segment values everything we built, and we treated it as secondary for too long.
We would do it again, with discovery earlier
Yes, we would do it again, with the sequence changed. The thesis about fees is correct and the arithmetic holds.
What we misjudged was that being right about pricing does not by itself make a marketplace work, because pricing is not the only thing a marketplace does.
The broader lesson generalises past this business. It is possible to identify a genuine structural unfairness, engineer a correct answer to it, and still be building the wrong thing first, because the unfairness that is most visible is not necessarily the constraint that binds.
The full build, including the escrow state model and the reconciliation design, is in the Open Lance case study. The platform itself is at Open Lance.
