Bolt Must Explain Its Algorithm — and So Will Employers
It's 6am in Tallinn and a courier can't log in. The app says the account is restricted. There's no name attached to that decision and no one to ring. Somewhere a rule fired, and the explanation on the screen is a single line.
From December 2026 that stops being acceptable. Under the EU Platform Work Directive, Bolt, Wolt, Uber and every other app that hands out work have to let a person make calls like that one. They have to put the reason in writing, and look at it again if the courier asks. Estonia's economy minister, Erkki Keldo, summed it up: "a situation cannot arise where an algorithm takes away a person's livelihood" (ERR News).
You probably don't run a delivery app. But if software decides who gets which shift or which job at your warehouse, plant or depot, the same question is on its way to you. And whether it's cheap or expensive to answer depends on one thing you can check today.
What the platforms now have to do
Directive (EU) 2024/2831 made the news for one reason: if a courier's working life looks like employment, the platform now has to prove it isn't. That's the half about contracts.
The other half is about the algorithm, and it covers everyone working through a platform, employed or self-employed. In plain terms, a platform has to:
- Tell people which automated systems watch them or decide things about them, what those systems decide, and what they weigh. In writing, on day one, and again before anything changes (Article 9).
- Keep a person in charge who can overrule the system, review its effect on workers at least every two years, and let a human make the call whenever an account is restricted, suspended or closed (Article 10).
- Give workers someone who can explain a significant decision, written reasons when an account is restricted or closed, and a proper answer to a review request within two weeks. If the decision was wrong, fix it or pay for it (Article 11).
- Leave some data alone: nothing on emotional state, private conversations, time off the clock, or who might join a union (Article 7).
Member states have until 2 December 2026. Estonia has said it will do the minimum, but even the minimum includes the human decision, the written reasons and the two-week answer. Bolt's Baltic policy lead told ERR that "nothing changes for us" (ERR). On contracts, maybe. On the algorithm, a lot does.
An explanation is a record you already have, or don't
Picture the two-week deadline from the other side. A decision was made a month ago, maybe at three in the morning. Someone wants to know why.
To answer properly you need three things: what the system knew at the time, which rule it applied, and who else it considered. If those were written down when the decision happened, the answer takes an afternoon. If they weren't, someone has to dig through logs and config history and ask around until the story holds together. That digging is the expensive part of any dispute about an automated decision. It's also the part lawyers charge for by the hour.
flowchart LR
D[System decides] --> R[Record written at the time]
R --> E[Explanation to the worker]
E --> H{Person reviews it}
H -->|stands| K[Answer within two weeks]
H -->|wrong| C[Fix it or compensate]
D -. nothing kept .-> X[Rebuild it from logs and memory]
X -.-> EThe dotted line is where the time and the fees go.
Why this reaches your site too
The Platform Work Directive is about platforms. A distribution centre handing picks to its own staff isn't one, and neither is a food plant whose rota comes out of a solver. Two other laws ask them the same question anyway.
GDPR, already in force. Article 22 says people shouldn't be subject to decisions made purely by automation that significantly affect them, and gives them the right to a human and the right to challenge. Article 15 gives them the right to understand the logic. In February 2025 the Court of Justice spelled out what that means: explain the procedure and principles you actually applied, in a way the person can follow. Not the source code, and not a vague paragraph either (Dun & Bradstreet Austria, C-203/22).
The AI Act, from 2 December 2027. Systems that hand out work based on a person's behaviour or traits, or that monitor and evaluate workers, count as high-risk (Annex III, point 4(b)). The date was pushed back from August 2026 (Gibson Dunn). Pushed back, not dropped.
So sooner or later someone on your floor will ask why they got the night shift, or why the forklift job went to somebody else. Can you show them?
What a good answer looks like
In Fivexer, work is handed out by rules you write down yourself: skills with weights, priorities, and hard stops. A weight of zero is a hard stop. That person never gets that kind of job, whatever else they score. Every candidate is scored, and anyone who's ruled out is recorded with the reason.
Here's a small example, run by the real engine when this page loads. There are three people and three jobs in a chilled warehouse. Ana has a forklift licence and cold-store training. Tomas has the licence, but he told his manager about a health reason to stay out of the cold store, so his cold-store weight is zero. Priya works cold-store picks and has no licence.
Every decision keeps its reasons, including who was ruled out — computed outcome:
- 3 agents, 3 tasks, same team — Best match. Every outcome below is computed, not drawn.
- Task dock load (flt) — Best match: A (score 100).
- Task chill pick (chill) — Best match: P (score 90).
- Task chill load (flt, chill) — Best match: A (score 180).
- The tally — Best match: A 2 · T 0 · P 1.
The dock load goes to Ana, who scores 100 to Tomas's 90. Priya has no licence, so she was never in the running. The chill pick goes to Priya, 90 to Ana's 80, and Tomas is ruled out by his zero. The chill load needs both skills. Ana scores 180, Priya 90, and Tomas is ruled out again.
Say Tomas asks why he never gets cold-store work. You point to one line in his profile, which he gave you himself, and the record showing it applied. No model's opinion of him comes into it. On the hosted platform that record sits behind GET /v1/decisions?taskId=…: who was considered, who won, and why (how matching works). It's written during the decision, not rebuilt afterwards, so it shows what the system knew at that moment.
What you'd hand over
| What the law asks | Where it comes from | What you'd show |
|---|---|---|
| Describe what the system weighs | PWD Art 9, GDPR Art 15 | Your own settings: skills, weights, hard stops, fairness mode. The engine is open source (MIT), so the procedure is public code, not a vendor secret |
| A person makes the calls that matter | PWD Art 10, GDPR Art 22 | A manager can overrule any match (POST /v1/tasks/{id}/assign), and the override is recorded too |
| Explain one specific decision | PWD Art 11, GDPR Art 15 | The decision record for that job: who was considered, their scores, and why each ruled-out person was ruled out |
| Check before you act | Sensible under all three | POST /v1/tasks/suggest-workers shows who would get a job without handing it out |
| Don't profile people for rotas | AI Act Annex III 4(b) | The rota planner uses declared skills, availability, legal limits, cost and hours actually worked. No reliability scores, no acceptance history |
| Keep workers' data in the EU | GDPR Chapter V | The platform runs in one EU region, Frankfurt (data residency) |
None of this makes you compliant. That's a judgement about your whole setup, and it belongs to you and your adviser. What changes is the question you take to them. "Here's the record, is our process right?" is an afternoon of their time. "Can you find out what our system did to this person in March?" is a project.
Where Fivexer stops
We won't write your Article 9 notice, your data protection impact assessment, or your reply to a worker. The decision record is written for a manager. Turning it into two sentences a picker understands is still your job.
There's also an optional learning layer that re-ranks eligible people using how past jobs went. That's exactly the kind of processing you'd have to tell workers about and weigh against the AI Act. It's off unless you switch it on, and the hard rules always run first either way.
And if you run a delivery fleet, we don't plan routes or count tachograph hours. We decide who does which job and which shift. Not all of what Bolt does.
One thing to try this week
Pick one decision your software made last month. A shift, a job, someone who got skipped. Try to explain it in writing to the person it affected. If that takes more than ten minutes, what you're missing is the record, not the lawyer. Then set up one real job type in Fivexer with your own skills and hard stops, and read the record it leaves behind. That's the page you'd hand over.