AI
Explainable, or it is not useful.
A ranking a coordinator cannot argue with is a ranking they will not trust at six in the morning. Everything below either shows its reasoning or is not shipped yet — and this page says which is which.
Available today
Ranked matching that names its reasons.
- 01 Ranked candidates for a shift Scored against what the site actually requires, with the reasons on every row — worked here before, same city, has the ticket, owns a vehicle.
- 02 Blockers shown, not hidden People who cannot take the shift stay on the list, greyed, with the reason in red. A coordinator forty minutes from a shortfall is entitled to know that the only person left needs boots.
- 03 Already-booked is visible Someone working elsewhere that day is shown as such rather than quietly dropped from the results.
How the ranking is built
It is arithmetic, and you are allowed to see it.
There is no model in the matching. A shift has requirements, a person has attributes, and the list is sorted by a score made of the points below — which is why every row can tell you exactly why it is where it is.
- Worked at this site before
- The largest single positive. Somebody who has done this job at this address is worth more than anybody who matches on paper, and every coordinator already knows it.
- Same city as the site
- A strong positive. Distance is the most common reason a confirmed person does not arrive.
- Own vehicle
- A positive on its own, and a larger one when the order says transport is required.
- Required ticket held
- Counts when the order asks for it. An order that does not need WHMIS gains nothing from a worker who has it.
- Safety footwear
- Same rule. A requirement on the order, matched against the person.
- Lift capacity
- Matched against what the order asks for, because it is the requirement most often discovered on arrival.
- Already booked that day
- A large negative — large enough to sink a row to the bottom, and deliberately not an exclusion. They stay on the list with the reason showing.
Why the arithmetic is published
A score you cannot interrogate is a score nobody uses twice.
These are today's weights. They are visible in the product on every row, so the list and this page cannot quietly disagree with one another.
A coordinator covering a shortfall at twenty to six has about ninety seconds and a phone. If the top of the list is wrong, they need to know why it is wrong faster than they can scroll — and the only way that happens is if the reason is on the row.
So the ranking names its reasons in the words the desk already uses: worked here before, same city, has the ticket, owns a vehicle, already booked that day. Not a confidence percentage, not a five-star rating, and not a number between zero and one that somebody would have to develop an intuition for over several months.
It also means the ranking can be argued with, which is the point. A coordinator who disagrees with it is disagreeing with something specific.
Also shipped
The assistant on your own careers page.
It runs on a third-party model. Anthropic or OpenAI, whichever the account is configured for. We say so here because a chat box that does not mention it is a chat box hiding something.
The public site Zawrix publishes for you can carry an assistant that answers the questions applicants actually ask — what you place, where, what shifts, how to apply. It is given your own published content to work from rather than the open internet.
It has a hard monthly spend ceiling rather than a warning, so an unusual month costs what you decided it could cost and then the assistant stops and hands over to a phone number.
And it refuses identity numbers. A careers chat box is the single most likely place in any staffing system for somebody to type their Social Insurance Number — not because anyone asked, but because every other agency has asked them for it and there is a text field. Those digits are stripped the moment the message arrives, before it is stored, before it reaches a model and before it appears in any log. There is no path where the original text is kept "just for the transcript".
In development
Not on your account yet.
Listed here rather than omitted. A roadmap you cannot see is a roadmap that gets sold as if it had already shipped.
Conversational search over your own pool — "who in Brampton has a forklift licence and is free Monday" — and drafting assistance for the messages a coordinator writes forty times a week. Both are being built against synthetic data. Neither is switched on for customers, and neither appears in a plan you can buy today.
There is no date on either, because we have not earned the right to give you one. When they ship they will appear in the section above this one, and this section will get shorter.
What we will not build
Nothing here decides who gets hired.
Not a positioning statement. An agency is the employer, and the employer is who the regulator, the client and the worker all come to.
Zawrix does not autonomously reject candidates, does not produce a score that stands in for a judgement about a person, and does not rank people on anything it cannot show you the reason for. Employment decisions are made by the agency, and a system that obscures why somebody was not called is a liability rather than a feature.
We also do not publish improvement percentages for any of this. We have no customer base to measure across yet, and a number we cannot substantiate is worth less than saying so.
Questions
The ones worth asking about any of this.
Is my data used to train a model?
No. The matching contains no model to train — it is the points arithmetic set out above, running against your own records inside your own schema.
The careers-page assistant sends the visitor's message to a third-party model to get an answer back. That is an API call, not training, and nothing from your pool, your clients or your invoices is in it.
Which model, and where does it run?
Anthropic or OpenAI, depending on how the account is configured. It is a hosted API, so those messages leave our infrastructure to be answered.
This is the one place in Zawrix where that is true, which is exactly why it is written on the page rather than in a settings screen nobody opens.
Can I turn the assistant off?
Yes. It is a feature of the public site we publish for you, and a site without it simply does not have a chat box. The matching is unaffected either way — they are separate things that the industry's habit of calling everything "AI" makes sound related.
Why is there no accuracy figure anywhere on this page?
Because we would have to make it up. Publishing "38% faster fills" requires a customer base to measure across, a baseline to measure against and somebody independent to believe. We have the first of those in one agency and none of the rest.
A number we cannot substantiate is worth less than saying so, and an agency owner who has been sold software before can tell the difference.
What happens when the ranking is wrong?
You look at the reasons on the row and you disagree with one of them. That is the whole design: a ranking that can be argued with, by name, in the words the desk already uses.
Nobody is removed from a list by the software. A person who cannot take the shift stays on it, greyed, with the reason showing — because the coordinator forty minutes from a shortfall is entitled to know that the only person left needs boots.
Does it screen out candidates automatically?
No, and it will not. Zawrix does not autonomously reject anybody, does not produce a score that stands in for a judgement about a person, and does not rank on anything it cannot show you the reason for. Employment decisions belong to the agency that makes them.
Ask it to rank a real shift.
Bring an order you are covering this week and watch the list explain itself.