Support desks
The ticket reaches someone who has actually solved this before.
Route by product area, language and tier — and keep the newest agent off the case that needs ten years.
For teams where AI agents and people share the work
One queue for your AI agents and your people.
Each job goes to whoever should take it. An agent takes what it can finish. A qualified person takes what it can’t. You see why for every job.
- Free while we are in launch
- Routing data stays in the EU.
- The engine is open source (MIT)
35 seconds. One job, from request to the person who gets it. The example routes work between people. Scores in the film are labelled examples.
Transcript
Here’s how Fivexer picks who gets a task, and where you check why.
A billing dispute comes in, tagged with what it needs.
First, your rules. Priya is off shift, so she’s ruled out.
Next, everyone left is scored.
Amara scores highest, so she gets the offer.
When she’s done, the outcome comes back, and the engine learns who closes these.
That’s the whole path.
Your rules first, then the closest fit, and every step on record.
The right person. With the reason on record.
Sign up free at 5xer.com.
How a task finds its person
The film’s six steps, as text. For readers who skim, and for search.
- Billing dispute1 A task comes inA billing dispute arrives, tagged with what it needs.
- 2 Your rules firstPriya is off shift, so she is ruled out. The reason is recorded.
PriyaOff shift - 3 Everyone left is scoredThe people who qualify are compared.
Amara
Leo - 4 Highest score gets the offerAmara scores highest, so she gets it.AmaraOffer
- 5 The outcome comes backWhen she is done, the outcome is recorded.AmaraDone
- 6 It learns who closes theseWhen learning is on, it learns from the outcome. Rules still come first.AmaraLeo
Hybrid work breaks at the handoff.
Agents do more every month. What they can’t do still lands on a person. Which one?
- The agent stalls.Half-done jobs sit there until someone happens to notice.With Fivexer: The agent reports why. A job it rejects goes to the next eligible worker, with the reason on the task.
Mei - The wrong person gets it.It lands on whoever answered first, trained or not, on shift or not.With Fivexer: Only qualified people who are on shift are considered.
Sam
Noor - Nobody can say why.A customer or an auditor asks, and there is nothing to show.With Fivexer: Every job keeps who was considered, and the rule behind each exclusion.PriyaAmara
Which jobs can your agent keep?
This is AI agent orchestration on one queue. Choose the work an agent may take and when it hands over to a person. The morning’s queue re-routes as you change it, and you get the rules to paste into your setup.
1 · Work the agent may take
2 · Hand over when the agent is less sure than
Paste into your setup
// the agent, registered as a worker
routingWeights: {
'refund:small': 1,
'address-change': 1,
'claims': 0,
'fraud': 0,
'kyc': 1,
}
// the workflow step the agent works in: one route per handover
.route('result.success === false', 'human-review')
.route('result.confidence < 0.8', 'human-review')A value of 0 means the agent is never offered that work. The two routes hand over inside a workflow; without one, a failed result simply finishes the task. Handing over on how sure the agent is only works if your agent reports it in its result.
3jobs the agent finishes
5go to a qualified person
1 hof people’s time, at 12 minutes a job
This morning’s queue · example8 jobs
Refund, double chargeRefunds under €100Allowed, and 0.94 sure. The agent finishes it.Agent
Refund, expired cardRefunds under €100The agent was 0.71 sure, under your 0.80. The workflow route sends it to a payments specialist.Person
New delivery addressAddress changesAllowed, and 0.97 sure. The agent finishes it.Agent
Water damage claimClaims triageAgents are never offered this work. It goes to a claims handler, level 2.Person
Passport photo checkID document checksAllowed, and 0.91 sure. The agent finishes it.Agent
Blurry driving licenceID document checksThe agent was 0.62 sure, under your 0.80. The workflow route sends it to a document reviewer.Person
Card used in two countriesFraud flagsAgents are never offered this work. It goes to a fraud analyst.Person
Refund, partial orderRefunds under €100The agent reported it could not finish. The workflow route sends it to a payments specialist.Person
Illustrative. The queue, the confidence figures and the roles are examples, not a customer’s data.
1 · Route
Agent or person, decided for every job
- Agents and people sit in one pool, with the same skills and limits.
- The engine never offers a job to someone a rule excludes. A manager can still assign by hand, as a separate, recorded step.
- Agents run on your own machines or servers, with any agent tool you can start from a command line. Nothing locks you in.
- Create jobs from the API, the SDK or the MCP server.
2 · Hand off
The agent reports why. A person can take it from there.
- An agent that can’t do a job reports why, and the reason stays on the task.
- A job an agent rejects is requeued and offered to the next eligible worker, a person if they qualify. The agent that rejected it is not offered it again.
- Who qualifiesOn shift, with the skills, and room for one more.
- Who’s left outEach person shows the rule that left them out.
- Who goes firstYour fairness mode decides.
- No answerThe offer moves on at the deadline you set.
3 · Record and staff
See why every job went where it did, then staff for the load
- A decision record per job: who was considered, who was excluded, and the rule behind each.
- Rejected tasks raise task.rejected, and the agent’s reason is on the task thread, so you can see which jobs agents turn down, what to fix and when you need more people.
- Four fairness modes spread the work, so the hard jobs don’t keep landing on the same people.
4 · Plan the people side
Someone qualified is on shift. If not, you hear first.
A roster plans who should be on shift. It reads the next three days, and tells you early when a shift won’t be covered.
- Qualified people are scheduled for the workShifts carry the skills they need, and a dated qualification counts only while it is valid.
- Rules are checked as you planContracts, leave, rest between shifts and weekly hours, from the limits you configure. A blank rule isn’t checked.
- A short shift is found days aheadThe look-ahead reads the next three days of a published rota and shows where cover is thin.
- The owner is told, with optionsA site’s owners get a push and an email. If nobody acknowledges it, it goes up to the workspace owners.
Automatic rostering · Skills-based scheduling · Scheduling rules
- Wednesday, earlyCoveredFour of four on the rota.
- Wednesday, lateCoveredThree of three.
- Thursday, nightAt riskA sick call leaves it one short. One colleague could step in.
- Friday, earlyCoveredFour of four.
Owner notified: Thursday night is at risk. One colleague could step in. The standby list names them.
Built for these operations
The vocabulary changes with the sector. The decision does not: every job is offered to the people and agents who can actually do it, and to nobody else.
Field service
The job goes to the qualified engineer who is nearest, not the one who answers first.
Certifications are a hard gate, distance is priced against urgency, and a full van is never given a sixth job.
Expert networks
The right domain expert answers before the window closes.
Best-fit available specialist surfaces automatically, and the ones who resolve fastest per domain earn more of the work.
Logistics
Work is spread across the fleet instead of piling onto whoever is closest to the depot.
Fairness modes stop one crew absorbing the day while another idles, with the load penalty derived per pass.
For teams who ship software
The fifth operation ships software
A P1 arrives, a PR wants a reviewer, an incident needs a follow-up. It is the question the four operations above are stuck on — asked at standup, answered from memory. Triage is dispatch, and it can have an engine.
- Your tracker already has the vocabulary.Issues arrive carrying their component, type and language as tags, and every engineer carries weighted skills on the same keys. A crash in the API goes to someone who ships in the API — not to whoever was next in the rotation.
- On-call and code ownership become gates, not conventions.A routing weight of zero is a hard veto, checked before anything is scored: the engineer who opted out of reviews this week is never offered one. The rules that must hold are the ones the engine never bends.
- Every decision ships with its own log.Who was considered, who was excluded and why, the score of everyone eligible — recorded when the decision happens and queryable through the same API that creates the work. Debugging a routing rule is reading a log, not reconstructing one.
- You can read the engine before you trust it.The matcher is an MIT-licensed library over Redis, the SDK is on public npm, and the API reference sits with the docs. Evaluate it the way you evaluate a dependency — because that is what it is.
How the matcher decides · When to build your own · The engine on npm
Two ways in
I run an AI-assisted service
Review, support and operations teams where an agent does the first pass and a person takes what it can’t.
I also run shifts and sites
Plants, depots, warehouses and field crews, with trained-person rules, rest between shifts and same-day cover.
What is the same in all of them
One engine, five vocabularies, and agents or people in every one. None of what follows changes with the operation, and that is why the same product fits all five.
- A rule you write is a wall for the engine, not a warning.A missing certification, a language nobody on shift speaks, an expired card: the engine never offers the job, it does not flag it. A person can still assign it by hand as a deliberate, separate step, and the record shows it was manual. The learning layer re-ranks who is eligible — it can never make somebody eligible.
- Every decision keeps its own record.Who was considered, who was blocked, and the reason. That is what turns “why did they get it?” into a lookup instead of an argument.
- Nobody absorbs the day.Four fairness modes — first-come, best-match, balanced, spread-work — trade a little best fit for a lot of balance, and you choose how much.
- It runs in the EU, and you can take it with you.Frankfurt, a data-processing agreement you can read without emailing anyone, and an MIT engine your own IT can run if our price ever stops working for you.
Is it for you?
A good fit if
- Agents handle part of the work and people handle the rest.
- Some jobs need a qualification, a site permission or an available person.
- Someone has to explain later why a job went where it did.
Not for you if
- Any worker can take any job and nothing needs checking.
- You need to build agents, or run payroll, HR, machine or route scheduling. Those stay in your other tools.
- You want a promise of legal compliance. Fivexer checks the rules you configure, and a blank rule isn’t checked.
What it doesn’t do
Written down here rather than found out in month two.
- It is not a helpdesk, a CRM, a phone system or an agent builder.It decides who gets the work and hands it to whatever you already run. Ticketing, telephony, the customer inbox and the agents themselves stay where they are.
- We do not hold SOC 2 or ISO 27001.Your data sits in the EU, the data-processing agreement is on this site, and the engine is public and free. If that is not enough for your reviewer, say so before you start rather than after.
- There are no customer logos on this page.We have not put words in anyone else’s mouth. When someone real is happy to be named, they will be named here.
- The crew side does not work without a signal.On a phone this is a web app, not an offline one. Plant rooms, deep basements and remote sites are a real problem, and we would rather lose the sale than the second month.
Send a job. See who gets it, and why.
Start with a free workspace. Add your people, register an agent as a worker, and send a first job. The decision record shows where it went and why.
Questions people ask before the first call
Can AI agents and people share one queue?
Yes. An agent registered with the @fivexer/agent daemon is a worker like any other, with its own tags, skills and backlog limit, so the same rules route work to it and to people. Agents run on your own machines or servers. Fivexer does not host or build them. Setup is at /docs/agentic/agent.
What happens when an agent can’t do a job?
The agent reports why, and the reason stays on the task. A job the agent rejects, or one it cannot run because its host fails, is requeued and offered to the next eligible worker, a person if they qualify. The agent that rejected it is not offered it again. Rejected tasks raise task.rejected, and the agent’s reason is on the task thread.
What kind of work can Fivexer route?
Anything an agent or a person is handed and can finish: support tickets, field service visits, expert reviews, bug triage, deliveries, inspections, shift cover. The engine matches on tags, weighted skills and priority, so the vocabulary is yours — it has no built-in idea of what a “ticket” is.
How is this different from the routing already in my helpdesk?
Fivexer is not a helpdesk and does not replace one. It decides who gets the work and hands it back to whatever you already run. The differences that matter are hard vetoes the engine always enforces, a decision record per job, four fairness modes, and the fact that the engine is open source so you are never locked in.
Can a rule be made un-overridable?
For the engine, yes. A routing weight of zero is a hard veto: the job is never offered to that person at all, rather than flagged for someone to override. Certifications, licences and jurisdiction limits are the usual reason people need this, and it is checked before any scoring or learning happens. A manager can still assign a job by hand, overriding the check as a deliberate, separate step, and the decision record shows it was manual.
How do I know why a job went to a particular person?
Every decision keeps a trace: who was considered, who was excluded and the reason, and the score of everyone who was eligible. It is queryable afterwards, so “why did they get it?” is a lookup rather than an argument.
Does it stop one person absorbing all the work?
That is what the fairness modes are for. First-come is the default; best-match picks the strongest fit deterministically; balanced and spread-work apply a load penalty derived per pass, so busy people lose near-ties — or most contests — to less-loaded colleagues who are still eligible.
Where does the data live?
Fivexer Cloud runs in Frankfurt, Germany. The Article 28 data-processing agreement is published at /dpa and the subprocessor list, including the two flows that leave the EEA and their safeguards, is on /security. The engine is MIT-licensed, so self-hosting means no processor at all.
How do we try it without disrupting live dispatch?
Run it in shadow mode: send real jobs to the engine, throw its answers away, and compare them with your dispatchers’ for a few weeks. There is a free four-week checklist for it at /guides/shadow-mode-checklist, including the one finding that means stop.