Twelve of our outcomesdescribe a machine.
Most systems call all of them no answer. So a voicemail, a phone menu, another AI assistant and a number that no longer exists look identical in the report, while each one wants a completely different next move.
25
system dispositions
4
outcome buckets
5
that stop the dialling
Voicemail, a phone menu, another AI assistant and a disconnected line all land in the same box.
You cannot fix a number you never knew was dead.
When voicemail, a phone menu and a disconnected line all report as “no answer”, the list decays quietly. Good numbers get dialled at the wrong hour, dead ones get dialled forever, and the report says the same thing about both.
Four buckets, and only one of themis about what was said.
Every code rolls up into one of four buckets, and the bucket is what most automation should key off. The largest is the one nobody else models.
| Bucket | Codes | What it means |
|---|---|---|
| reached_human | 9 | A real person answered. The only bucket where what was said carries meaning. |
| machine_answered | 12 | Something picked up and nobody was there. The largest bucket, and the one everyone else collapses. |
| no_connect | 3 | The call never established in the first place. |
| unclassified | 1 | Not confident enough to say. Preserved rather than guessed. |
All twenty-five, and what each one does next.
Each disposition carries a label, a description, a recommended action, and three switches: whether it ends the sequence, whether it stops the number being dialled, and whether it needs a person to look. Pick any code to see its own.
Two layers, and the second onehas to show its working.
Known wordings are matched against the transcript first. Anything the patterns cannot settle goes to a model, which is required to return the exact sentence that convinced it.
Layer one, patterns
Known wordings, each carrying a weight and a confidence threshold at which it can decide the call outright. Running hit and false positive counts mean patterns are tuned on evidence rather than on opinion.
Layer two, with proof
The model must return a quote copied verbatim from the transcript. If that quote is not genuinely present, the model is guessing, and the answer is downgraded to needs review rather than stored as a confident error.
Only the answering side
The classifier reads the caller’s turns and never our agent’s, so an agent saying "sorry I missed you" can never be misread as evidence of a voicemail.
More reluctant, not less,when it cannot be undone.
Five codes permanently stop a number being dialled. A false positive on any of them destroys a real customer record for good, so they are held to a 0.9 minimum confidence where ordinary codes need far less.
This is a trust argument rather than a technical one. The system is more cautious precisely where the consequence cannot be reversed, which is the opposite of how most automation behaves. Confidence is easy to spend when nobody is counting the cost of being wrong.
The system remembers howthis number behaves.
Beyond a single call, every contact keeps a live state: current disposition and bucket, total attempts, consecutive machine answers, whether it needs attention, the recommended next action, whether it is suppressed and why, plus first seen, last called and last campaign.
The consecutive machine counter is the quietly valuable one. A number that has hit voicemail four times running is a different proposition from one being tried for the first time, and the system knows the difference without anybody running a report to find out.
Add your own, and change what they do.
The system set is a starting point, not a fixed list. The vocabulary is per account.
Your own dispositions
Sitting alongside the system set, in the same reports and the same automations.
Per-disposition behaviour
Label, description, recommended action, and the three switches: ends the sequence, stops the dialling, needs attention.
Your own ordering
So the outcomes your team records most often sit at the top of the list rather than alphabetically wherever they land.
Corrections are recorded
When a human overrides a classification it is kept, which is what makes the pattern layer improvable rather than static.
A disposition is an input, not a full stop.
The code decides what happens next, on this call and on every future one to the same number.
What the extra twenty-one codes buy you.
Stop dialling dead numbers
Disconnected lines, fax tones and blocked routes take themselves out of the list.
Retry the right way
A voicemail, a phone menu and a busy tone each deserve a different next attempt, and get one.
Evidence for compliance
Do not call and complaints stop the dialling on the spot, with the quote that proved it attached.
Honest uncertainty
Needs review exists so that nothing confident is ever invented to fill a gap.
What people ask about call outcomes.
A call disposition is the structured outcome recorded when a call ends, such as interested, booked, voicemail, wrong person or do not call. Because it is a code rather than free text, connect rate and conversion become numbers you can report on and automate against. Sysevo ships 25 system dispositions in four buckets, and you can add your own alongside them.
Stop reporting four outcomes.Start acting on twenty-five.
A short demo on your own call recordings. We will classify a handful live and show you the sentence behind every verdict.
Book a demo- 01Why Lead Disposition Tracking Cost Keeps RisingLead disposition tracking costs £800–£5,000 monthly depending on call volume and CRM depth. Find hidden fees and build an accurate budget.CRM
- 02How to Set up AI Outbound Calling with ZanusLearn how to deploy AI outbound calling with Zanus for sales teams. Covers setup, integration, real-world results, and when to use this approach.AI & Automation
- 03Does Zanusai Help with AI Outbound CallingEvaluate whether Zanusai suits your outbound calling needs. Compare features, pricing, and integration gaps against your campaign automation requirements.AI & Automation