Give your AI access to
human intelligence.
One API call sends a call recording or transcript to a vetted human reviewer. Structured JSON back in minutes, matching a schema you define.
Stop paying engineers to listen to calls. Stop standing up a QA team to grade your agents. No headcount, no minimum, and nothing to talk to a salesperson about before you start — one endpoint.
Asynchronous by design. This is human review in minutes, not a real-time channel — so we publish the queue instead of promising a number. Nobody is at the queue right now; a backup reviewer is paged the moment a task lands.
Try it — no account
Pick a call. Watch it become JSON.
This runs the real API. Real schema validation, real response shape, real task id. The human answer at the end is a replay of an actual review — we label it, because faking the human would defeat the point.
Instructions for the reviewer
plain languageListen to this outbound sales call handled by an AI agent. Determine: - Did the customer clearly agree to a demo? Only mark true if they committed to a specific time or explicitly said yes. - Which objections did they raise? - Was pricing discussed with a specific number? - How confident are you in your reading of this call?
{
"type": "object",
"properties": {
"demo_booked": {
"type": "boolean",
"description": "True only if the customer committed. 'Send me some times' is not a commitment."
},
"objections": {
"type": "array",
"items": {
"type": "string",
"enum": [
"price",
"integration",
"timing",
"authority",
"incumbent",
"trust"
]
},
"description": "Every objection the customer raised."
},
"pricing_discussed": {
"type": "boolean",
"description": "True only if a specific number or range was said out loud."
},
"confidence": {
"type": "string",
"enum": [
"low",
"medium",
"high"
],
"description": "Your confidence in this assessment."
}
},
"required": [
"demo_booked",
"objections",
"pricing_discussed",
"confidence"
]
}Press Send to a human to POST this task to /v1/tasks.
Install
Three ways in. Same API.
Pick the one that matches how you build. The skill teaches an agent when to reach for a human; MCP gives it the tools; REST is there when you want to wire it yourself.
# Drop the skill into your agent. That's the install.
mkdir -p ~/.claude/skills/zelto-humans
curl -sL https://humans.zelto.ai/skill \
-o ~/.claude/skills/zelto-humans/SKILL.md
# Your agent now knows:
# - when a decision needs a human instead of a model
# - how to quote a task for free before spending anything
# - how to sandbox-test the integration with no API key
# - what a review costs, and to tell you before it charges you
The file is versioned and served from our origin, so an agent reading it can replace its own stale copy. No package to install, nothing to upgrade.
How it works
Four steps, one endpoint.
No workflow builder, no annotation project, no onboarding call. If you can describe what you want in a sentence and a schema, you can use this.
01
Submit
POST a recording or transcript, plain-language instructions, and the JSON Schema you want back. Price it for free first with /v1/quote.
02
We hold, not charge
Credits are reserved at submit and charged at completion, against the real call length. Cancel while queued, or hit an unreviewable file, and you pay nothing.
03
A human reviews it
Your schema becomes a form. A vetted reviewer listens, answers each field, and writes their reasoning. Median turnaround is minutes.
04
Structured result
Poll the task or take a signed webhook. You get JSON that validates against your schema, plus the reviewer's notes — which is often the part you needed.
Pricing
Per reviewer-minute. Not per seat.
Priced on the thing that actually costs money — a person's time. No subscription, no minimum, no per-seat fee. Credits do not expire.
Transcript review
Reading is 3–4× faster than listening. If a transcript answers the question, send the transcript.
$0.75
per task
Audio review
Then $0.25 per additional minute. Pass audio_duration_seconds for an exact hold.
$2.00
first 5 min
Re-review of a wrong answer
A different reviewer answers independently. Recourse is the whole reason to trust a stranger with your calls.
Free
always
Quotes, sandbox runs, polling, cancels
Find out what something costs, and prove your integration works, before you spend anything.
Free
unlimited
An in-house QA analyst costs roughly the same per call. We are not cheaper than a QA team you already have — we are cheaper than hiring one. Zero fixed cost, no headcount req, no ramp, and it arrives as JSON.
Everything above is self-serve and stays that way. Book 15 minutes if you need a committed SLA, volume past the hourly cap, reviewers trained on your domain, a task type we don't support yet, or a DPA before sending regulated audio.
Book a discovery callBefore you integrate
The things a demo would skip.
Everything below is the sort of detail you'd normally find out in week three. It's cheaper for both of us if you find out now.
Turnaround
Asynchronous. Minutes, not seconds. If your agent is mid-call and needs an answer now, this is the wrong tool and we would rather you know before you integrate.
Who reviews
A small vetted team, not an open marketplace. Small enough that submissions are capped per hour — a cap that protects your turnaround, not our margin. Steady volume above it is a conversation, not a paywall.
When it's wrong
POST /v1/tasks/:id/redo. Free, independent, different reviewer. We track redo rate as our primary quality metric.
Your audio
Copied to our storage at submit, because signed URLs expire before humans get to them. Deleted after 30 days; every task reports its own retention deadline. Sending recordings from a regulated domain? Email us for a DPA before you send volume.
Schemas
A flat object of typed fields, up to 25 of them. Nested objects and oneOf are rejected with a message naming the exact path — because a human has to fill this in, and a form we guessed at is worse than a clear error.
What we're not
Not a labeling vendor, not an outsourcing agency, not a freelancer marketplace. There is no account manager and no statement of work. It's an endpoint.
Stop listening to calls. Start querying humans.
Sign up, add credits, get a key. First real review in about five minutes — or run the sandbox right now and pay nothing.
Custom SLA, high volume, or regulated audio? Book a 15-minute discovery call.