It reads every CV. Just not those four fields.
Iris scores every application against your competency framework, watches your pipeline for stalls against your own SLAs, keeps every candidate informed at each stage, and books interviews with the AI interviewer or a human panel. It declines nobody — anything that would close an application goes to you.
- Declines nobody
- Starts on a supervised trial
- Every score shows its reasoning
I’ve set up a starting configuration from your hiring data. Here are the five things I’ll take on.
Working on live data from your hiring module. Every run is written to the log under agent://cv-screening@tanca v3.1.
- 01Read and score every CV against the competency frameworkInternal
- 02Watch candidates in the pipeline against your hiring KPIs and SLAsInternal
- 03Tell each candidate where they are at every stageExternal
- 04Book interviews with the AI interviewer or a human panelExternal
- 05Measure candidate satisfaction and how the process is performingInternal
Screening is where the time goes, and where the trouble starts.
Three hundred CVs per opening, read at speed, late in the day, by whoever is free. Nobody applies the same standard twice, and nobody can reconstruct afterwards why the third candidate didn’t go forward.
Automating it is the obvious answer and also the frightening one, because the failure mode isn’t slowness — it’s a system quietly declining people for reasons nobody chose, and no record of having done it.
Which is why this agent’s configuration opens with what it won’t do, and why it starts on a supervised trial instead of going straight to work.
Five tasks, each labelled with what it can reach.
Turn any of them off individually.
Score every application
Reads each CV against the skills and weights in your competency framework and returns a level per skill, with the axes and the evidence source shown.
Never: scores on name, age, gender or place of origin.
Watch for stalls
Compares every candidate’s time in stage against your own SLAs and flags roles going cold before you notice.
Never: moves a candidate to close a stall.
Keep candidates informed
Tells each person where they are at every stage, on your behalf, with the sender stated in two layers.
Asks first: before any message where consent status is unclear.
Book the interviews
Finds slots across panel calendars, sends the invitation, chases the confirmation, and hands off to the AI interviewer where you’ve chosen that.
Asks first: before inviting anyone externally for the first time.
Measure the process
Candidate satisfaction after each stage, and how the pipeline is performing against your targets.
Never: shows an individual’s satisfaction response attributed to them [CONFIRM].
Nothing in this list can close a candidate’s application. That class of action exists, and it always belongs to a person — in trial and live alike.
Name, age, gender, place of origin.
Not weighted lightly. Not “considered in context”. Not inputs at all. And the fourth one is on the list because it’s the one this market actually has a problem with, and nobody names it.
Name
Including what it implies about family, region or ethnicity.
Age
Including the proxies: graduation year, years since first job.
Gender
Including honorifics and pronouns in the document.
Place of origin
The field that costs candidates the most in this market and appears on almost every Vietnamese CV.
It scores against what you wrote.
The skills and weights in your competency framework, and nothing else. If a requirement isn’t in your framework, it isn’t scored.
Every score shows its reasoning.
The component axes and the evidence behind each one, always, with one click to the full profile. A number you can inspect is a number you can correct.
Proxies are the real risk, and we won’t claim to have solved them.
Removing four fields doesn’t make a model neutral — a school name or a district can carry the same information. What we can do is show you every axis behind every score so you can see what it’s actually rewarding, and give you the framework controls to change it.
This describes how the product works and is not legal advice. [CONFIRM WITH COUNSEL before publishing any statement about compliance with employment law.]
Below the line means a person should look, not that nobody will.
A low score sends a candidate to a review list. It never closes an application, never removes anyone from a board, and never hides a row from a search result.
The bottom band is called worth a look because that’s the accurate description — the evidence on file doesn’t yet show what the role asks for, which is a different statement from a judgment about the person. Career changers and people with badly written CVs live in that band, and a tool that quietly buries them costs you hires you’d never know you missed.
Stage-closing actions exist as a category in the system, and they are permanently reserved for a person. There is no threshold you could set that would change this, because it isn’t a setting.
- No automatic decline, at any threshold
- Nothing hidden from any view
- One click from any score to the person
You wouldn’t let a new recruiter loose on day one either.
The agent runs in a trial period with a visible countdown. During it, everything it produces goes to a queue for a person to review — including the internal, reversible work that wouldn’t normally need approval.
You read its output for a week. The scores it gave, the messages it drafted, the stalls it flagged. If it’s getting things right, you promote it. If it isn’t, you change the configuration or leave it in trial for longer.
There’s no automatic promotion at the end of the countdown. It stays in trial until a person says otherwise, which is the opposite of how trial periods usually work in software.
- Everything queued, including internal work
- Promotion is a deliberate act
- No automatic graduation
Everything Iris produces is going to the queue for review, including internal actions. Promote it when you’re satisfied.
Promote to liveThree classes of action, and one of them is never automatic.
The useful question isn’t “is it autonomous”. It’s “what specifically can it do without me”, and that has a precise answer.
During the trial period, all three classes queue for review — including the first.
A message at every stage, sent on your behalf, saying who it’s from.
Most applicants never hear anything. Not a response — nothing at all, which is worse, because they keep checking. It’s the single most consistent complaint about hiring in this market and it’s entirely a coordination problem.
The agent sends a short, plain update whenever a candidate’s stage changes: received, being reviewed, invited to interview, decision made. Every message states its sender in two layers — on behalf of your company, sent via Career GPS. Never disguised as a personal email from someone who doesn’t exist.
And it never contacts someone whose consent status doesn’t allow it, which is checked per message rather than assumed once.
- An update at every stage change
- Two-layer sender, always
- Consent checked per message
Thank you for applying. We’ve reviewed your application and would like to invite you to the next stage.
You’ll receive a calendar invitation shortly with the details of the interview time and format.
It notices the stage that’s slowing before you do.
You set the targets — days in stage, time to first shortlist, time to hire. The agent compares every live role against them continuously and raises a flag with the specific number attached.
It flags. It doesn’t fix. A stalled interview stage usually means a person is busy, and moving candidates along to clear a metric is the worst possible response.
- Your targets, not ours
- The specific number, not a colour
- Flags, never resolves
Average 6 days, up from 2 last month. 4 candidates waiting, longest 9 days.
Four sources, read-only. Every run on the record.
The same scope and the same log as every other agent, because a different answer per agent would be a bad sign.
Read-only. Its outputs come to you; nothing writes back to these.
Agent runs use your monthly allowance.
Shared across all six agents.
Draws on your AI allowance
- Trial-period runs count [CONFIRM]
- Scores already produced stay readable at zero remaining
- Credits are self-serve
Unlimited and free
- Configuring the agent
- Reading the run log
- Reviewing the queue
- AI Screen intake scanning
- Everything in Pipeline, Talent Pool, Assessments, Competency Framework, Pre-Onboarding and Reports
At zero remaining, the agent pauses. Nothing it already did is rolled back, and your pipeline works exactly as it did before you turned it on.
This is one of five.
Zola
Adapts and publishes roles across every channel, tracks cost per qualified applicant.
Learn more →Iris
Reads CVs against your competency framework and routes candidates to the right next step.
You’re reading this oneOrin
Conducts a scored first-round interview against the role’s competency requirements.
Learn more →Remy
Drafts and tracks offer letters, with salary band checks before anything goes out.
Learn more →Hana
Manages the checklist from signed offer to the end of probation.
Learn more →One allowance, one log, one queue. Turn on as many or as few as you want.
Measured inside the product.
Food retail chain · 22 stores
of screening removed per opening after the first month [CONFIRM]
Logistics company · 3 warehouses in the south
of applicants received a stage update at every step [CONFIRM]
Contact centre · recruitment hub
hires came from the bottom two score bands in the first quarter [CONFIRM]
You’ll see the same figures in your own run log from the first role.
Questions people ask before switching it on.
Can it reject a candidate?
No. Stage-closing actions are a permanent category reserved for a person, in trial and live alike. There’s no threshold that changes this because it isn’t a setting.
What does it not score on?
Name, age, gender and place of origin. Those fields aren’t inputs, including their common proxies.
So it’s unbiased?
No, and we won’t claim that. Removing four fields doesn’t remove every proxy. What we do is show every axis behind every score so you can see what it’s rewarding, and give you the framework controls to change it.
What is a trial period?
The agent starts with everything queued for your review, including internal work, with a visible countdown. You promote it to live when you’re satisfied. It never promotes itself.
What are internal, external and stage-closing actions?
Internal happens in your system and is reversible. External reaches a person. Stage-closing would end an application and always needs you.
Will it message our candidates?
Yes, stage updates on your behalf, with a two-layer sender, and only where consent allows — checked per message.
What does it read?
Four named sources, read-only, listed on screen and removable: candidate records and CVs, your competency library, the interview panel calendar and salary bands by job grade.
What happens to a low score?
It goes to a review list labelled Worth a look, with the reason attached. Nothing is hidden or closed.
How many runs are included?
[CONFIRM] per month, shared across the six agents.
Is there a record?
A read-only log per run with the action, its class, the time, the approving person and the agent version.
Put it on trial before you put it to work.
Read its configuration, watch a week of its output, promote it when you’re satisfied. It declines nobody at any point.
Start free- Declines nobody
- Four fields never scored
- Every action logged