Talk to an engineer Book a demo
AI employees

AI employees, hired like people.

A role. A manager. KPIs that are outcomes, not minutes. A review every Friday and a performance file you can open. FormantAI's whole machine — governance, memory, learning, evaluation — seen as one employee on your collections floor.

Open an employee fileThe 6-week onboarding

Nobody hires a person by the minute. So why buy a voice agent that way? An employee has a role, a boss, a review and a file — and so should the model.

The four things every hire has

Role. Manager. KPIs. Permissions.

Role

One job, named.

Collections Officer, bucket 1, Hindi/Hinglish. Not "an AI". A role has a journey, a portfolio, calling hours and a tone — and it does not wander outside them.

Manager

A human who signs.

A named owner on your side — Head of Collections, QA lead — who reads the weekly scoreboard and approves or declines every release. No release ships without a signature.

KPIs

Outcomes, not minutes.

Promise-to-pay kept, right-party contact, complaints per thousand calls, cost per resolved call. Talk time is a cost line, never a target.

Permission set

See, say, do — bounded.

Reads due_amount, never the card number. Captures a promise; cannot waive a fee. Frozen lines it may never rephrase. Governance

The employee file

Open the file. Read the record.

Every FormantAI agent has one: its role and reporting line, its permission set, the version it runs today, the last review it passed, and the KPIs it is held to. It is the same record your risk team asks for about a human — and every line is backed by the audit log.

The performance file is the evaluation scoreboard; the weekly review is the learning cycle; the permission set is governance. Nothing on this page is a new system — it is the machine, read as a person.

The performance file The weekly review The platform
Employee file · AI-EMP-0042 · Riya
EmployeeRiya AI-EMP-0042
RoleCollections Officer · bucket 1 (DPD 1–30) · retail EMI
Reports toHead of Collections (named owner) · signs every release
Languagesहिन्दी Hindi · Hinglish · தமிழ் Tamil
Permission setpolicy: collections-b1 · see: due_amount, dpd_bucket · say: frozen lines v3 · do: ptp, pay_link, callback · deny: fee_waiver → human
Current versionv42 live since 14 Sep · canary 5 → 25 → 100%
Last reviewFri 12 Sep · scoreboard red→green +31 · green→red 0 · frozen lines exact-match illustrative
Onboarded6-week pilot · shadow → assisted → live on one cohort
KPIs
PTP keptoutcome
Right-party contactoutcome
Complaints / 1,000outcome
Minutes talkednot a KPI
ID badge"Main AI assistant hoon." — said on every call, frozen, never learned
Exitone-click rollback to any prior version · or retire the role; data stays in your tenant

Illustrative file. replace with a live employee record

A week in the life

Monday it ships. Friday it is reviewed.

The learning cycle, seen from the employee's side of the desk: release, work, get graded, get corrected, get reviewed — and rehearse the next version before anyone hears it.

Week in the life · collections-hi · v42 → v43
Mon

Release day

  • v42 candidate → canary 5%
  • guardrails watched · auto-rollback armed
  • 25% by noon · 100% by close
Tue

On the floor

  • calls 09:00–19:00 · RBI FPC hours
  • bucket-1 EMI reminders · Hindi / Hinglish
  • legal threat → warm transfer to a human
Wed

Graded

  • 100% of calls scored: ground · flow · comply · persona
  • failures clustered and named
  • 400 calls drawn for blind human labels
Thu

Corrected

  • your QA team grades the sample
  • corrections → next training set
  • Aadhaar / PAN redacted before anything is stored
Fri

Weekly review

  • manager opens the performance file
  • reads red→green · green→red
  • signs — or declines — v43
Sat

Half day

  • calls till 13:00 · callbacks the customer asked for
  • promise-to-pay follow-ups
  • after-call work: notes · disposition · CRM
Sun

Simulated

  • no outbound campaigns
  • simulator replays 1,240 scenarios against v43
  • parity check: Hindi · Hinglish · Tamil

Illustrative week for one collections role. confirm cadence with ops

Onboarding · badge · exit

Hired in six weeks. Fired in one click.

1
Weeks 1–2 · Shadow

Listens, never speaks

Trained on your redacted calls inside your tenant. Frozen lines, glossary and escalation matrix loaded. Graded on replays only.

2
Weeks 3–4 · Assisted

Speaks with a human beside it

One cohort, one language. Humans take over mid-call at will; every takeover becomes a correction.

3
Weeks 5–6 · Live

Own the role, canary first

5% → 25% → 100% with auto-rollback. First weekly review signed by the manager. How we work

4
Any week · Exit

Roll back, or retire

Back to any prior version in seconds; or retire the role. Recordings, transcripts and weights stay yours.

ID badge · worn on every call
R
Riya
AI assistant · Collections Officer
live
🔒 "Main AI assistant hoon." — AI disclosure, every call
🔒 "Yeh call record ho rahi hai." — recording disclosure
🔒 Grievance line · calling hours · tone by DPD bucket
Version-hashed · exact-match on every releaseNever learned
Exit · rollback and retire
$ formant employee --id AI-EMP-0042 --rollback --to v41
✓ traffic v42 → v41 · 100% in 38 s · no calls dropped
→ reason logged: "tone guardrail breach · bucket-1" · approver: Head of Collections
$ formant employee --id AI-EMP-0042 --retire --keep-data
! role retired · recordings, transcripts, weights stay in your tenant
audit both actions appended to the performance file

Illustrative run. replace with a live rollback log

The roll-up

One employee, six pillars.

Each pillar of FormantAI is one part of the same hire. Read them in any order — they all describe the same person.

Governance

Permission set

What the employee may see, say and do — field-level, with approval tiers. Deny rules win.

governance
Personalization

How they speak

Language, name, pace, tone by history — from governed fields only, never inferred.

personalization
Memory

What they remember

The customer never repeats themselves. Redacted, retained per journey, forgettable on request.

memory
Coworkers

Their seat on the floor

Warm transfer with a brief, whisper to a human, after-call work done. Humans take escalations.

coworkers
Digital twins

Who trained them

A twin of your best agent; a twin of your customers to rehearse against; a twin of the journey.

digital twins
Learning + evaluation

Their weekly review

Graded every call, improved every week, signed by a named human. The performance file.

learning
FAQ

What your risk team will ask.

Who is accountable when the AI employee gets a call wrong?
The named manager on your side, exactly as for a human agent — and the file shows what the agent was permitted to say, which version said it, and who approved that version. FormantAI is accountable for the engine behaving as the file says.
Does the customer always know they are talking to an AI?
Yes. The AI disclosure is a frozen line on the ID badge: verified verbatim on every call and on every release, and never part of what the model learns. The same applies to the recording disclosure and grievance line.
Can it ever act outside its permission set?
No. The permission set is enforced by the governance engine on every turn, not by the prompt. Deny rules win over allow; anything not granted — a fee waiver, a settlement offer — is routed to a human with a brief.
What happens to the data when we retire the role?
Nothing leaves. Recordings, transcripts, labels and fine-tuned weights live in your tenant, in the region you chose, and are retained or deleted on your schedule under DPDP. Retiring the role is a permission change, not a data transfer.
Under NDA

Read one employee file.

Open an employee file
Readiness

Draft the role you would hire

Get a Call Readiness report