01 · What breaks
Design for the failure modes, not only the happy path.
Placement figures draw scrutiny from funders and auditors, so shortcuts in verification come back later as disputes. Four mistakes cause most of the trouble.
Contacting an employer without permission
Calling an employer can reveal that a worker came through a programme, or expose them where the job is informal or the relationship is fragile. Ask the graduate first, record the permission on the case, and leave the employer out whenever the graduate declines or a counsellor flags a risk.
Informal work treated as unverifiable
Shop assistants, home-based workers and the self-employed rarely have letters or HR teams. Marking them unverifiable pushes programmes towards formal jobs only. Define acceptable evidence for them in advance: customer or revenue records, a joint certificate with the employer, or repeated self-report over several months.
Nonresponse marked as false
A graduate who does not answer has not lied about a job. Logging silence as not placed or as a false claim damages trust and the record. Keep a separate unreachable state, log each attempt, and let a person decide what happens after the programme's attempt limit.
AI resolving wage or identity conflicts
When a salary slip shows a different name, employer or take-home pay from the claim, the mission must not choose which is right. It shows both values side by side with the source document and routes the case to a reviewer. Silent auto-correction is how a placement record stops being evidence.
02 · Minimum viable model
Keep the claim, the evidence and the verdict as separate records.
The ladder below runs from the weakest evidence to the strongest. Each rung proves something specific and none proves everything, which is why a DozAI verification case stores the evidence type next to its state instead of a single yes or no.
| Evidence | What it shows | What it does not show |
|---|---|---|
| Self-report by call or WhatsApp | The graduate says they are working | Employer, pay or start date, independently |
| Offer or appointment letter | An employer intended to hire | That the graduate joined |
| Day-one joining call | The graduate turned up on the first day | That they stayed, or what they are paid |
| First salary slip | Pay was issued; CTC and take-home pay can both be read | That the slip matches the claim, until a person checks it |
| Employer confirmation | The employer recognises the worker | Anything reliable, if identity was not confirmed first |
| Enterprise records over time | A self-employed graduate is earning | A stable income, if only one month is seen |
03 · Implementation sequence
Move through reviewable gates.
Verification moves faster when the evidence rules are settled before the first call. Each gate below leaves a record a reviewer can check later.
Define acceptable evidence
Write down, per funder, which evidence makes a case verified: an offer letter, a salary slip or a number of days in the job. Keep one outcome record per graduate with a funder tag and funder-specific rules, so the team stops running a separate tracker for each donor.
Capture permission for contact
Ask the graduate, in their language, whether the programme may contact their employer and for what purpose. Record yes, no or only with notice, with the date. A refusal is respected even when it leaves the case at self-reported.
Verify identity before disclosure
Before sharing anything, confirm that the person on the line is the graduate, using details already on the record, and that the employer contact is a real HR person or supervisor. Disclose nothing about the programme or the graduate until both checks pass.
Collect evidence across channels
Request documents over WhatsApp, take answers by voice call and accept uploads from field staff. The mission reads CTC and take-home pay from salary slips, shows a confidence score for each field and waits for a person to confirm before saving. Slips never go into public AI tools.
Route conflict and low confidence
Send three kinds of case to review, each labelled: a record conflict, where the data on file is wrong; disputed evidence, where two sources disagree; and nonresponse. Each needs a different action, so one mixed queue slows the reviewer and hides the real disputes.
Approve outcome and preserve trail
The reviewer chooses approve, edit, ask for evidence or reject, and gives a reason. The approved state, evidence links, reviewer and date are locked together. Evidence links stay restricted to the roles that need them, so salary slips do not travel in shared spreadsheets.
As of October 2026, DDU-GKY training partners must place at least 70% of trained candidates. Since the DDU-GKY 2.0 guidelines (May 2025), a placement counts after six months of employment, with wages at or above the applicable minimum wage. The programme SOP accepts salary slips plus proof of salary payment, such as a bank transfer or passbook entry, or a joint employer-candidate certificate where the employer has no HR department. Confirm the latest guidelines before relying on these norms.
04 · Evidence
Measure whether the work became clearer.
A verification mission is working when more cases reach a clear state with evidence attached, and reviewers rarely have to reverse it. Read these measures by centre and trade, next to the placement funnel in the youth employability playbook.
05 · Responsible automation
Let systems prepare. Keep consequence with people.
Systems can support
- Requesting documents and answers in the graduate's language
- Reading CTC and take-home pay from salary slips
- Sorting cases into evidence-based states
- Labelling record conflicts, disputes and nonresponse
People should decide
- Final verification state
- Disputed evidence
- Employer contact exceptions
- Support after adverse findings
06 · Common questions
Practical questions teams ask.
Questions placement teams and MEL leads ask about verifying jobs with an AI mission.
Can AI verify job placements for a skilling programme?
It can do most of the legwork. A mission can call or message graduates, request salary slips or letters, read the key fields and sort each case by the evidence it holds. It should not declare a placement verified on its own. A person approves the final state, resolves conflicts and decides what meets each funder's definition of a placement.
What counts as proof of placement?
It depends on the funder, so write the rule down for each one. Common evidence runs from self-report, through an offer or appointment letter and a day-one joining call, to a salary slip and employer confirmation. Self-employed graduates need different evidence, such as customer or revenue records over several months. Record which evidence type supports each case.
Is an offer letter enough to count a placement?
Usually not. An offer letter shows that an employer intended to hire, not that the graduate joined or stayed. Candidates sometimes decline once they learn the shift, the commute or the deductions. A day-one joining call and a salary slip within the first month or so are much stronger, and many funders ask for proof of employment over time.
Should a programme contact employers to confirm placements?
Only with the graduate's permission, recorded on the case, and only after confirming who is on the other end of the call. Some graduates have good reasons to decline, for example in informal or fragile jobs. In those cases accept other evidence, or leave the case at self-reported, rather than overriding the graduate's choice.
What happens when a graduate does not respond to verification?
The case moves to an unreachable state after the programme's attempt limit, with every attempt logged. It is not marked as a false claim or as not placed. A coordinator decides the next step, such as another channel or a field visit, and the case stays visible in reports so nonresponse can be tracked by centre and subgroup.