01 · What breaks
Design for the failure modes, not only the happy path.
Retention numbers go wrong quietly. The usual causes are not technical: questions drift between rounds, hard cases fall out of the base and a normal career move gets logged as a loss.
Different questions at each interval
If day 90 asks "are you still working?" and day 180 asks "are you with the same employer?", the two rates measure different things and the trend means nothing. Fix a short core set asked in every window, then add window-specific probes, such as pay or progression, after it.
Unreachable graduates removed
Graduates who stop answering are more likely to be out of work. Drop them from the base and retention rises for the wrong reason. Keep them in the cohort with an unreachable state, the attempt log as evidence and an alert to a coordinator, and report coverage next to the rate.
Job change counted as failure
A graduate who moved to a better-paid job in the same trade has been retained in work. So has one whose employer name changed after a contractor switch at the same site. Record changed job as its own state, and correct the employer record instead of logging that the graduate left.
Sensitive reasons shared with employers
Exit reasons such as harassment, unsafe conditions or ill health stay in the graduate's restricted record. Employer feedback uses themed counts only, never one person's words. The mission flags these answers for a trained staff member and never repeats them in any message to an employer.
02 · Minimum viable model
Model each check-in as its own episode, tied back to one placement.
In a DozAI mission each graduate is a case and each window is an episode inside it. The six records below keep the answer, its evidence and the reviewer's call together. The outcome states in the table keep free text out of status fields.
| Outcome state | What it means | Counted as in work? | Next action |
|---|---|---|---|
| Same job | Same employer as the baseline, including a promotion there | Yes | Wait for the next window |
| Changed job | Working for a different employer or in a new role | Yes, retained in work | Update the employer record and note whether the trade matches |
| Not employed | Left work and not yet in a new job | No | Personal call and a re-placement task |
| Education or training | Left work to study or train further | Reported separately | Record the course; no re-placement push |
| Unreachable | No answer after the programme's attempt limit | No, and stays in the base | Coordinator alert; retry in the next window |
| Unclear or disputed | Answers conflict with the baseline or with each other | Pending review | Reviewer packet with both versions |
03 · Implementation sequence
Move through reviewable gates.
Set the rules before the first message goes out. Most of the work is deciding what counts, who is contacted and how. The mission then repeats it the same way for every graduate in the cohort.
Define follow-up windows and states
Generate windows from the confirmed joining date, not the batch end date, on whatever schedule the funder sets, such as 90 and 180 days. Agree the outcome states as a fixed dropdown, and write down what retained means for each funder before any contact starts.
Verify safe contact channels
Before outreach, check that the number on file still belongs to the graduate and that their consent covers follow-up. A shared family phone, or a number now held by an employer, needs a different approach. Record opt-outs so the mission stops at once and never retries them.
Schedule multilingual outreach
A common pattern is WhatsApp first, then a voice call in the graduate's language within agreed hours. As of October 2026, WhatsApp Business rules allow messages only to people who opted in and, outside the 24-hour window after their last message, only pre-approved templates.
Collect comparable evidence
Keep the instrument to about four questions: still with the employer, current employer and role, income band, and whether they need help. Store the transcript or thread against the episode. Where a funder needs proof, ask for a recent salary slip rather than accepting a spoken figure.
Route support needs and conflicts
Any lost-job or need-help answer creates a personal call by a counsellor within a time the programme sets, two working days being a common default, plus a re-placement task linked to the first placement. Conflicts with the baseline go to a reviewer, not back to the graduate.
Report retention with nonresponse
Freeze a snapshot when each window closes, so later edits cannot change a filed report. Report two numbers together: graduates in work as a share of everyone placed, and follow-up coverage. Break exits down by employer, trade, shift and centre to see where people leave.
Under the DDU-GKY 2.0 guidelines (May 2025), placements are tracked for 12 months, and a candidate counts as placed once they complete six months of employment, which need not be continuous and can be wage, self- or gig employment. Those are the norms as of October 2026; check the latest guidelines before relying on them. Other funders use their own windows, so keep one episode schedule per funder rule on the same graduate record.
04 · Evidence
Measure whether the work became clearer.
Judge the mission on whether the retention figure became honest and useful, not on the number of calls made. The worked example shows why one number is never enough. The youth employability playbook covers the wider placement funnel.
These numbers are invented for illustration. A programme placed 100 graduates. At day 180 the mission reached 85 of them: 68 were in work (same or changed job) and 17 were not. The other 15 stayed unreachable after the attempt limit. Report 68 of 100 placed in work (68%), with 85% coverage beside it. Quoting 68 of 85 (80%) describes only the graduates who answered.
05 · Responsible automation
Let systems prepare. Keep consequence with people.
Systems can support
- Generating check-in tasks from joining dates
- Sending opted-in WhatsApp messages and retrying calls
- Transcribing answers and proposing an outcome state
- Flagging help requests and baseline conflicts
People should decide
- Responding to distress or exploitation
- Resolving employment conflict
- Offering programme support
- Approving the reported status
06 · Common questions
Practical questions teams ask.
Questions programme heads and placement teams ask before handing follow-up calls to an AI mission.
Can an AI mission run 90-day and 180-day follow-up calls for a skilling programme?
Yes, for the repetitive part. The mission can schedule check-ins from joining dates, message and call graduates who agreed to contact, ask a fixed set of questions in their language and propose an outcome state. Staff still approve the status that gets reported and personally call anyone who has lost a job or asks for help. Start in draft mode, with a person checking every classification.
What questions should a retention follow-up ask?
Keep a short core that never changes between windows: are you still with the same employer, who is your employer and what is your role, which income band are you in, and do you need any help. Add window-specific probes after the core, such as progression at six months. Changing the core questions breaks the comparison between day 90 and day 180.
How should unreachable graduates be counted in retention reports?
Keep them in the denominator as unreachable and never remove them. People who stop answering are more likely to be out of work, so excluding them inflates retention. Report retention as a share of everyone placed, with follow-up coverage beside it, and keep the attempt log as evidence that the programme tried within its agreed contact limits.
Does changing jobs count as a retention failure?
No. A graduate who moves to another employer, especially in the same trade or for better pay, is still in work. Record it as a separate changed-job state so it can be counted on its own. Also look for record conflicts, such as a contractor change at the same workplace, which need the employer record corrected rather than a left-job entry.
How do you protect graduates' privacy during follow-up?
Contact only people who consented to follow-up, on the channels they chose, and stop at once when someone opts out. Keep contact details and sensitive answers, such as harassment or health reasons for leaving, in a restricted part of the record. Employers should see only themed, aggregate exit patterns, never an individual graduate's reasons or words.