AI Overviewfor the morning.
A daily AI insight and action plan inside an automotive dealer group’s internal dashboard, so dealer leaders know which KPI is behind, why, and who does what today.
- 01I came up with the concept: the insight logic and the UI.
- 02I mapped every use case across 4 roles and their KPIs: 99 in total.
- 03I spent time with frontliners, and it shaped the design.
SPK tertinggal karena follow-up leads lambat
- 42 dari target berjalan 48; butuh 3 SPK/hari.
- 23 leads belum di-FU, 9 lewat SLA.
- FU ini bisa menutup ±4 dari gap 6.
- Budi & Dedi FU leads lewat SLA, sebelum 12.00.
- SS bagi 5 prospect test drive ke AE.
DO tertahan di leasing
Many reports,one morning.
The client is a national car dealer network in Indonesia, run as Head Office › Region › Area › Branch. Its leaders work from an internal analytics dashboard, with reports for sales, aftersales, finance and customers.
AI Overview is a feature inside that dashboard. Every day it reads the reports for a user’s KPIs and writes a short insight and action plan, for four roles: Branch Heads (Kacab), Workshop Heads (Kabeng), Area Heads (ABH) and Head Office.
- TimeLeaders had plenty of reports but little time to diagnose them each morning.
- Generic AIThe first AI output gave soft advice (“Saran saya…”) and repeated the same root cause across several KPIs.
- No contextNumbers came without a comparison. Actions came without an owner or a deadline.
- No specThe dev team had no clear definition of a “good insight”, or of how the feature should look.
I went to the branchesbefore the prompt.
I spent time with [who: e.g. Branch Heads, Workshop Heads, Sales Supervisors, Service Advisors] at [X branches], by [how: e.g. sitting in morning briefings, interviews, watching how they read the dashboard].
What I saw became the standard the playbook asks the AI to meet. Every rule below traces back to how leaders actually run their day.
Key decisionAn insight that stops at “KPI X is behind, please check” doesn’t help anyone.
Leaders act in the morning briefing.
Advice without a person and a time goes nowhere once the briefing ends.
→ So every action names who, what (a countable object) and whenThey want the point first.
No opening sentences, no “saran saya”, no hedging. Encouragement is welcome, but not mixed into the analysis.
→ So headlines are conclusions, and encouragement lives only in the greetingNames point to a PIC, not a verdict.
Naming the person behind a number helps a leader know who to ask, as long as it never reads as a judgment.
→ So names appear only to assign the action99 use cases,4 roles, one map.
27
Whole branch: sales, aftersales overview, cost, receivables, NPS.
23
Unit entry, CPUS, revenue, gross profit, down to SA and process.
24
The same KPIs, read across the branches in an area.
25
Every KPI, compared across regions, areas and branches.
A signal can only be named as the cause of a KPI if it leads that KPI. That rule alone stopped the AI from blaming one root cause for everything. Simplified here; the full tree lives in the playbook.
| KPI | Detailed for the owner | Summary for |
|---|---|---|
| SPK, DO | Kacab: per sales exec, funnel stage, outstanding SPK status, stock | ABH, HO: per branch, area, region |
| Unit Entry, CPUS, Revenue, Gross Profit | Kabeng: per SA, bookings, reminders, WIP, lead time | Kacab: status, main cause, action routed through the Kabeng. ABH, HO: per branch, area, region |
| NPBT, OPEX, Absorption Rate, AR | Kacab: cost lines, margin, receivables per customer | ABH, HO: per branch, area, region |
Same KPI, different depth by role. The map lives in a master Excel, the playbook and a stakeholder deck.
One KPI,one story.
The AI only uses real dashboard data. No internal targets are hard-coded; comparisons are relative: last month, last year, the 4-week average, the area average.
Each insight is a story. The headline is the conclusion. Up to three points follow: status, main cause, impact. Then up to three direct commands, each with who, what and when, ending with “check tomorrow: this number should move from X to Y”. If nothing is wrong, it is one line: on track.
Key decisionThe user picks what matters. The system makes sure what’s worst is never missed.
- 01Your picks come first, in the order you chose them.
- 02Empty slots fill with the KPIs furthest behind target, ranked by % of target (flipped for lower-is-better KPIs).
- 03The same root cause is never shown twice.
- 04Empty slots are never padded with on-track KPIs.
Every elementearns its place.
Sets the day: working days left. Encouragement lives only here (Finding 02).
KPI tags, a conclusion headline, up to 3 bullets and a separate action box (Finding 01).
Auto-filled cards say “picked automatically · behind target”, so nothing appears without a reason.
A one-line on-track row, a warning slot and a status strip for KPIs outside the 3 slots. Cards collapse; a timestamp shows the daily refresh.
The whole morning, on one screen.
The AI summary sits on top of the reports a Branch Head already opens every day.
One line to start the day.
A conclusion first: what is behind, and where the problem actually sits.
A number, with what it takes.
Status against the running target, plus the pace needed for the working days left.
Then a task with an owner and a time.
Who and when on every action, with a check to mark it done.
The cause, not the symptom.
Orders are up but deliveries are down: the hold-up is leasing and stock, not leads.
Quiet about what’s fine.
Improved items are shown, unchanged ones collapse. Colour always comes with a label.
The whole branch.
Deliveries behind target, with the hold-up after the order: leasing and stock.
Down to each advisor.
Unit entry behind, closed by tomorrow’s bookings and customers due for service.
Regions side by side.
The same KPIs, compared across regions, areas and branches.
Explored, then redirected.
Prototyped here; in the final playbook, finance executes branch actions instead of getting its own summary.
These are the prototype screens from an earlier exploration, live and anonymized. The 3-slot layout in the mockup at the top is the concept the playbook now specifies.
“Saran saya, MRA dan Service Advisor menghubungi pelanggan… fokus meningkatkan kedatangan unit.”“My suggestion: MRA and Service Advisors contact customers… focus on increasing unit arrivals.”
- Soft advice, no owner
- No number, no comparison
- No deadline, no way to check
SPK tertinggal karena follow-up leads lambat ConclusionSPK is behind because lead follow-up is slow.
- SPK MTD 42 dari target berjalan 48 (87,5%); butuh 3 SPK/hari di 8 hari kerja tersisa.Orders month-to-date: 42 of 48 so far; 3 a day needed over the 8 working days left.
- 23 leads belum di-FU, 9 lewat SLA, terbanyak di Budi (4) dan Dedi (3).23 leads not followed up, 9 past SLA, mostly Budi’s (4) and Dedi’s (3).
- Dengan konversi biasa 18%, FU leads ini menutup ±4 dari gap 6 SPK.At the usual 18% conversion, these follow-ups close about 4 of the 6-order gap.
- Minta Budi dan Dedi FU leads lewat SLA sebelum jam 12.00.Have Budi and Dedi follow up past-SLA leads before 12:00.
- Minta SS bagi 5 prospect test drive ke AE di briefing pagi.Have the Sales Supervisor assign 5 test-drive prospects to sales execs at the morning briefing.
- Cek besok: Not Yet FU turun dari 23 ke 0.Check tomorrow: not-yet-followed-up drops from 23 to 0.
Business context, data catalog, KPI driver tree, insight and action-plan rules, the 3-slot concept.
Across 4 roles, plus worked examples per KPI.
The classification behind the map, and the stakeholder slides.
Greeting panel, insight card, action box, KPI picker, on-track and warning states.
The feature is rolling out to branch, workshop, area and head-office users, built from the Foundation playbook as its spec.
Whether leaders act on the plan the same morning, and whether the “check tomorrow” number actually moves the next day.
- Time with frontliners changed what “useful” meant.On paper, a good insight was accurate. In a branch, it was one a leader could act on before the briefing ended.
- Constraints made the AI more trustworthy.“Only what the dashboard shows” cut use cases, but every number it writes can be checked against a report.
- An AI feature needs a writing system.Not just a prompt: a use case map, a driver tree and rules for what a good sentence looks like.
- Immersion photos or notes, anonymized
- The full use case map (roles × KPIs) from the master Excel
- The complete KPI driver tree
- Real names, branch data and numbers blurred or replaced with dummy data