Issue 01
High manual effort
Every cycle started with gathering data by hand from fragmented sources: plant-to-plant dependencies, packaging flows, lane data.
WorkAI Network Optimizer
AI Network Optimizer (0→1) | Unilever
Designing an AI logistics planning tool that planners could trust, even when the AI couldn't explain itself.
Network Optimizer is an AI planning tool for Unilever India's transportation planners. It replaces the spreadsheet process for costing inbound shipping routes.
- Enterprise
- AI
- Logistics
- FMCG
- 0→1
Tools
Figma
FigJam
Excel
Word
Outcome & impact
- Measured40–50%Shorter planning cycle (7 → 4 days)
- Measured60%Less rework
- Projected8–10%Cost savings identified
TLDR
Got a minute?
Planners already knew the routes. They couldn't prove which one was cheapest. I designed a 0→1 tool that shows them the cheapest option and leaves the final call with them.
With it, planners consistently uncover cost-saving opportunities, refine route decisions with confidence, and produce execution-ready shipping plans with far less manual effort.
Problem
Planners built inbound shipping plans in spreadsheets every month and quarter. Each cycle took days of data gathering and validation. It still couldn't tell them whether a cheaper route existed.
My role
I led UX for this 0→1 MVP as the only designer, from early alignment to developer-ready designs. I worked with Product, Engineering and Data Science to turn optimization logic into a workflow planners could trust, within real limits on model explainability and data readiness.
Context
What was it about?
- What it is
- An AI planning tool for Unilever India’s transportation planners, in place of their costing spreadsheets.
- What planners get
- An automated baseline, a map of the network, and scenarios to compare before locking the month’s plan.
- Why it matters
- Once the month’s network is set, every extra kilometer repeats as cost across routes, regions and volumes.
The opportunityFind more efficient routes and show the true cost of every shipping decision, without heavy manual planning.
Why · the problem
Every month, planners rebuilt the same plan from scratch
The network it covers
- Processing plants
- 15+
- Depots
- 25+
- Suppliers & packaging units
- 80+
- Lanes
- 300+
- Annual inbound logistics spend
- ~₹150 cr
Issue 02
No memory between cycles
The same workflow ran monthly for execution and quarterly for forecasting. Nothing carried over, so nothing got faster.
Issue 03
No way to see a better route
Formula-driven costing was error-prone and hard to audit. The plan got finished, but it never showed whether a cheaper route existed.
fx
=IFERROR(VLOOKUP(B12,'Lanes (2)'!A:F,6,0)*D12,"CHECK")| A | B | C | D | E | F | |
|---|---|---|---|---|---|---|
| 1 | Lane | From | To | Qty (t) | Rate / t | Cost |
| 11 | L-208 | PL-03 | DP-11 | 1,240 | 2,180 | 27,03,200 |
| 12 | L-211 | PL-03 | PK-02 | 860 | #N/A | CHECK |
| 13 | L-214 | PK-02 | DP-07 | 1,050 | 1,940 | #REF! |
| 14 | L-219 | PL-05 | DP-11 | ?? | 2,410 | #VALUE! |
| 15 | L-221 | PL-05 | DP-04 | 720 | 2,050 | 14,76,000 |
| 16 | L-226 | PK-04 | DP-04 | 1,310 | 1,860 | 24,36,600 |
Plan_JunPlan_Jun (2)LanesLanes (2)Depots_OLDPkg flows+
RE: RE: FW: updated depot volumes (use this one)
Plant ops · 11:48 PM
Attached the latest. Please ignore the file from Tuesday, the packaging numbers were wrong.
depot_volumes_v3.xlsx
EXPORT_ZMM_INB_0625.csv
MATNR;WERKS;LGORT;MENGE;MEINS 10045;PL03;DP11;1240;TO 10045;PL03;PK02;0860;TO 10071;PL05;;1050;TO 10071;PL05;DP04;0720;TO
- Inbound_Plan_v6.xlsx
- Inbound_Plan_v7_FINAL.xlsx
- Inbound_Plan_v7_FINAL_revised.xlsx
- Lane_rates_Q2 (1).xlsx
check row 214 again??
Users
Planners knew the network. They couldn't prove the best route.
They plan the inbound network for Unilever India's supply chain. Their plans go to a logistics manager, who reviews and approves each one and needs confidence in the AI output.
Users
Transportation and logistics plannersUnilever India · Supply chain operations
- Monthly planning
- Cost forecasting
- Operational feasibility
- Geography
- IndiaPrimarily India-based teams managing the national supply network
- Planning cycle
- Monthly and quarterlyExecution plans every month, forecasts and pre-plans before each quarter
- Network scope
- End-to-end routingPlants, depots, suppliers, packaging units and inbound legs
Every decision sits between three pressures
- Cost controlEvery routing decision carries a cost that repeats across routes, regions and volumes.
- Planning speedPlans must be ready before each month or quarter. Delays cascade into operations.
- Delivery reliabilityRoutes must work in practice, not just be cheapest on paper.
Proto-persona
Arjun Mehta
Senior Transportation Planner · Unilever India
- Age 34
- Mumbai, India
- 8 yrs experience
- Supply chain / Logistics
Persona typeThe Meticulous Planner
Built from recurring planner feedback, not interviews.
I already know the network. I just need a faster way to prove which route is actually the best, and to show the numbers to my manager.
Arjun has managed inbound logistics across Unilever India for close to a decade. He builds monthly shipment plans and quarterly forecasts by routing flows between plants, packaging units, depots and suppliers. He knows the network deeply, but most of his time goes into calculations he knows could be automated.
His goals
- Finalize monthly plans quickly, without spending days on spreadsheet calculations.
- Identify the most cost-efficient route scenario with confidence.
- Reduce rework caused by errors found only after plans are submitted.
- Reuse past plans as a starting point rather than rebuilding from scratch.
His pain points
- Data is scattered across emails, SAP exports and stakeholder inputs. Collecting it takes hours before planning starts.
- Spreadsheet costing only validates. It never shows whether a better route exists.
- Errors surface after export, which forces last-minute rework under deadline pressure.
What he needs most, ranked
- Cost visibility
- Speed of output
- Scenario comparison
- Editable scenarios
- Network map view ranked last. See decision 01
Tools he uses today
- Microsoft Excel
- SAP
- WhatsApp (coordination)
- Manual data validation
Core tensionCost control, planning speed and delivery reliability. Arjun balances all three every cycle, with no tool that shows him the trade-offs clearly.
Research
No formal study. The evidence came from four live planning cycles.
This MVP had no dedicated research phase. The planning cycles themselves were the research: the same friction came up every month and every quarter.
Source 01
Recurring planner pain points
The same friction surfaced in every monthly and quarterly cycle, which gave a clear signal on what mattered most.
3 monthly cycles and 1 quarterly cycle
What it told usWhich steps were most effort-heavy and error-prone.
Source 02
Stakeholder workshops
Cross-functional alignment sessions with stakeholders across the supply chain.
3 sessions with planners, the logistics manager, the product owner and data science
What it told usWhere planning broke down across the supply chain.
Source 03
Operational reviews
6 spreadsheet audits of past monthly plans
What it told usMost errors started in manual lookups between tabs, and validation alone added 2–3 days to every cycle.
This gave us clear direction on scope quickly, without waiting for formal research. The recurring nature of the pain points meant the signal was already rich and validated by operational reality.
Before
Planners spent more time verifying than optimizing
- Normal01
Collect network inputs
Gather data from plants, depots, suppliers and lanes.
PainData scattered across emails and SAP exports, in inconsistent formats.
- Normal02
Build route plans by hand
Map lanes: processing → packaging → depots.
PainNetwork complexity makes this slow and repetitive every month.
- High friction03
Calculate costs in spreadsheets
Compute inbound costs with formula-driven sheets.
PainError-prone and hard to audit. Errors stay hidden until downstream.
- High friction04
Validate repeatedly
Cross-check outputs against source data several times.
PainMore time verifying than optimizing.
- Normal05
Finalize the monthly plan
Lock the execution plan and the quarterly pre-plan.
PainNext month starts from scratch, with no reuse and no learning.
- longer turnaround
- 35–45%
- of validation loops before a plan was ready (step 04)
- 2–3 days
- route optimization: alternative routes were never costed
- 0%
- restarts from scratch
- ↻ Monthly
Steps 03 and 04 were the bottleneck. Planners spent more time checking numbers than improving them.
Core insight
The problem wasn't routing. It was proof.
Planners didn't need help finding routes. They already knew them. The question they had to answer every month was:
“Are we shipping in the most cost-efficient way possible, given our real network constraints?”
They needed to answer it with confidence, and show the numbers to their managers.
Constraints
Explainability and lane‑level costs had to wait
To ship a reliable 0→1 MVP on time, we focused on the core optimization workflow. Two capabilities needed more mature models and data, so we deferred them.
Guiding principle, from the stakeholder workshopDeliver core value first, defer complexity. Ship fast, learn, then expand.
Deferred 01
AI explainability
- Intent
- Show planners why the AI chose a route.
- Constraint
- The optimization layer didn't consistently expose decision-level reasoning in a structured, trustworthy way.
- The decision
- Planned for a later release, once explainability data is stable.
Deferred 02
Lane-level cost breakdown
- Intent
- Show route-level and leg-level cost drivers.
- Constraint
- Needed consistent lane mapping, a mature data model and performance tuning for large datasets.
- The decision
- Deferred to keep the MVP simple.
Deferring explainability created the central design problem
How do you get planners to trust a recommendation the AI can't explain?
Designing for trust
If the AI can't explain itself, let planners check its work
✕The goal was not
To replace planners or automate their decisions away.
✓The goal was
To cut the time spent on repetitive calculation and validation, and give it back to making the best decision.
I designed three ways for planners to check the AI against what they already knew.
AI proposesThe best-cost plan
Check 01 · baseline
Anchor it to their own numbers
The AI's best cost always sits next to the planner's baseline.
Check 02 · map
Show it on the network they know
On a map, an odd or unrealistic route stands out in seconds.
Check 03 · edit
Let them change it
Any value can be edited, and the cost updates instantly.
Planner decidesLocks the plan
Three design pillars
Pillar 01 · Speed
Less time calculating and validating
Pillar 02 · Optimization
Reveal the best-cost scenario
Pillar 03 · Control
Keep planners' judgment in the plan
Key decisions
Three calls that shaped the product
The stakeholder workshop produced a long list of ideas. These three decisions shaped the product most, and each one had an alternative we chose against.
Decision 01
The tensionPlanners ranked the map last among their needs. Cost visibility came first.
Alternative
A cost table or dashboard as the home screen (confirm)
Chosen
The map is the working canvas. The three-scenario cost strip is pinned in the header on every screen.
Why?Cost never leaves the screen, so the top need is always met. The map adds what a table can't: a fast way to sanity-check AI routes against a network planners know by heart.
Decision 02
Alternative
Show only the AI-optimized plan.
Chosen
Baseline, Edited and Best cost side by side.
Why?A single AI number asks for blind trust. The same number next to the planner's own baseline makes the gap concrete and checkable.
Decision 03
Alternative
Export to Excel, edit, re-upload: the planners' existing habit.
Chosen
Edit any value in place and recalculate instantly.
Why?Every round trip through a spreadsheet was a chance for errors to creep back in, which was the exact problem we were solving.
From pain to decision
Every planner pain from the old cycle, the decision that answers it, and where it lives in the product. It condenses the stakeholder workshop board.
| Planner pain | Design decision | Where it lives |
|---|---|---|
| Days spent collecting and costing data | Upload baseline data, get costs instantly | Upload and instant cost |
| No way to know if a cheaper route exists | AI best-cost scenario shown against the baseline | Three-scenario view |
| Network invisible inside spreadsheets | Plants, depots and routes on a map | Map view |
| Hard to explain the "why" behind numbers | Drill down to any depot or plant | Location drill-down |
| Errors found only after export | Edit in the tool, recalculate instantly | Edit and refresh |
| No way to weigh options | Scenarios side by side | Scenario comparison |
Solution
One planning cycle, start to finish, in one tool
Six features, in the order a planner moves through a cycle. Each one is tagged with the pillar it serves, so the thread runs on to the outcomes.
01Upload
SpeedUpload data. See costs instantly. No spreadsheets.
Planners upload baseline datasets and see the overall inbound shipping cost right away. Manual spreadsheet costing is gone.
Impact → Less time on repetitive cost calculations. Planners move from collecting data to making decisions faster.
02Map view
ControlThe network, finally visible, not just calculated.
A map plots shipping plans across India, with plants, depots and routes as real network points and clear cost differences between them.
Impact → High-cost lanes are easier to spot, and route patterns are easier to trust.
03Three-scenario view
OptimizationThree scenarios. One view. Cost optimization made visible.
Baseline, Edited and Best cost sit together in the header on every screen: the uploaded cost, the cost after manual edits, and the AI-optimized cost.
Baseline
₹12.20 crUploaded planEdited
₹11.60 cr↓ 4.9% vs baselineBest cost
₹11.10 cr↓ 9.0% vs baselineImpact → Faster comparison, more confidence in the AI output, fewer manual recalculations.
04Location drill-down
ControlEvery depot. Every plant. Drill down to check at the source.
Planners can open any depot or plant (processing, new or closed, packaging) and check the details behind the numbers.
Impact → Less manual cross-checking, and faster answers when planners need to explain the "why".
05Edit and refresh
ControlEdit the scenario. Refresh the cost. No rebuild.
Planners edit baseline data directly in the tool and get an updated Edited-scenario cost instantly.
Impact → Planners can test real-world adjustments without rebuilding the plan from scratch.
06Scenario comparison
OptimizationPut scenarios side by side. See what changed.
A Compare view shows the total cost of several scenarios side by side, broken down by cost type.
Impact → Planners choose between outcomes backed by evidence, not just between routes.
Outcomes
Faster cycles, less rework, and a savings gap planners could finally see
Each outcome maps to one design pillar, and says whether it was measured or projected.
SpeedMeasured
40–50%Shorter planning cycle (7 → 4 days)Instant cost generation on upload replaced the manual spreadsheet loop that used to take days.
MethodHow it was measured, and over how many cycles
OptimizationProjected
8–10%Potential cost savings identifiedThe AI best-cost scenario surfaced savings planners couldn't see before, and made the gap between the current plan and the optimal one visible for the first time.
MethodGap between the baseline and the best-cost scenario across X cycles, before planner edits. Realized savings, if known
ControlMeasured
60%Less reworkMap visibility and editable scenarios let planners check and adjust before committing, which cut the post-submission corrections that used to restart the cycle.
MethodPost-submission corrections per cycle, before and after. How they were counted, and over how many cycles
A quote from a planner, the product owner or a supply chain lead about the tool's impact
What happened next
- Launch
- Month Year, used by X planners across X planning cycles
- AI explainability
- Shipped, in progress or still deferred
- Lane-level cost breakdown
- Shipped, in progress or still deferred
- Beyond the MVP
- Any expansion: other regions, longer-term forecasting, new users
Reflections
What I'd keep, and what I'd change
What I'd keep
0→1 MVPs demand ruthless prioritization
I learned to prove the core value early instead of trying to solve every edge case in the first release. Choosing what to defer, and designing around that gap, shaped this product as much as any screen did.
Speed and alignment matter more than perfection
Quick stakeholder alignment and clear milestones moved the project forward in a complex enterprise setup.
What I'd change
I'd measure the baseline before launch
Some outcomes here are estimates because the old cycle time wasn't tracked systematically. Next time I'd record the "before" state in week one.
I'd push for a minimal cost-driver view in the MVP
Comparison, the map and editing gave planners enough to check the AI, but not to see why it chose a route. Even the top three cost drivers per route would have answered most of those questions without full explainability.
Final thoughts
This project made me confident designing 0→1 enterprise products where complexity is high, timelines are tight and the AI can't do everything yet. The work was less about screens and more about shaping a workflow, aligning Product, Engineering and Data Science, and turning an optimization model into something planners would actually use.
More work
Thanks for reading.
If you'd like more detail, get in touch, or see my other work below.
Open for opportunities
Building something complex? Let's make it clear.
I partner with teams to turn complex problems into clear, human-centred solutions.