Stuck onboarding 25 partners a month. For a year.It wasn’t the agents. It was the tools.

Every theory anyone had heard for that gap put the blame on the agents — undertrained, undermotivated, working a harder segment of the market. None of it survived five minutes of actually talking to the people doing the job.

So the brief wasn’t “redesign the tool.” It was narrower and harder than that: find out what’s actually broken, prove it with evidence, and only then decide what — if anything — to build.

That’s the process I ran — end to end, from a fuzzy productivity problem to a validated platform tested across seven markets. I followed the double diamond: diverge to understand the problem, converge on a diagnosis, diverge again on solutions, converge on what actually shipped.

Four phases, one continuous loop. Everything below follows this path — research and diagnosis first, design and validation after.

01 · Discover

Diverge — exploring the problem

Where the friction actually lived

Before drawing a single screen, I needed to see the problem with my own eyes, not read about it in a process doc. So I watched how sales reps, QC agents, and content agents actually spent their day inside the existing Salesforce ecosystem — where they slowed down, doubled back, or worked around something broken. Then I looked outward: how tools built for exactly this kind of work — pipelines, dashboards, handoffs — had already solved the same problems for other teams.

Three roles, one shared complaint. Reps were juggling Salesforce, Excel, a call tool, Tableau, and a back-office system just to move a single deal forward. None of these tools talked to each other, so every agent quietly built their own patch — usually a spreadsheet only they could read. That wasn’t a training gap. It was five systems doing the job of one.

One deal, ten tools, one ordinary morning

Numbers like “five disconnected tools” are easy to read past. So put yourself in the chair for a morning. Not a bad day — an average one, the kind that happened on repeat, for a year, across seven markets.

1. 8:52 AMOpen your personal Google Sheet to check this month’s target. It says 14 of 25. It doesn’t say what to do about the other 11.
2. 9:10 AMOpen Salesforce for today’s leads. Nothing ranks them, so you start from the top and hope.
3. 9:24 AMCross-check the lead in your own Excel tracker, because Salesforce flagged three “duplicates” this week that weren’t.
4. 9:41 AMMessage the partner on WhatsApp, since it’s the only channel they actually answer — then copy that reply into Gmail, because the system of record still wants a written trail.
5. 10:52 AMDial the in-system call tool, switch to Viber when that stalls, then place a personal call from your own phone when Viber goes quiet too.
6. 2:40 PMMessage your manager on Slack for an answer mid-call, submit the contract in the back-office system, then log that same approval in the Google Sheet — because that’s still the process too.

Ten tools, one deal, and the day isn’t over — the sheet still needs updating by hand before you log off. Multiply that by every lead, every agent, every market, and 25 a month stops looking like a motivation problem. It looks like the only number that was ever possible.

The as-is: Salesforce, before any of this changed

This is the system it replaced — the full partner sign-up and pipeline flow agents worked inside every day: cockpit, activity logging, partner sign-up, tag and status handling, upsell. Nothing simplified or staged for this case study, just the screens as they existed.

Studying how the best already solve this

In parallel, I audited category-leading tools for patterns worth borrowing rather than reinventing: Jira and Asana for how they surface priority in a dashboard, HubSpot and Pipedrive as the closest existing analogs for pipeline and deal management, and WhatsApp and Telegram for the messaging mental model every agent already had muscle memory for.

02 · Define

Converge — committing to a problem

Everything was tracked. Nothing was guided.

Salesforce already had a pipeline — a full list of leads, stages, filters for grade, object type, and competitor. The gap wasn’t that nothing existed. It was that none of it told an agent what to do next. The structure was dense: rows of tags, dropdowns, and metadata, with no way to filter for what actually mattered — which leads were hot, which ones were about to go cold. Finding that meant scanning the whole list by eye, every time.

Stakeholders asked why agents on identical leads were producing wildly different results — and read it as a people problem. Talking to the people doing the job told a different story: the tooling forced every agent into their own undocumented workaround, and results tracked how good that workaround happened to be — not effort, not skill.

Salesforce had a pipeline. It just never told anyone what to do with it. Agents logged every lead in the system of record, then re-sorted the ones that actually mattered into a personal spreadsheet on the side — because Salesforce had no way to flag a hot lead, and no read on what needed attention today. Three numbers show what that gap cost:

None of it surfaced above the spreadsheet. Every lead looked the same in Salesforce, so most got treated like none of them mattered — no ranking, no suggested next action, no signal on where to start. Even monthly and quarterly targets lived in that same sheet, and tracking isn’t the same as guiding: it could show “17 of 25 done this month,” never what to do about the other 8 today. A target just kept drifting further out of reach, and nothing in the system said so until it was too late.

A heuristic evaluation, screen by screen

Rather than guess at fixes, I ran a structured heuristic evaluation of the existing Salesforce ecosystem — every screen tagged with what wasn’t working and, next to it, what to do instead. No wireframe existed yet; this was purely diagnosis.

Three roles, one broken system

The evaluation started as a sales problem. But a workshop with product and leadership stakeholders, walking through the same findings together, showed the pattern didn’t stop at sales: a Quality Checker cross-checking every IBAN, tax number, and barcode by eye, and a Content Agent chasing menus and photos across scattered uploads, were hitting the same wall from a different angle — no system catching what a person shouldn’t have had to.

That reframed the fix. This wasn’t a sales-tool patch to ship — it was a shared infrastructure gap, and closing it would lift all three roles, not just one.

Design principles, not just opinions

Every proposed fix got checked against Nielsen Norman’s usability heuristics before it earned a place in the brief — a way of making sure “this feels better” was actually “here’s why this is better, and for whom.”

Three of those mattered more than the rest. Reps wanted one cockpit — in their own words, from the research board — not “5 disconnected tools.” Dashboards ran on Tableau data that lagged 24 hours behind reality. And QC wanted errors like an invalid IBAN caught before rejection, not after.

Mapping the structure before the screens

With the diagnosis and the principles in place, the last step before any pixel work was structural: a full information architecture for what would become SalesHero, so every screen I designed next had an agreed place to live.

03 · Develop

Diverge — exploring solutions

Three numbers to move. Eight objectives to move them.

The objective was blunt: close the gap. Three numbers had to move, and nothing else about the redesign mattered if they didn’t.

Hitting those numbers meant designing a platform, not a feature. Eight objectives, each solving a specific piece of the funnel:

• 1
  Hot Leads Visibility — surfaced first, the moment a rep logs in

  The exact filter Salesforce never had. Instead of scrolling a full list every morning, the leads that actually need attention today are the first thing on screen — no spreadsheet required to know where to start.

• 2
  Mass Outreach — one bulk send, instead of message after message

  Replaces a manual chase — one partner at a time — with a single reminder sent to a whole segment at once. What used to eat an agent’s morning now takes minutes.

• 3
  Comms Hub — WhatsApp and email, native to the platform

  Ends the exact redundant work one agent called out directly: copying an email from another platform just to send it through Sales Ecosystem. Every call now closes with an automated summary and a clear next step, so nothing depends on memory.

• 4
  Self-Service Onboarding — a magic link connects Salesforce straight to the vendor’s sign-up page

  A rep sends one link; the vendor lands on the same self-sign-up flow to upload a document themselves — no paperwork chase, no re-typing between two systems that used to work in isolation.

• 5
  Duplicate Detection — flags an existing lead before a second one gets created

  If a lead’s already in the system, the rep works that one or marks it lost before starting fresh — one clean record per partner, instead of the same conversation split across three.

• 6
  Menu Digitization — OCR turns a paper menu into digital content automatically

  Answers the Content Agent’s exact complaint: a dine-in menu is scanned and converted the moment it arrives, instead of being retyped by hand from scratch.

• 7
  AI Smart Actions — automation built into every task, plus an assistant that’s always on

  A live AI agent answers a rep’s questions on demand and prompts them in real time during a call with what to say next — help that shows up in the moment, not after it.

• 8
  Pipeline Redesign — a cleaner layout for the view reps live in all day

  The same pipeline, rebuilt around fewer clicks and clearer next steps — built to move fast through 25 partners a month, or 40.

One of these eight was never a sales tool on its own. Menu Digitization exists because the same workshop that reframed the problem also reframed the roadmap — building relief for Content Agents into the platform instead of filing it as a separate backlog. Two more closed the exact gap Define surfaced on the sales side: a spreadsheet that tracked goals but never nudged anyone toward them. Hot Leads Visibility and Mass Outreach gave the dashboard the intelligence that spreadsheet never had, turning a static “17 of 25 this month” into “here are the 12 partners you haven’t touched in 20 days — message them now.”

The direction itself was never seriously contested: extend what Salesforce already had instead of replacing it, which kept feasibility risk low and tied the business case to leadership’s own numbers, not a hypothetical one. The real debate happened one level down — component by component, tested on paper before a single pixel got polished.

From whiteboard to wireframe

None of it started in Figma. The first pass at the architecture happened on paper and whiteboard, in a room, working through what a lead actually needs to move through — intent capture, qualification, outreach — before a single screen existed.

Mapping the flows that mattered most

Before finalizing any screen, I diagrammed the three multi-step flows most likely to break in production: moving a lead across pipeline stages, resolving missing documents, and messaging leads in bulk.

The lo-fi layer behind it

Pipeline, kanban, and filtered-leads views, plus the four AI-assistant placement options mentioned earlier — roughed out in Figma before any of it moved into Figma Make.

04 · Deliver

Converge — committing to scale

Five pillars carrying the whole workflow

Five pillars carry the entire workflow: dashboard, pipeline, lead detail, communication, quality check. Each one obeys the same rule — clarity first, the next action always visible, nothing buried a click deep — and each one had to trace back to Define’s heuristics before it counted as Design Ready.

The Quality Check screen carries the other half of the story from Define: rejected cases used to disappear into a stage nobody owned — the same invisibility problem sales reps had with their own leads. Giving it a reason and an owner on-screen was the fix a Quality Checker needed just as much as a Sales Rep needed a Recommendations strip.

Testing it with the people who’d actually use it

A blueprint is a hypothesis until someone outside the design team tries to use it. Between March and April 2026, I ran 14 moderated usability sessions on a live Figma prototype of this architecture — sales agents, team leads, and managers across Talabat, Foodpanda, and Yemeksepeti, in 7 markets, thinking aloud while they worked through their actual day: dashboard, pipeline, communication, onboarding, and an AI calling feature.

Interactive prototype · Figma Make Open the live prototype in Figma Make

Most of it validated on contact. Three findings stood out as proof the architecture itself needed no rework — only its edges did:

What validated

• Communication Hub — the highest-rated area in the study

  14 of 14 participants rated it 5/5, unanimous across every market tested — WhatsApp, email, calls, and AI finally living in one thread instead of five.

• Pipeline’s core actions, near-perfect on first try

  Adding a lead, signing a partner up, and calling from the pipeline each succeeded in 13 or all 14 sessions — the exact tasks agents do most, working without a single walkthrough.

• Hot Leads — the exact fix Define called for

  The filter agents used to fake with a personal spreadsheet, built in instead: 12 of 14 tapped “Reach out now” on their very first try, no explanation needed.

The architecture held up. Three specific decisions didn’t — and two of those three were brand-new ideas I deliberately put in front of users, not settled patterns just being tweaked. Not every change you make lands; testing is what tells you which ones didn’t, before a real user ever hits them in production:

What we learned

• 1
  “Qualified” pipeline stage

  A new label introduced to standardize terminology across markets, tested to see if it would travel. It didn’t: confused 11 of 14 participants, each market reading it differently — QC approval in Malaysia, a quote approval in Pakistan, a signed contract in the Philippines. Renamed to “Verbal Agreement.”

• 2
  Pipeline filters hidden in the sidebar

  9 of 14 agents never found them, defaulting to Salesforce muscle memory that expects filters at the top of the page.

• 3
  “Stalled Deals (Past EDC)” KPI label

  A new metric added to give managers pipeline-health visibility, tested as new terminology rather than an established one. It stopped almost every agent cold: internal jargon (EDC means Expected Deal Close) with no definition anywhere on-screen. The data made sense the moment someone explained it; only the label didn’t. Renamed to “Overdue Deals.”

Communication Hub came out of the study essentially untouchable — the instruction to the team was to protect it, not iterate on it. Everything else got a fix, a rename, or a market-specific configuration, and none of it required touching the underlying architecture.

Read the full UXR report

Validated the model. Then chose the platform.

Once the architecture, the feature set, and the flows were signed off — by product, and by internal and external stakeholders across markets — one decision was still open: what to actually ship first.

The 14 sessions had validated something that’s largely independent of screen size: the model underneath it. Whether “Qualified” means something different in every market, whether filters live where Salesforce muscle memory expects them, whether an always-visible payout number creates a privacy problem — none of that changes based on whether the screen is a laptop or a phone. A single, moderated desktop prototype was the fastest way to get consistent, comparable signal across seven markets in two weeks, including markets where connectivity alone was already a fight.

What that testing didn’t need to settle yet was the platform. That decision came after, and it wasn’t close: field and telesales agents run their entire day from their phones, not a laptop bag. So the build target became a Progressive Web App — installable straight to a rep’s home screen, no login dance, no laptop required — and the mobile workflow is what got built first, in Cape, Delivery Hero’s real design system, ready to hand straight to engineering.

Mobile-first, built in Cape

Dashboard, pipeline, lead detail, communication, and quality check — the same five workflows, rebuilt mobile-first in Delivery Hero’s real design system, ready for engineering to build straight from.

Roadmap & Reflection

Converge — committing to scale

Ranked, not rushed. Talabat is already live.

This is where the story stands today, not where it ends. Talabat is live, Foodpanda and Yemeksepeti are next, and nineteen ranked fixes are how the rest of the rollout gets there.

Targets for the full rollout — not results banked yet.

The first room: Talabat goes live

Talabat went first. I put SalesHero directly into the hands of more than 30 sales reps — their own phones, the real PWA, no slides. We walked the dashboard, the pipeline, a live call logged into the Communication Hub, and I talked a few people out of habits built on years of Salesforce muscle memory. Reps who’d spent a year hand-stitching five tools together watched Recommendations do in one glance what used to take a spreadsheet. They left satisfied — not because they were told it was ready, but because they’d just used it themselves.

Build Ready: the 14 sessions produced 19 ranked fixes. Five are launch blockers — renaming “Qualified” to “Verbal Agreement,” moving pipeline filters to the top, hiding payout behind a toggle, fixing Egypt’s stage order, and a Salesforce reversion bug. Six are high priority for V1.1, like a manager dashboard and role-based KPIs. The rest are market-specific — messaging channels for the Philippines, multi-contact support for Malaysia — differences no brief could have predicted.

Release Ready: the rollout itself is phased across seven entities and four languages, not big-banged. Every market runs old and new side by side for 3–4 weeks before the switch is permanent — a safeguard that came directly out of Turkey’s sessions.

The 25-partner ceiling was never about volume — it was about whether the tool got out of the way. Fourteen sessions confirmed the architecture was right; what needed fixing lived in the details no market shared with any other. And it was never just a sales problem: the same platform gave Quality Checkers a reason on every rejection and Content Agents one workflow across markets — one root cause, closed for three roles instead of one.

If I ran this again, I’d pull market-specific testing forward — Turkey’s sharpest gaps only surfaced in week two of two, with no cycle left to react. And two of these fixes were ideas I was confident about walking in; testing said otherwise. That’s not a miss. It’s the process working — catching the wrong bet before it ships to seven markets instead of one.

Talabat is proof of that, live now. The 40+, the sub-2-day, the sub-15% — still ahead of this case study, not behind it. What’s already true is smaller and more convincing: a room of agents who’d spent a year being told to work harder picked up their own phones, used the fix, and didn’t want to put it down.

Credit where it’s due

One name is on this case study; the outcome isn’t the work of one person. Engineering de-risked the platform call early. Regional sales and operations leads across Talabat, Foodpanda, and Yemeksepeti opened the door to real reps and made the Talabat pilot possible. Fourteen research participants gave honest feedback about tools they rely on daily. And the stakeholders in that early workshop are the reason this became a fix for three roles, not one.