Voice Order Entry in Pakistani FMCG: What I've Learned After 18 Months

A rep in Faisalabad once told me he takes 42 orders a day. Forty-two. And each one used to eat about 90 seconds of tapping through SKUs on his phone while a shopkeeper stared at him getting impatient.

That math never sat right with me.

90 seconds times 42 outlets is roughly an hour of pure thumb-scrolling every single day. Multiply that across a distributor with 60 reps and you're bleeding 60 hours of selling time daily. That's the problem voice order entry actually solves in Pakistani FMCG — not some futuristic AI story, just giving reps back their hours.

But I got this wrong at first. I'll explain.

Why typing-based order apps were always going to struggle here

When we started building Zivni, I assumed the biggest issue with Pakistani field sales was connectivity. Bad 3G in rural Sindh, dead zones in interior Punjab, that kind of thing. And yeah, that matters. But the real bottleneck sits somewhere else entirely.

It's the SKU count.

A typical FMCG distributor in Lahore or Karachi carries anywhere from 300 to 1,200 SKUs across categories. Biscuits, beverages, personal care, dairy — a Unilever or Peek Freans distributor might carry more. Now imagine a rep at a kiryana store trying to scroll through 800 SKUs while the shopkeeper is calling out orders in rapid Urdu mixed with English. "Do carton Tapal Danedar, aik Lipton yellow label, chhe packet Slanty masala…"

By the time the rep finds Slanty in the app, the shopkeeper has moved on to three other items and forgotten what he wanted.

So reps did what humans do. They wrote orders on paper, then re-entered them into the app later in the evening — usually with mistakes, usually skipping the geo-tag, usually inflating volumes to hit targets. The whole point of field sales technology in Pakistan gets defeated the moment reps go back to notebooks.

Voice order entry breaks that loop. The rep just talks. The app parses "do carton Tapal Danedar" into 2 units of the right SKU, matches it against the distributor's price list, and drops it into the cart. Shopkeeper keeps rattling off items. Rep keeps confirming. Order done in under 40 seconds.

The stuff nobody warns you about

Honestly, the first version we shipped was rough. Really rough.

Our speech model was trained heavily on urban Urdu and standard English. Which meant it worked beautifully in DHA Karachi and completely fell apart in Multan and Peshawar. Regional accents, Saraiki-influenced Urdu, Pashto loanwords in product names — the model would hear "chai patti" and log "chapati." Reps hated it. One distributor manager in Gujranwala sent me a WhatsApp voice note that I still remember: "Sufyan bhai, ye app humein pagal kar de gi."

Fair.

We rebuilt the acoustic model with roughly 340 hours of field-recorded audio from actual reps across 11 cities. Not studio recordings — the real thing, with horns honking, generators running, kids yelling in the background. Accuracy jumped from something like 71% to 94% on mixed Urdu-English orders. Still not perfect. But usable.

Here's the thing about voice in FMCG Pakistan — you can't build it in a Silicon Valley lab and expect it to work in Sadar Bazaar. The linguistic reality is too specific. Reps code-switch constantly. Product names are half-English, half-Urdu, sometimes with brand nicknames the shopkeepers made up ("lal wala Tang" for the red packaging).

Our competitors — FieldAssist, BeatRoute, some of the newer players — mostly punt on this. They offer voice as a checkbox feature but the underlying model is English-first with Urdu bolted on. It shows.

What actually changes on the ground

A distributor we work with in Rawalpindi runs 34 reps. Before voice order entry, average orders per rep per day was 38. Six months in, it's 51. Same reps. Same beats. Same territory.

That's a 34% jump in productive outlet coverage without hiring anyone new.

But the number I care about more is data quality. Order accuracy — meaning what the shopkeeper actually wanted versus what got logged — went from around 82% to 97%. Returns and disputes dropped. Distributor's ops team stopped spending Friday afternoons reconciling paper slips against app entries.

And reps started liking their job more. That sounds soft but it isn't. Field sales attrition in Pakistani FMCG runs brutal — 40 to 60% annually at most distributors. When you take away the two hours of evening data entry a rep used to do at home, you're giving him back his life. He tells his friends. Referrals go up. Recruiting costs drop.

One small thing I didn't expect: shopkeepers started respecting the tech. A rep talking to his phone in Urdu and having it instantly show a printed order confirmation on screen — that's a status thing at the kiryana counter. Sounds silly. It's real.

Where voice still doesn't work

Look, I'm not going to pretend this is magic.

Voice struggles when three people talk at once, which happens in busy wholesale markets. It struggles with brand-new SKUs the model hasn't seen yet — you need to update the vocabulary every time a distributor onboards a new principal. And it will never fully replace the tap interface for edge cases like promotional bundles or split-payment orders.

We treat voice as the default, not the only path. Rep can always fall back to typing when he needs to. Best tools don't force one workflow — they meet the rep where the moment is.

If you're running a distribution business in Karachi, Lahore, Islamabad, or anywhere across Pakistan and your reps are still tapping through 800 SKUs to log an order, you're leaving hours on the table every day. Do the math on your own team. 90 seconds times how many outlets times how many reps.

What's that number for you?