Techbanq

AI agents · orders

An order is a dozen decisions taken at speaking pace

Taking an order is not answering a question. It is resolving items, sizes, modifiers and quantities against a live menu, pricing the result, and writing it into the POS so the kitchen makes exactly what the customer said — across phone and chat, in Arabic and English, at the pace of the dinner rush.

Serving Saudi Arabia from our Riyadh office and our engineering hubs in Lahore and Karachi

Shoppers moving through a modern retail concourse
  1. 01

    Market

    Saudi Arabia

  2. 02

    Written for

    Commercial Director, COO

  3. 03

    Decision

    Automate order capture across phone and chat without corrupting the POS

  4. 04

    Languages

    EN · AR

Enterprise delivery credentials

  • ISO

    ISO/IEC 27001

    ISO/IEC 27001 certified

Certifications & registrations

The transaction

From “I want the large one” to a line the kitchen can cook

The same spine runs whether the order arrives by phone or by chat. The agent’s job is to end every exchange with a structured, priced basket in the POS — or with a person holding the conversation and the basket built so far.

A flow from resolving the item, settling modifiers, reading back and pricing, to writing structured lines into the POS — with a branch handing ambiguous orders to staff along with the basket built so far.

  1. 01

    Resolve the item

    The customer’s words — a colloquial name, an Arabic-English mix, “the usual one” — are matched to a real item on the live menu, not to a guess.

  2. 02

    Settle the modifiers

    Size, spice, extras, removals and quantity are confirmed against what the menu actually allows for that item, so an impossible combination is caught in the conversation, not in the kitchen.

  3. 03

    Read back and price

    The basket is read back with its total before anything is committed, because the read-back is where a misheard item is cheapest to fix.

  4. 04

    Write into the POS

    The order lands as structured lines with the POS’s own item codes and prices — the same shape a cashier would key — so the kitchen display, the receipt and the report all agree.

Always available

A person takes the order over

A complaint, a catering enquiry, an allergy question or an item the agent cannot resolve hands the conversation to staff with the basket built so far, so the customer never repeats the half of the order that went well.

The hard part

The menu is a data model wearing a laminated card

What defeats order automation is rarely comprehension; it is that the menu the agent needs does not exist as data. Items have colloquial names the printed card never mentions, in two languages and every mix of them. Modifiers interact — this size excludes that extra, this combo fixes the drink. Availability changes mid-shift when an item runs out. An agent working from a PDF of the menu will take orders fluently and wrongly; one working from a structured model synchronised with the POS takes them correctly, because it is choosing from what can actually be sold at that minute.

This is also what separates order taking from its neighbours. A booking agent writes one appointment into a calendar; an order agent writes many priced lines into a transaction system, and every line is a chance to be wrong. Where the phone is the whole channel and the reservation is the product, our restaurant voice agent work covers that ground; this page is about the transaction itself, on whichever channel it arrives.

Where it goes wrong

Four order failures that are menu data, not language

Each of these looks like the agent misunderstood the customer. Each is actually a gap between the menu the agent believes and the one the POS is selling.

  1. 01

    The name on the street is not the name in the POS

    Customers order by the colloquial name, the old name, or the Arabic name of an English-listed item. Without an alias table maintained per item, every one of those becomes “sorry, we don’t have that” for something the shop sells daily.

  2. 02

    A modifier the kitchen cannot make

    An agent that accepts any modifier phrase will cheerfully sell a combination the item does not allow. The rules live in the menu model, so the impossible request is caught and renegotiated while the customer is still on the line.

  3. 03

    Sold out at seven, still selling at eight

    An item marked out in the POS must vanish from the agent’s menu within the same shift. An agent that keeps selling it manufactures a refund and an angry call per order — a worse outcome than the queue it replaced.

  4. 04

    The agent’s price and the receipt disagree

    If the agent prices from its own copy of the menu, promotions and price changes drift. The fix is structural: the agent quotes the price the POS computes for that basket, so the number said aloud is the number printed.

Our role

What we wire between the conversation and the till

  1. 01

    The menu model and its upkeep

    We build the structured menu — items, aliases in both languages, modifier rules, availability — and the synchronisation that keeps it honest against the POS, because that model is the agent.

  2. 02

    The write path into the POS

    Orders land as structured, priced lines through the POS’s own interface, tested against its item codes and promotions, so nothing the agent takes needs re-keying and nothing it prices contradicts the receipt.

  3. 03

    The handover, designed not hoped for

    We define with your floor team which situations go to a person — complaints, allergies, large or unusual orders — and build the transfer so the basket and the context arrive with the call.

The menu, its prices and what may be promised to a customer stay decisions of your operation; we build the agent that carries them accurately from the customer’s words into the systems your kitchen and finance already trust.

Common questions

What buyers ask

Which POS systems can the agent write orders into?+

Any POS that exposes an interface for creating orders — most modern cloud POS platforms used in the Gulf do, and that interface is a scoping question we settle in the first conversation. Where a POS offers no order-creation route at all, we say so early, because the alternative — printing a ticket for staff to re-key — keeps the labour the agent was meant to remove. The write path into your specific POS, tested against its item codes and promotions, is the core of the build rather than an add-on.

Customers here order in a mix of Arabic and English. Can it follow?+

It has to, because that mix is how orders are actually spoken and typed in the Gulf — an English item name inside an Arabic sentence, quantities in either language, dialect rather than textbook Arabic. The agent is built and tested on that reality, and the menu model carries names and aliases in both languages so “the large one” resolves to the same POS code whichever language surrounds it. An agent evaluated only on clean single-language test phrases will fail on the first real Thursday night, which is why we test on transcripts of real orders before going live.

What happens to orders already in progress when an item runs out mid-shift?+

The availability change reaches the agent’s menu within the same synchronisation cycle, so new conversations stop offering the item at once. A conversation already holding it in the basket gets the honest treatment: the agent says it has just sold out, offers the nearest alternative from the same menu category, and reprices the read-back. What the agent must never do is complete the sale and let the kitchen discover the gap — that failure converts one disappointed customer into a refund, a call and a lost repeat order.

Scope it with us

A scoping session that walks one real order — with its modifiers and its Arabic-English mix — from the customer’s words to a priced line in your POS.

Start the conversation