AI-Powered Conversational Food Ordering with iLo

ifood · E-commerce · 2026 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Food delivery recommendation and ordering

iFood wanted to move beyond product search into conversational, hyper-personalized ordering that understands food preferences, budget, channel constraints, and ambiguous user requests at large scale.

Constraints:

Food preferences vary by taste, price sensitivity, location, and purchase history
Voice, WhatsApp, and app experiences have different latency and detail constraints
Users may treat a food-ordering chat as a general iFood support channel

Architecture

iFood's iLo sits on top of the company's LCM recommendation profile and channel-specific conversational surfaces. The system combines historical user characteristics, recommendation logic, LLM reasoning, WhatsApp/app/voice interfaces, payment/order flow, and experimentation through the jet-ski operating model. The architecture is deliberately hybrid: shared intelligence and user data feed different channel experiences because voice, WhatsApp, and app interfaces have different latency and presentation constraints.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  ilo-conversational-agent["ilo-conversational-agent"]
  lcm-user-profile["lcm-user-profile"]
  recommendation-and-ranking-layer["recommendation-and-ranking-layer"]
  channel-experience-layer["channel-experience-layer"]
  payment-and-order-flow["payment-and-order-flow"]
  jet-ski-experimentation-loop["jet-ski-experimentation-loop"]
  support-handoff-router["support-handoff-router"]
  ilo-conversational-agent --> lcm-user-profile
  ilo-conversational-agent --> channel-experience-layer
  ilo-conversational-agent --> payment-and-order-flow
  ilo-conversational-agent --> support-handoff-router
  lcm-user-profile --> recommendation-and-ranking-layer
  recommendation-and-ranking-layer --> ilo-conversational-agent
  channel-experience-layer --> payment-and-order-flow
  jet-ski-experimentation-loop --> ilo-conversational-agent

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`ilo-conversational-agent`	User-facing food-ordering agent that interprets natural-language requests and drives recommendation and ordering flows.	`lcm-user-profile`, `channel-experience-layer`, `payment-and-order-flow`, `support-handoff-router`
`lcm-user-profile`	Recommendation profile that summarizes many recommendation approaches and user characteristics for personalization.	`ilo-conversational-agent`, `recommendation-and-ranking-layer`
`recommendation-and-ranking-layer`	Hybrid recommender and optimization layer that balances food preference, price, timing, and context.	`ilo-conversational-agent`
`channel-experience-layer`	WhatsApp, app, and voice interface logic that adapts response detail and latency expectations to the channel.	`ilo-conversational-agent`, `payment-and-order-flow`
`payment-and-order-flow`	Checkout path, including Pix payment on WhatsApp, that turns a recommendation into an order.	—
`jet-ski-experimentation-loop`	Small-team experimentation model for testing iLo quickly before investing at larger scale.	`ilo-conversational-agent`
`support-handoff-router`	Emerging routing concept for complaints or customer-support intents that users send to the ordering agent.	`ilo-conversational-agent`

Data flow

A user asks for food in natural language through WhatsApp, app, or voice. iLo uses the LCM profile to understand the user's characteristics and likely dish preference, applies conversational reasoning to narrow the request, presents a channel-appropriate recommendation, and can continue into checkout, including Pix payment on WhatsApp. Non-ordering complaints may need a handoff or orchestration path.

“we use the base of LCM which is which has the the full characteristics of the user to understand what is the best request what is the best dish for you at that time”transcript, stated

“Because you can send an audio saying, "Okay, I want a pizza." And he knows what pizza you like. And you can pay by Pix in WhatsApp like very fast”transcript, stated

Why this architecture

iFood chose a hybrid architecture because recommendation quality depends on existing company data and recommendation systems, while conversational interaction depends on channel-specific LLM behavior. The jet-ski model keeps early product risk cheap, and the channel layer prevents a single interaction design from serving app, WhatsApp, and voice poorly.

“you have your idea, you build your agent, but when you try to prototype it, so you have much more than the AI agent. You have the back end that you you consume your data.”transcript, stated

“So, we've tried we tried to to to That's that's another another hint, right? If you are working on a voice to voice message, please do not use one model”transcript, implied

Models used: LCM recommendation model · LLMs · voice-capable multimodal/generative models

Scale

iFood operates at hundreds of millions of orders per month.
“I know you have like iFood doing hundreds of millions of orders per month.”transcript, stated
The iLo experiment was running with about half a million users.
“We try it uh with half a million users right now.”transcript, stated
Benchmarks showed 16% faster order completion from search to checkout.
“we started finding some interesting things like it's 6 16% faster to complete an order from a search to a checkout with Ailo than iFood uh product search”transcript, stated
The benchmark also showed a 35% higher probability of search becoming cart addition.
“it's 35% higher the probability that a search became a cart addition as well.”transcript, stated
The product reasoning includes concrete order constraints such as delivery in 40 minutes and cost below 50 reais.
“I want a pizza that could be delivered in 40 minutes, should cost less than 50 reais”transcript, stated

Alternatives considered (roads not taken)

Big-ship product planning before user validation

Rejected because: The jet-ski model tests theses quickly and cheaply, then invests big only after evidence appears.

“testing things very fast. And if it doesn't work out, that's fine. You go to another thesis and try it again. So, the thing is you need to do it fast, cheap”transcript, stated

Identical responses across WhatsApp, app, and voice

Rejected because: The team recognized that each channel has different latency tolerance and presentation affordances.

“if you were working on WhatsApp, you can take a little bit more because you're expecting this to be more like a sequence. However, if you are on a voice mode”transcript, stated

Voice pipeline that separately transcribes and then interprets speech

Rejected because: The speaker warns that voice-to-text-to-reasoning chains introduce problems and prefers direct voice-capable models when possible.

“please do not use one model to just to translate through your voice and then go to another model to interpret from voice to text. Try to do this directly”transcript, stated

Treating iLo only as a food-ordering bot

Rejected because: Users report general iFood problems to the chat interface, so the team is considering support handoff or broader orchestration.

“I don't know, iFood because every company has a problem with a customer. And they report this problem by using ilo.”transcript, stated

Patterns

`retrieval-ranking` (personalized-recommendation)

iLo ranks food options by combining user taste, location, budget, purchase history, and similarity to other users/restaurants/dishes.

“I need to understand not only your characteristics of what your tastes are, but your economic profile, what you were aiming to to buy, what is your historical purchases”transcript, stated

`canary-rollout` (product-experimentation)

The jet-ski model tests AI ordering bets cheaply and quickly, then invests heavily only when a thesis works.

“testing things very fast. And if it doesn't work out, that's fine. You go to another thesis and try it again.”transcript, stated

`latency-optimization` (conversation-channel-design)

iFood varies response detail and processing by channel because voice needs faster, terser answers than WhatsApp or app UI.

“if you were working on WhatsApp, you can take a little bit more because you're expecting this to be more like a sequence.”transcript, stated

“if you are on a voice mode, for example, where you are calling and you're calling the AI, that should be very fast.”transcript, stated

`model-routing` (multi-channel-agent)

The team discusses separate or flexible systems for voice, GUI, and WhatsApp because each channel requires different intelligence/detail tradeoffs.

“you almost have to have two different systems because what you were saying with the latency requirements and the experience of a voice agent versus the GUI or versus WhatsApp”transcript, implied

`multi-agent-orchestration` (cross-domain-support)

The product discussion anticipates routing food-ordering and customer-problem intents either to specialized agents or one orchestrated experience.

“If you want If you have a customer problem, talk with my other friend here that's going to help you.”transcript, implied

“Or you're going to arc straight everything. Yeah, that's what I was thinking. In one experience, one voice tone, one kind of messaging”transcript, implied

Tradeoffs

How much intelligence to spend in a voice recommendation.

Chose a sufficiently good, concise recommendation over the best fully detailed recommendation.

Rationale: Voice latency and attention constraints make perfect-but-slow recommendations less usable than fast, clear suggestions.

“What can I simplify in terms of my whole intelligence to make sure that the user will have a very good experience, but not the perfect one”transcript, stated

How to handle non-ordering intents inside iLo.

Chose explore handoff or orchestration over ignoring customer-problem reports.

Rationale: If the bot fails to respond to support complaints, users may perceive the whole assistant as unintelligent.

“If you answer nothing, Yeah. probably the client will think you are dumb.”transcript, implied

Gotchas (what bit them)

Personalized food recommendation has no silver bullet.

Root cause: The system must relate users, nearby restaurants, dishes, budgets, preferences, and historical behavior in a fast-changing marketplace.

Resolution: iFood treats recommendation quality as trial-and-error experimentation over many signals rather than a one-shot model choice.

“there is no there is no rocket rocket science. There is no easy way. A silver bullet.”transcript, stated

Promising metrics did not eliminate scale and cost problems.

Root cause: LLM-based conversational ordering adds infrastructure, inference, and product-market-fit complexity beyond classic search.

Resolution: iFood continues shaping user behavior and product-market fit under the jet-ski experimentation model.

“We still uh have some things that we are struggling like scalability, uh some costs and I think everybody that's working with AI are facing some you know, challenges.”transcript, stated

Voice agents cannot present the same detail as a screen.

Root cause: Voice has one channel and long descriptions become boring or slow, while app UIs can display many data points.

Resolution: The voice flow should be assertive and concise, presenting the most important recommendation rationale.

“voice cuz you only have one channel. And with an interface, you have many different data points that can pop up.”transcript, stated

Users may send support complaints to the ordering bot.

Root cause: From the user's perspective, chat is one iFood interface rather than a set of verticalized bots.

Resolution: The team is considering fast customer-support handoff or orchestration across multiple agents.

“they have a problem with I don't know, iFood because every company has a problem with a customer. And they report this problem by using ilo.”transcript, stated

Evaluation

iFood evaluates iLo through benchmark comparisons to product search, conversion from search to cart, order-completion speed, user research, conversation deep dives, and ongoing product-market-fit experiments.

Metrics: half-million-user experiment · order-completion speed · search-to-cart conversion · scalability · cost · product-market fit

“We try it uh with half a million users right now.”transcript, stated

“you need to deep dive the conversation itself because people say, I hate this recommendation. But why?”transcript, stated

Outcomes

iLo is being tested with about half a million users.
Orders complete about 16% faster than with iFood product search in benchmarks.
Search-to-cart-addition probability is about 35% higher.
The system supports conversational ordering through WhatsApp and other channels.

iFood operates at hundreds of millions of orders per month, making the conversational assistant a large-marketplace personalization problem.

3 supporting quotes

“it's 6 16% faster to complete an order from a search to a checkout with Ailo than iFood uh product search when we benchmark.”transcript, stated

“it's 35% higher the probability that a search became a cart addition as well.”transcript, stated

“iFood doing hundreds of millions of orders per month. So, it's a big ship and a big ship is very hard to maneuver.”transcript, stated