OfficeAI: AI Voice Agents for Restaurants and Takeaways in Ireland
As I continue researching AI companies across Ireland through AI Dubliners, one thing keeps becoming clearer to me: the value of AI often comes from solving very practical problems in everyday workflows, rather than from big, abstract ideas.
Dublin-based OfficeAI is a good example of that.
Founded by Safaiet Hosan, OfficeAI is an AI voice agent built for restaurants and takeaways. It is particularly focused on smaller businesses where one or two people may be working in the kitchen.
The problem is simple.
The phone often rings at exactly the busiest moment. Someone has to stop what they are doing to answer it. If the call goes unanswered, that can mean a lost order.
The other option is often to rely on delivery aggregator platforms, which can take around 25 to 30 percent in commission. OfficeAI is trying to address the gap between those two realities.
More than an AI that answers the phone
OfficeAI is not designed simply to pick up incoming calls.
When a customer rings, the agent can manage the full conversation.
It takes the order, handles sizes and extras, works out whether the order is for delivery or collection, checks whether the address is within the delivery area, asks how the customer wants to pay, reads the order back and sends the ticket directly to the kitchen.
It can also send an SMS confirmation to the customer.
The system goes beyond order taking too.
It can handle table bookings and answer the kinds of questions customers regularly ask restaurants.
For example:
“What’s in it?”
“Is there nut in the korma?”
“Do you deliver to Cahir?”
“What time do you close?”
The aim is to reduce the need for the owner to handle the phone during service.
The restaurant can teach the agent itself
One of the most interesting parts of OfficeAI, at least to me, is that the business owner can teach the agent directly.
There is no need for a developer, prompt engineering knowledge or an ongoing support ticket.
The owner uploads the menu and can then teach the agent new information in plain English through the dashboard.
That information can include permanent facts.
It can include rules about how the agent should behave.
Or it can be something temporary.
“We’re out of king prawn tonight.”
A note like that can be added with an expiry, so the information only applies for as long as it is relevant.
Over time, this allows the agent to become more specific to each restaurant. After a couple of weeks, it can already know things that were not possible to anticipate at the beginning.
What happens when the AI does not know the answer?
This is where an important guardrail comes in.
If the agent does not know the answer to a customer’s question the first time, it asks for a human transfer rather than guessing.
The reason for the transfer then appears in the dashboard.
The business owner can teach the agent the correct answer, so when the same issue comes up in a future call, the system already knows how to respond.
I find that approach particularly interesting because the goal is not to make the AI answer everything at all costs. It is to know when a human should step in, and then turn that moment into something the system can learn from.
“The model turned out to be the easy part”
One of the most interesting lessons OfficeAI founder Safaiet Hosan has taken from building the product is this:
“The model turned out to be the easy part.”
Most of the real-world problems happen upstream or downstream of the model.
Pricing is a good example.
At OfficeAI, prices come from the database, never from the language model.
A model improvising a price is not an acceptable mistake in this context. The price the customer hears has to be the price they are actually charged.
The same principle applies to other critical steps in the order flow.
Things such as choosing a size, capturing the address or selecting a payment method are enforced on the server, rather than being left only to prompt instructions.
That matters because prompt instructions can drift during a long call.
A “500 error” feels very different on a phone call
Error handling also becomes more important when AI is working through voice.
In a web application, a user may simply see an error message when something fails.
On a phone call, there is no screen.
If a tool or service fails, what the customer experiences is the assistant suddenly breaking down.
That is why OfficeAI needs tool failures to be recoverable during the conversation itself.
And then there is the Irish reality
Making voice AI work in production is not only about choosing a good model.
OfficeAI also has to deal with challenges that are specific to Ireland.
These include understanding Irish accents over poor phone lines, regulatory approval for phone numbers and Eircodes.
OfficeAI is currently live with takeaways in Ireland and is particularly focused on restaurants and takeaway businesses with small kitchen teams.
From model to real product
What stood out to me most while learning about OfficeAI is how clearly it shows that the value of AI in the real world is not only about how capable the model is.
Where the prices come from matters.
How critical steps are enforced matters.
Knowing when a human should take over matters.
Handling failures matters.
And giving the business owner a way to teach the system over time matters too.
Perhaps that is where much of the applied AI story is heading.
The model is becoming the starting point.
The real difference comes from how reliably it can be connected to the daily operations of a real business.
OfficeAI is one of the Dublin-based companies trying to do exactly that, through a very ordinary but very real problem faced by restaurants and takeaways.
