Dashboard

How she works

Plans

Talk to Clara

Book Meeting

Talk to Clara

Book Meeting

Dashboard

How she works

Plans

Talk to Clara

Book Meeting

← All posts

Comparison

2 min read

Why Clara uses several AI models, not one

Why Clara uses several AI models, not one

Why Clara uses several AI models, not one

Clara Mind gives every job its own part: one hears, one understands, one writes, one speaks, and plain code keeps the rules. Each is tested on realistic conversations, so every job is done by the part that does it best.

Clara Mind gives every job its own part: one hears, one understands, one writes, one speaks, and plain code keeps the rules. Each is tested on realistic conversations, so every job is done by the part that does it best.

Clara Mind gives every job its own part: one hears, one understands, one writes, one speaks, and plain code keeps the rules. Each is tested on realistic conversations, so every job is done by the part that does it best.

By Gonçalo Neto

Ask a general AI to answer your phone and it will try. It will also try to understand a voice note in a Porto accent, write a WhatsApp reply in the right tone, and decide whether “não” means cancel. One model, every job. Some of those jobs it will do badly.

Clara Mind is the AI employee who answers when your team can’t, and she is not one model. Each job she does has its own part, chosen for that job.

What happens to one message

Each step has its own part

  1. A message arrives

    WhatsApp or the phone, at any hour.

  2. Hearing
    Model

    A voice note becomes text, in Portuguese, accents and all.

  3. Rules in code
    Fixed rule

    Cancelling a booking is never guessed.

  4. Understanding
    Model

    What the customer actually wants, not what a keyword suggests.

  5. Writing
    Model

    The right facts and the right tone, in the customer’s language.

  6. Reply

    Sent on the same channel. On a call, a voice chosen for speed does the talking.

Figure 1.Every step of a message has its own part, and the step that must never be guessed is plain code.

Close is not good enough

Models from the same family can look interchangeable. On our own test conversations they were not.

We sent the same English test messages, some captured from the live system, to several models and checked every reply for mistakes a customer would see (source: Reply test). A cheaper model from the same family as the one we chose answered every one of them in Portuguese: six samples out of six, against none for the model we chose.

English messages answered in Portuguese

Each dot is one test message. A filled dot is a message answered in Portuguese.

  • A cheaper model from the same family6 of 6
  • The model we chose for the job0 of 6

Figure 2.On the same English messages, the model picked for the job never answered in the wrong language, and its close cousin did every time. Source: Reply test, 5 September 2026.

That is the whole case for choosing per job. On paper the two were nearly the same. In front of a customer, one of them answered in the wrong language, six times out of six.

Some jobs should not be AI at all

Some decisions must never be guessed. Cancelling a booking is one.

A customer once corrected the time of a booking with a message like “Não, é às cinco”, no, it’s at five. A rule that guessed from the first word read it as a cancellation.

So we wrote a test of 26 messages, including that one: 18 where the customer never asked to cancel and 8 genuine cancellations (source: Cancellation test). The guessed rule cancelled 13 of the 18 by mistake and missed 7 of the 8 genuine cancellations. The fixed rule, written as code, got all 26 right.

Wrong answers on 26 test messages

Each mark is one message. Hover or tap one to read it.

Corrections7
Denials6
Plain no5
Genuine cancellations8
Guessed rule
20of 26 wrong
13 of 18 bookings cancelled by mistake, 7 of 8 genuine cancellations missed
Fixed rule
0of 26 wrong
0 of 18 bookings cancelled by mistake, 0 of 8 genuine cancellations missed

“Não, é às cinco”The customer never asked to cancelGuessed rule: booking cancelled (wrong). Fixed rule: booking kept (right).

Figure 3.Switching from the guessed rule to the fixed rule turns twenty wrong answers into none. Source: Cancellation test, 20 September 2026.

Building it yourself

Connecting one AI to a phone line takes an afternoon. Making it right for real customers means deciding, alone:

  • Which model hears Portuguese voice notes, which understands, and which writes.

  • Which voice is quick enough on a call.

  • Which decisions must never be guessed, written as code.

  • A set of your own customers’ real messages to test every choice on, again with every new model.

  • Where your customers’ messages go. Some cheap models are paid for with the right to learn from them.

Each choice is a place to get it wrong, quietly, in front of a customer. Testing does not make her perfect; it shows us where she is weakest, so that is the part we fix first.

Hear the difference. Call her on +351 308 806 081, in Portuguese or in English.

Talk to Clara

Book a meeting

Clara Mind is made by Quadrant Sunshine, in Lisbon. She answers on the phone and on WhatsApp, in Portuguese or English.

Sources

  1. Reply test: the same English test messages, some captured from the live system, sent to several models on 5 September 2026; every reply checked for mistakes a customer would see. Six samples per model.

  2. Cancellation test: 26 test messages written to cover corrections, denials, plain no’s and genuine cancellations (18 where the customer never asked to cancel, 8 genuine cancellations), run against the guessed rule and the fixed rule on 20 September 2026.