A conversation that acts

It does not tell you where to click. It does the thing.

Ask for lenders and the list arrives in the conversation. Give it a figure and the field is filled. Drop in a statement and it is read, not just attached.

We prepare the file. The lender interviews, verifies and decides.

Open a conversation How it works

What it can put in front of you

From the product

The assistant does not only write sentences. It emits typed pieces of interface that render inside the conversation, and these are the kinds it can produce.

01A table you can read and sort
02A form you fill in without leaving the chat
03A confirmation step before anything is done on your behalf
04A chart over time
05A statistical chart
06A grid of figures
07A map
08A diagram
09Formatted text and documents

Every one is checked against a schema before a pixel is drawn, and the set is closed: a new kind that has not been declared in the type fails the build rather than reaching you.

Why this exists

Most assistants beside a product are a help desk wearing a costume.

They explain where the button is, and you still have to go and press it. That is a search box with manners. If the conversation is the way you actually get to a lender, it has to be able to list them, filter them, open the right session, fill the field, read your document and hand the file on. Otherwise it is one more thing to learn on the way to the thing you wanted.

How it works

Six properties, and the last two are the safety ones.

01
It acts, it does not point.
From your own session it can look up loan products across lenders, fill a field in your application, read a document you just dropped in, teach a concept, check what you have actually understood, and search live public sources when what we hold is thin.
02
It draws, inside the conversation.
A comparison arrives as a table. A set of questions arrives as a form. A projection arrives as a chart. You are not sent to another screen to see the answer to the thing you just asked.
03
What it draws can act back.
A form rendered in the conversation submits back into the conversation and runs the action it was presented for. The piece of interface is part of the dialogue, not a dead illustration of one.
04
What you say becomes part of your file.
There are three ways into one record: what you say, what you upload, and what you type into the stepper. A figure mentioned in conversation is captured with its source, so nothing has to be entered twice.
05
Nothing arbitrary is ever rendered.
Every piece of interface is parsed against a schema first, and the renderer is a closed set of vetted components. Raw markup from a model never reaches your screen, by construction rather than by review.
06
Your own answers cannot become commands.
When you submit a form in the conversation, your values are passed to the assistant wrapped and explicitly marked as data, never as instructions. A sentence typed into a field cannot become an order the assistant follows.

A conversation that can act needs a line drawn around it.

The more a chat can do, the more carefully it has to be bounded. Ours is bounded twice: by what your session is allowed to reach, and by what may appear in anything it draws.

It can

Bring you lenders and products that fit your size, your sector and your need.
Fill your application, generate your documents and surface the next gap.
Read what you upload and tell you what it found.
Teach, and check whether the teaching landed.

It will not

Speak for the lender, or tell you what their answer will be.
Tell you that you are verified, cleared or approved.
Act on an instruction hidden inside a document or a form value.
Advise on an attachment it could not actually read.

What people ask before they trust this.

Is this the same assistant everywhere, or a smaller one for the public?: It is bounded by what your session is entitled to reach, not by a cut-down version of the mind. A public visitor can be advised and taught and can research a venture; filling your application needs your session, because it writes to your file.
Can I do everything by talking, or is the chat a shortcut?: The target is that anything you can reach by clicking, you can reach by asking, and a click path with no spoken equivalent is treated as a defect rather than a design choice. The dashboard and the tabs are conveniences.
What happens to a document I paste in?: It goes through the same categorise, extract and capture path as a formal upload, and what was found comes back to you in your language. Where only the file's outer details could be read, it says so instead of advising on something it did not read.
If I stop halfway, is the conversation gone?: No. What you said is already part of your file, with its source recorded, so picking the conversation back up later or on another phone continues from where you left it rather than starting again.

The rest of what is built

Say what you are trying to build.

In English or Swahili, in your own words. It works out which capability that is.

Open a conversation All capabilities

A conversation that acts

It does not tell you where to click. It does the thing.

Ask for lenders and the list arrives in the conversation. Give it a figure and the field is filled. Drop in a statement and it is read, not just attached.

We prepare the file. The lender interviews, verifies and decides.

Open a conversation How it works

What it can put in front of you

From the product

The assistant does not only write sentences. It emits typed pieces of interface that render inside the conversation, and these are the kinds it can produce.

01A table you can read and sort
02A form you fill in without leaving the chat
03A confirmation step before anything is done on your behalf
04A chart over time
05A statistical chart
06A grid of figures
07A map
08A diagram
09Formatted text and documents

Every one is checked against a schema before a pixel is drawn, and the set is closed: a new kind that has not been declared in the type fails the build rather than reaching you.

Why this exists

Most assistants beside a product are a help desk wearing a costume.

How it works

Six properties, and the last two are the safety ones.

01
It acts, it does not point.
From your own session it can look up loan products across lenders, fill a field in your application, read a document you just dropped in, teach a concept, check what you have actually understood, and search live public sources when what we hold is thin.
02
It draws, inside the conversation.
A comparison arrives as a table. A set of questions arrives as a form. A projection arrives as a chart. You are not sent to another screen to see the answer to the thing you just asked.
03
What it draws can act back.
A form rendered in the conversation submits back into the conversation and runs the action it was presented for. The piece of interface is part of the dialogue, not a dead illustration of one.
04
What you say becomes part of your file.
There are three ways into one record: what you say, what you upload, and what you type into the stepper. A figure mentioned in conversation is captured with its source, so nothing has to be entered twice.
05
Nothing arbitrary is ever rendered.
Every piece of interface is parsed against a schema first, and the renderer is a closed set of vetted components. Raw markup from a model never reaches your screen, by construction rather than by review.
06
Your own answers cannot become commands.
When you submit a form in the conversation, your values are passed to the assistant wrapped and explicitly marked as data, never as instructions. A sentence typed into a field cannot become an order the assistant follows.

A conversation that can act needs a line drawn around it.

The more a chat can do, the more carefully it has to be bounded. Ours is bounded twice: by what your session is allowed to reach, and by what may appear in anything it draws.

It can

Bring you lenders and products that fit your size, your sector and your need.
Fill your application, generate your documents and surface the next gap.
Read what you upload and tell you what it found.
Teach, and check whether the teaching landed.

It will not

Speak for the lender, or tell you what their answer will be.
Tell you that you are verified, cleared or approved.
Act on an instruction hidden inside a document or a form value.
Advise on an attachment it could not actually read.

What people ask before they trust this.

Is this the same assistant everywhere, or a smaller one for the public?: It is bounded by what your session is entitled to reach, not by a cut-down version of the mind. A public visitor can be advised and taught and can research a venture; filling your application needs your session, because it writes to your file.
Can I do everything by talking, or is the chat a shortcut?: The target is that anything you can reach by clicking, you can reach by asking, and a click path with no spoken equivalent is treated as a defect rather than a design choice. The dashboard and the tabs are conveniences.
What happens to a document I paste in?: It goes through the same categorise, extract and capture path as a formal upload, and what was found comes back to you in your language. Where only the file's outer details could be read, it says so instead of advising on something it did not read.
If I stop halfway, is the conversation gone?: No. What you said is already part of your file, with its source recorded, so picking the conversation back up later or on another phone continues from where you left it rather than starting again.

The rest of what is built

Say what you are trying to build.

In English or Swahili, in your own words. It works out which capability that is.

Open a conversation All capabilities

It does not tell you where to click. It does the thing.

Most assistants beside a product are a help desk wearing a costume.

Six properties, and the last two are the safety ones.

It acts, it does not point.

It draws, inside the conversation.

What it draws can act back.

What you say becomes part of your file.

Nothing arbitrary is ever rendered.

Your own answers cannot become commands.

A conversation that can act needs a line drawn around it.

It can

It will not

What people ask before they trust this.

The rest of what is built

Say what you are trying to build.

It does not tell you where to click. It does the thing.

Most assistants beside a product are a help desk wearing a costume.

Six properties, and the last two are the safety ones.

It acts, it does not point.

It draws, inside the conversation.

What it draws can act back.

What you say becomes part of your file.

Nothing arbitrary is ever rendered.

Your own answers cannot become commands.

A conversation that can act needs a line drawn around it.

It can

It will not

What people ask before they trust this.

The rest of what is built

Say what you are trying to build.