Your company’s AI model
The AI that has learned your business.
HUB AI trains a model on your ways of working, tests it before entrusting it with a file and puts it to work in your software tools. On machines reserved for your company, hosted in Canada.
A good answer takes three things. A model alone brings only one.
Think of an experienced employee. She has the reflexes of the trade. She opens the right file before answering. And she knows what she is allowed to sign. HUB AI builds all three, each one separately, so each improves at its own pace.
-
How it speaks and reasons
The adapted model
Trained on examples your teams have validated: your vocabulary and the reflexes of your trade. We train it when the trade has to go into the model itself.
Changes by tested versions
-
What it knows today
The company memory
Facts, files, decisions and rules, each with its source and its permissions. It follows the company day by day, with nothing to retrain.
Changes every day
-
What it is allowed to do
The agents’ skills
Procedures, tools and roles: what to do, for whom, with what approval. We improve an agent without touching the model.
Changes according to your rules
The adapted model, in practice
It learns your business from examples your teams have validated.
Your company’s model
Training examples
-
Procedures
Step requested:
Steps to follow
-
Technical files
Project data:
Written file
-
Trade language
Generic term:
The company’s term
-
Quoting process
Price request:
Structured quote
A trained model, on its own, is just a file. In HUB AI, it gets to work.
And because the three are separate, we pull the right lever for each need. Often, the memory or an agent is enough. We will tell you.
Computing power
The computing power to train your model
Hub AI has its own computing power, dedicated to training business-specific models for companies. Your model is trained and hosted at Hub AI, in Quebec: your data does not leave the country.
- Infrastructure
- Dedicated to training
- Hosting
- In Quebec and in Canada
- Hardware
- None to manage on your side
Excellent for a general question. Not for your company.
A consumer chatbot is excellent for a general question. On its own, it knows neither your clients, nor your rates, nor the way you really deliver. And what you give it goes to the provider, depending on its plan.
- It doesn’t know your business in advance.
Your codes and your exceptions have to be explained to it again.
- It has no company memory.
The company’s files, decisions and rules stay out of its reach.
- It works beside your software tools.
Without configured integrations, you copy, you paste, and the file doesn’t move forward.
- Nobody measures it on your cases.
If it gets worse after an update, you hear about it from your clients.
Three ways to have an AI
| Point of comparison | A consumer assistant One subscription per person | A single piece A trained model or a server | HUB AI Your company’s model |
|---|---|---|---|
| Knows your business | Depends on what you give it in each exchange | Partly, if it is trained | Yes: adapted model, memory and skills |
| Works in your software tools | Depends on the integrations configured | You connect it yourself | Through HUB AI Connect, according to the permissions granted |
| Serves your teams | Everyone uses it their own way | It takes a technical team | In their screens, with permissions by role |
| Quality tracked | Depends on the provider | Rarely measured on your cases | Every version tested in the Creuset, rollback possible |
| Your data | With the vendor, depending on its offer | Depends on the hosting chosen | On machines reserved for your company, in Canada |
| To get started | Simple, but generic | You assemble it yourself | A diagnostic, then a first need |
HUB AI Creuset
No version goes to work before it has proven itself.
Each new version is tested on your real cases, side by side with the one in place. It has to do better, without losing what the other one knew how to do. Then a person decides.
Example: version A learns the business better but forgets what it knew; it is rejected at the second gate. Version B passes all four gates and goes into service, first on a small share of the work.
- 01Tested
On cases set aside, never seen during training.
- 02Compared
With the version in place, criterion by criterion, and on what it already knew how to do.
- 03Approved
A person decides, and the decision is recorded.
- 04Reversible
It starts on a small share of the work. We roll back in one move, without touching the company memory.
A version that learns your business better, but forgets something else? It doesn’t pass.
Your knowledge serves your company. Only yours.
The adaptation to your business learns only from what you entrust to it. You know what goes in, you can say no, and your model stays on your private infrastructure, in Canada.
- Only what you authorize.
Examples your teams have validated. Never your emails in bulk.
- Your consent, and you can withdraw it.
The owner of the company gives consent before any training. If the owner withdraws it, a version built on that consent is not put into service, or is taken out of it.
- Personal information, masked.
It is masked in the examples before training.
- Never for another client.
Your examples train only your model, and they stay walled off.
- Hosted in Canada.
On machines reserved for your company, which HUB AI hosts and maintains.
- The memory stays separate.
The model learns ways of working. Your facts and your files stay in the memory, with their sources and their permissions: changing versions erases nothing.
We start by listening. Not by training.
Sometimes the diagnostic concludes that an adapted model is not needed. We will tell you.
-
Diagnostic
3 to 4 weeks
We interview several people and gather tasks, tools and examples. We see whether an adapted model is really the right lever.
-
A first need
A single case to start: the one where a model that speaks your trade changes the day-to-day.
-
Pilot
Validated examples, training, exam in the Creuset, go-live on your machines. Your teams use it in their real work.
-
Improvement
The corrections come back. We measure at 30, 60 and 90 days, then every quarter, and we withdraw a version if it gets worse.
The questions we get asked. Our straight answers.
“Do we need a technical team?”
No. HUB AI prepares, trains, tests, hosts and maintains the model. Your teams validate examples and correct answers, in their screens.
“What if it gets it wrong?”
It works within limits: what it is allowed to do, and with what approval. It is tested on knowing how to say “I don’t know,” or to refuse. And every correction feeds the next version.
“What if a new version does worse?”
It doesn’t go into service. If a problem shows up afterward, we go back to the previous version, without losing anything from the company memory.
“Will our data train someone else’s AI?”
Never. Your examples train only your model, and they stay walled off, in Canada.
Your business deserves better than an AI that gives everyone the same answer.
Tell us about a file that costs you time. We will tell you whether an enterprise model is the right lever, or whether something else is enough.
- First.
We talk: together we choose the first file, then the start date of the diagnostic.
- 3 to 4 weeks.
Your teams describe and show their real work.
- Then.
Your proposal: scope, price, schedule and measurable objectives.