Any model. Any provider. Chosen by evidence.
We build and run production AI on OpenAI, Anthropic, Google, and open-weight models - on Azure, AWS, Google Cloud, or your own infrastructure. The model is picked per use case by testing it on your data, and the system is built so you can change your mind later.
The leading model families, and what each tends to win at.
These are starting points, not rules. Every recommendation we make is confirmed against your own evaluation set before it reaches production.
OpenAI
GPT and reasoning models · Realtime voice · embeddings · gpt-oss
Agents and tool use, structured extraction, real-time voice, and high-volume API workloads.
Delivered via: OpenAI API · Microsoft Azure
Anthropic
Credentialed →Claude models · Claude Code · Cowork
Long, multi-step instructions, document-heavy reasoning, coding agents, and Claude Team and Enterprise workspaces.
Delivered via: Anthropic API · AWS Bedrock · Google Cloud Vertex AI
Gemini models
Multimodal work across images, video, and audio, very long context, and low-latency, low-cost tasks at volume.
Delivered via: Gemini API · Google Cloud Vertex AI
Open-weight
Llama · Mistral · gpt-oss · Qwen and others
Workloads where data cannot leave your environment, high-volume tasks where per-token pricing stops making sense, and fine-tuned specialist models.
Delivered via: Your cloud or your own hardware
Voice systems add a speech layer - OpenAI Realtime, Deepgram, ElevenLabs, and similar - chosen the same way: by testing on recordings of your real calls.
Deployed where your data and contracts already live.
If your organization already runs on Microsoft, AWS, or Google Cloud, the AI can run there too - inside your tenant, under your security controls, on your existing agreements. Canadian hosting regions are used where the provider offers them.
Direct provider APIs
OpenAI, Anthropic, and Google APIs - the fastest path to new models and features.
Microsoft Azure
OpenAI models inside your Azure tenant, under your existing Microsoft agreements and controls.
Amazon Web Services
Claude, Llama, Mistral, and other models through Bedrock, inside your AWS account.
Google Cloud
Gemini and partner models through Vertex AI, inside your Google Cloud project.
Your own infrastructure
Open-weight models self-hosted on your cloud or on-premises, for data that cannot leave.
Measured on your work, not on a leaderboard.
Define the task and its constraints
What good output looks like, the latency and cost budget, and where the data is allowed to go - including Canadian privacy obligations such as PIPEDA and Quebec’s Law 25.
Build an evaluation set from your data
Real examples from your own work, scored the way your team would score them. Public benchmarks do not tell you how a model handles your invoices.
Test the candidates side by side
Models from several providers run against the same evaluation set, compared on quality, latency, cost, and data handling. The numbers decide, not habit or vendor preference.
Route per step, and keep measuring
Different steps often get different models - a small, fast one for classification, a frontier one for the hard reasoning. When a new model ships, it runs against the same evaluation before anything changes.
Built so you can switch provider without starting over.
The best model this year is rarely the best model next year. Every system we build assumes the model underneath it will change.
A provider abstraction layer
Your application talks to one internal interface, not to a vendor SDK scattered through the code. Switching a model is a configuration change, not a rewrite.
Evaluations that travel with you
Prompts, evaluation sets, and model configuration live in your repository, under version control. Any new model can be measured against your baseline in an afternoon.
Fallback providers on critical paths
Where an outage would hurt, a second provider sits behind the first. If one has a bad day, your customers do not.
Your accounts, your keys, your code
Provider accounts and contracts are in your name, and you own everything we build. Nothing about the system depends on Nisco staying involved.
Already on one provider and want to move? AI Platform Migration → Spend growing faster than usage? AI FinOps →
Models & providers - common questions
Which AI models does Nisco AI Systems work with?
Nisco AI Systems is model-agnostic and delivers on any model from any provider: OpenAI GPT and reasoning models, Anthropic Claude, Google Gemini, and open-weight models such as Llama, Mistral, and gpt-oss. The model is chosen per use case by evaluation on the client’s own data, and different steps of one system often use different models.
Does Nisco AI Systems only build with Claude?
No. Nisco AI Systems holds Anthropic credentials - the Claude Certified Architect certification and the Claude Partner Badge: Claude Code - but delivers on OpenAI, Google, and open-weight models just as readily. A credential reflects depth on a platform, not a limit on which platforms Nisco recommends.
Can Nisco build on Azure OpenAI, AWS Bedrock, or Google Vertex AI?
Yes. Nisco AI Systems deploys inside the client’s own cloud tenant when that is where their data, security controls, and existing agreements live - OpenAI models on Microsoft Azure, Claude, Llama, and Mistral on AWS Bedrock, and Gemini on Google Cloud Vertex AI - as well as on the providers’ direct APIs.
Can our data stay in Canada?
Often, yes. Nisco AI Systems uses Canadian hosting regions where the chosen provider or cloud offers them, and self-hosted open-weight models when data must not leave the client’s environment at all. Data residency is assessed against PIPEDA and, where applicable, Quebec’s Law 25 before a model is chosen.
Can we switch AI providers later?
Yes, and systems built by Nisco AI Systems are designed for it. A provider abstraction layer keeps vendor code out of the application, and versioned evaluation sets let any new model be measured against the current baseline before a switch. Nisco also migrates existing AI workloads between providers as a contracted service.
Does Nisco earn more if we choose a particular provider?
No. Nisco AI Systems recommends the model that performs best on the client’s evaluation, and where it supplies licences it passes them through at cost with no markup. There is no commercial reason to steer a client towards one vendor.
Not sure which model fits?
Bring the use case. We will tell you which models are worth testing, what it would take to find out, and whether AI is the right tool at all.
Book a Discovery Call