Frequently asked questions about Artificial Intelligence

It depends on the task. For tabular predictive models we usually need at least a few thousand rows of clean, labelled historical data. A churn prediction model for a subscription business, for example, worked well with 8,000 customer records spanning two years. Computer vision projects require several hundred annotated images per class at minimum; for a defect-detection system we recently built, the client supplied around 1,200 images and we augmented them to 5,000. Our AI readiness audit will tell you exactly where you stand before any money is committed to model development.

An AI readiness audit takes two weeks. Most build projects run between six and fourteen weeks depending on complexity. A straightforward demand-forecasting model for a single product line can be done in six weeks. A multi-language chatbot with integration into three back-end systems will take closer to twelve. We break every engagement into phases with fixed milestones so you always know what is coming next and can plan internal resources accordingly.

Not during the project. We handle all model development and deployment. If you want to bring maintenance in-house later, we provide a handover package that includes full source code, architecture diagrams, a model card documenting training data and known limitations, and two half-day training sessions for your engineering team. Several of our clients run their models independently within six months of handover.

We deploy on AWS, Google Cloud, and Microsoft Azure. The choice usually depends on what the client already uses. If you have an existing AWS account with VPCs configured, we deploy there. If you have no cloud presence yet, we recommend AWS because our team has the deepest experience with its ML-specific services like SageMaker and Bedrock. We never lock you into a proprietary platform; all our code runs on standard open-source frameworks.

Yes. We sign a non-disclosure agreement before seeing any data. All data transfers happen over encrypted channels, and we work inside your cloud environment rather than copying data to our own servers. For clients in regulated industries like healthcare or finance, we follow sector-specific compliance requirements and can provide evidence of controls for your auditors. Once a project ends, we delete all local copies of client data within 14 days and provide written confirmation.

We agree on a minimum performance threshold during the scoping phase. If the model does not meet that threshold after a reasonable number of training iterations, we present the findings transparently: what worked, what did not, and why. Common reasons include insufficient data volume, noisy labels, or a target variable that is genuinely unpredictable. In those cases we recommend concrete next steps, which might be collecting more data, relabelling, or reframing the problem. We do not charge for additional iterations needed to hit the agreed threshold.

That is the whole point. Every model we build is wrapped in a REST API with documented endpoints, authentication, and rate limiting. If your ERP, CRM, or internal dashboard can make an HTTP request, it can call our model. We also provide sample integration code in Python, JavaScript, and C#. For more complex integrations, such as embedding predictions inside a Salesforce workflow or a Power BI dashboard, we scope the integration work as part of the project.

All build projects are fixed-price. We quote after the discovery phase, and that quote does not change unless you request additional scope. We invoice in three stages: 30% at kick-off, 40% at the midpoint milestone, and 30% on delivery. The ongoing model management retainer is billed monthly. Cloud infrastructure costs are separate and billed directly by the cloud provider to the client, so there is no markup from us.