Have you found a technology partner capable of delivering your AI roadmap?

The key checks to confirm that your technology partner has the AI expertise it claims to have before you trust it with your roadmap.

Today, almost every software provider claims to have AI engineers. The title is easy to adopt: a chatbot demo can be built in a weekend and familiarity with the latest models proves very little. Before entrusting your company’s roadmap, data or architecture to an external team, the question you need to ask is not, “Do they know the newest model?” but, “Can they safely operate an AI-dependent system within our product and engineering environment?”

Here is a practical way to find out before signing a contract.

Evaluate the whole company’s AI capability, not just one specialist’s expertise

Ask for a few concrete examples of applied AI projects that show the business problem, the architecture selected, how quality was measured and what was ultimately deployed in production. Then determine whether that expertise is distributed across the organization or concentrated in one person’s knowledge: who reviews model and architecture decisions, how engineers keep their skills up to date and what happens if the assigned expert leaves the project.

At this point, a structured internal program matters more than a job title. Cactus develops and certifies its engineers through the CactAI Excellence Program, an internal program that provides a defined path for developing and assessing AI skills, rather than relying on the informal and unstructured learning that often takes place within companies.

Interview the engineers who will actually work on your project

Speak with the proposed technical lead or at least one of the engineers who will join the project, not only the sales team. The aim is to assess their real judgement across four areas: data, model and architecture trade-offs, evaluation and production and MLOps.

Security is not an afterthought

Get clear answers on whether client data is shared with third-party models, whether those providers can retain it or use it to train their models and how data residency, encryption, credentials and keys, access controls, retention periods, and ownership of code, prompts and derived assets are managed.

For LLM and agent-based systems, the team should understand risks such as prompt injection, sensitive-information disclosure, unsafe output handling, supply-chain vulnerabilities, excessive permissions and uncontrolled model costs. These are some of the main areas OWASP currently highlights in relation to generative AI.

Cactus’s AI principles, set out in the Cactus AI Manifesto, also cover human involvement, privacy, security, accountability, client intellectual property,and transparency.

The definitive test

Once you have worked through the points above, the decision comes down to four questions:

• Can they define the right problem?

• Can they provide measurable evidence?

• Can they build reliable production software around the AI component?

• Can they protect your data, users, and architecture?

A credible partner answers these questions with transparency and confidence.

Not sure whether your AI roadmap has a talent, architecture, or delivery gap? Talk to Cactus about the main bottleneck and the engineering capabilities your project actually needs.