From Test Automation to AI-Native Development
By: Hélder Ferreira, Director of Product Management, Testing Tools, Sembi
Tech News Europe - AI adoption across Portugal is prominent. For instance, 62% of Portuguese respondents of a recent study use AI tools regularly, well above the European average of 52%. This active adoption is translating into the software development lifecycle (SDLC). Around 65% of developers expect their roles to be redefined in 2026, shifting from routine coding toward architecture, integration, and AI-enabled decision-making.
The challenge is that, as of now, testing workflows mainly employ AI for individual tasks. It can generate test cases, turn manual tests into automation scripts, and prioritize which tests to run based on risk and relevance. While helpful, this narrowed focus can be limiting when organizations try to extend AI capabilities to other parts of the development process.
Reaching the next phase of development requires software leaders to understand the complexities of expanding AI’s role across the SDLC and establish an architectural framework that supports its continued evolution.
The AI Adoption Problem Isn’t Generational - It’s Context:
As organizations apply AI for more SDLC tasks, they tend to face two major constraints. The first is that generic models still don’t know how a product is supposed to behave. For example, models like Copilot and Cursor lack a nuanced understanding of the context required for an end product to function correctly. Developers end up pointing their AI agents at code, expecting them to infer intent. The result may be plausible work, but the software often drifts from how it is actually supposed to operate.
On top of this, tracing every requirement, test, and existing behavior that an AI-generated change touches is increasingly difficult. AI has enabled asynchronous agents to increase lines of code output by 658% in some cases. The result is that there are often more changes to review than human developers can keep up with.
Feeding AI agents regression tests that cover business rules, edge cases, historical failures, and expected behaviors early in the development cycle is one possible solution. However, this only addresses the behavioral pain point, and even then, tests may still be stale or incorrect. To make tests an effective lever in AI operations, testing itself will have to become a continuous part of how AI builds, evaluates, and eventually acts across the SDLC.
Three Steps to Connect AI Across the SDLC:
Transitioning to the next stage of the AI SDLC requires organizations to evolve their current development infrastructure. This shift is often most effective over a three-part process:
● Step one:
Connect AI to the testing knowledge enterprises already have. The most effective way to address the challenge of AI software drift is to provide the system with the information required to build correctly. This means the test cases, requirements, execution results, and other test data that developers deem necessary. Of course, this must be done securely, which is why many organizations that initiate this step do so via a model context protocol (MCP). This connection layer links AI tools to relevant testing context directly within their existing workflows, enabling AI to pinpoint coverage gaps, translate requirements into functional tests, and continuously update the testing suite.
● Step two:
Turn tests into a live context layer for development. Once development teams feed this information into AI models, they can begin building. This technology can leverage the identified tests to work toward the expected behavior before and during development, while execution results enrich the available context. The existing requirements, tests, and code can identify the risks associated with a proposed change, giving AI enough context to preserve existing behavior during development.
● Step three:
Put agents to work in context. The eventual goal should be to turn this foundation into connected workflows that require less manual intervention. AI could move beyond generating code or drafting tests to evaluating how a change stacks up against relevant tests, surfacing potential risks and missing coverage, keeping test suites up to date, and helping implement code against test-defined behaviors. Developers could then spend less time on broad manual review and debugging, instead focusing on applicable markers where the human perspective remains vital.
Consider how this progression might play out in practice. A bank wants to create a new password reset flow that supports SMS-based resets. To enable AI to build this new capability, the organization could leverage an MCP to provide it with the appropriate testing context. This means requirements and tests that define how long the reset link remains valid before expiring, how many reset requests a user can make before being locked out, or whether a successful password change requires the user to log in again.
With that information, AI could begin the actual development. It could create the new feature while accounting for existing functionality and behaviors. If execution results reveal that the feature isn’t meeting an expected behavior, AI could use that feedback and adjust its approach. This technology could handle the entire process, flag any relevant issues for developers to assess, and then wait for developer approval before shipping the release.
The scenario is hypothetical, but the principle behind it is not. It highlights how AI agents can build toward the expected behavior while keeping existing operations intact when software development teams give agents context up front.
From AI Assistance to a Connected SDLC:
AI adoption in Portugal may be prominent, but adoption alone does not guarantee its effectiveness. An agent can generate passable code, while still lacking the nuanced understanding of requirements, historical behaviors, edge cases, and dependencies that determine whether the code functions correctly in production.
To move on to the next phase of software development, enterprises must integrate all relevant information, securely feed that context to AI agents, and enable them to act on it. This foundation allows AI to build and evaluate changes against expected behavior. Eventually, this baseline will facilitate connected workflows that can operate with greater autonomy.
About the Author:
Hélder Ferreira is Director of Product Management at Sembi. As a Product Director with a background in Mathematics & Computer Science, Hélder brings over a decade of experience building impactful software products in the internet and enterprise space. His journey has taken him from hands-on product ownership to leading cross-functional teams and driving strategic vision at scale. Today, Hélder focuses on shaping product strategy with a strong emphasis on innovation, user-centric design, and the transformative potential of AI. Hélder is particularly passionate about leveraging AI to deliver smarter, more efficient, and deeply intuitive experiences that solve real customer problems.
