At Jama Software, we apply Spec-Driven Development (SDD) within our own engineering and product development organizations. We use Jama Connect® and AI engineering workflows internally across our own engineering and product development teams, giving us firsthand insight into how requirements, traceability, verification, and human oversight help turn AI coding speed into product velocity.
As we’ve expanded SDD across real projects, we’ve established a disciplined approach for turning AI coding speed into greater product velocity. Requirements quality, engineering context, human judgment, testing, and traceability all become more important as development accelerates.
Here’s how the approach works in practice and the results we’re seeing.
How We Scaled SDD Across Engineering
We introduced Spec-Driven Development through a phased, practitioner-led rollout across Engineering. Senior engineering leaders started by applying SDD to real project work, including production-oriented features, and used those implementations to establish practices that could scale across teams.
Hands-on workshops allowed engineers to apply SDD to active projects, develop shared practices, and bring those practices back to their teams. This phased approach allowed us to standardize the workflow around real development work and scale proven practices across teams. Today, roughly 80% of our Engineering teams have applied SDD to real project work, and adoption continues to expand across larger initiatives.
One lesson from the rollout became clear: AI adoption is a workflow change before it’s a tooling change.
Better AI Starts With Better Requirements
As part of SDD adoption, we strengthened business systems analysis practices across participating teams. Teams practiced breaking problems down, defining them clearly, and representing workflows before asking AI to act on them.
The objective was straightforward: understand the problem before asking a machine to help solve it.
Our internal use has reinforced the value of defining product intent before asking AI to generate requirements, specifications, or prototypes. When teams provide clear objectives, constraints, and expected behavior, AI-generated work is more relevant, predictable, and actionable.
Clear intent allows AI to accelerate well-defined engineering work while maintaining alignment with product objectives.
Unambiguous, clearly-defined requirements are the starting point, but AI also needs the engineering context surrounding them. Requirements, specifications, architecture decisions, tests, verification evidence, and the relationships between them help AI understand what the product is supposed to do, which constraints matter, and how the work fits into the larger system.
Strong AI outcomes begin with clear product intent, not prompting alone.
For us, AI has only increased the importance of requirements quality.
How We Give AI the Context It Needs
High-quality requirements become even more effective when AI can access the broader engineering context surrounding them.
Much of that context already lives in Jama Connect. Requirements, specifications, tests, verification evidence, reviews, and the relationships between them provide a persistent record of what we’re building and why.
Through the Model Context Protocol (MCP), our AI coding agents can retrieve relevant requirements and related engineering context directly from Jama Connect as they work, rather than reconstructing that context in each prompt.
The Traceability Information Model™ (TIM) gives that context structure. It preserves the relationships between scenarios, features, requirements, specifications, tests, and verification evidence, so AI isn’t working from a collection of disconnected artifacts.
AI produces stronger results when engineering information includes the relationships that give it meaning.
How Spec-Driven Development Changed Our Workflow
One of the biggest changes we’ve seen with SDD is how engineers are spending their time.
As AI takes on more implementation work, the specification becomes a more important engineering asset. Engineers spend more time upstream examining what the system should do, questioning assumptions, identifying edge cases, and making sure the intent is clear enough to implement and verify.
In practice, the workflow looks something like this:
- Start with clear requirements. The team reviews the intended behavior and relevant product context before moving forward.
- Develop and review the specification. AI helps turn requirements into a more detailed specification, while engineers look for missing constraints, assumptions, edge cases, and other gaps.
- Generate the implementation. AI produces code based on the reviewed specification.
- Verify and validate the result. Tests verify that the implementation satisfies the specification, while human review validates that the resulting behavior meets the intended need.
When engineers identify a missing condition or incorrect assumption, the goal is to correct the specification and underlying context rather than the code alone. That correction can then inform the implementation and future work.
This creates a feedback loop:
Intent → context → specification → implementation → verification → feedback → better context
Over time, the lessons engineers uncover during implementation and verification become part of the product context instead of disappearing in a coding session or temporary AI conversation. The result is a workflow that gets better at expressing engineering intent before AI acts on it.
Human Judgment Becomes More Strategic in AI Engineering Workflows
As AI takes on more execution, human judgment moves earlier in the workflow.
Engineers review requirements and specifications before implementation, challenge assumptions and missing conditions, and determine whether generated work reflects the intended behavior. Verification and human validation provide additional checks downstream.
Product teams use AI to develop and explore ideas while people remain responsible for identifying the customer problems worth solving. People still need to talk with customers and go-to-market teams, establish priorities and boundaries, and give Engineering enough context to build the right thing.
AI changes how the work is performed while human accountability remains essential. People still have to answer the questions that matter most:
- Are we solving the right problem?
- Have we defined the intended behavior clearly enough?
- Have we accounted for every important assumption, constraint, and edge case?
- Does the resulting product actually meet the intended need?
AI can perform more of the work between those decisions, while engineering expertise increasingly shifts toward intent, architecture, tradeoffs, edge cases, and verification.
AI Accelerated Testing and Raised the Bar for Verification
Testing is one area where we’ve already measured substantial productivity gains from AI. Our broader use of AI-generated test cases has delivered:
- 60% faster time-to-coverage
- 11X test coverage
- 3X test creation velocity
These results demonstrate AI’s potential while highlighting the value of connecting every generated test to the requirement or specification it is intended to verify.
Verification has to remain connected to product intent. In our SDD workflow, tests trace back to the requirements and specifications they’re intended to verify. Generated code is accompanied by unit tests, while human validation provides another check that the resulting behavior meets the intended need.
This matters even more as development accelerates. Our SDD workflow scales verification alongside the volume and speed of AI-assisted implementation.
Faster generation increases the value of a strong verification loop.
Early Results: Engineering Throughput Increased Nearly 6X
The improvements are now visible in broader engineering outcomes.
In one recent release, Engineering dramatically increased project delivery by nearly 6x compared with the prior-year release.
These improvements reflect SDD working alongside broader advances in our engineering practices, AI-assisted development, and team workflows. The combination is an important signal: Engineering is delivering substantially more work.
We’re tracking these outcomes through an internal dashboard built from data in Jama Connect, including delivery, throughput, portfolio progress, and quality. As SDD adoption expands, the dashboard allows us to connect workflow improvements directly to measurable delivery, throughput, portfolio, and quality outcomes.
That’s the standard that matters to us. Speed is valuable when it translates into greater product velocity without sacrificing quality or control.
Why Traceability Matters More at AI Speed
As the volume and speed of engineering work increase, maintaining control becomes more important. AI can rapidly increase the volume of requirements, specifications, code, tests, and changes moving through the engineering lifecycle.
Traceability gives engineers the evidence needed to assess an AI-generated output, understand its source, and identify affected downstream work. A specification stays linked to the requirement behind it, a generated test stays connected to what it’s intended to verify, and an upstream change can be followed into affected downstream work.
Traceability provides the evidence engineers need to evaluate that output, understand where it came from, and determine what else may be affected.
As AI generates more work, preserving those relationships becomes increasingly important.
How Product Teams Apply AI-Assisted Workflows
Engineering and Product apply SDD at different points in the development lifecycle. Product teams use the approach to strengthen definition, design context, and expected feature behavior before work moves into implementation.
Product uses AI extensively for upstream work, including use cases, strategy documents, and lightweight prototypes. Teams are applying SDD workflows to bring product definition, design information, and expected feature behavior together before Engineering moves into implementation.
AI-assisted workflows and Spec-Driven Development are enabling Product teams to bring suitable initiatives into monthly release cycles, with smaller efforts moving even faster.
We evaluate SDD as part of the broader AI-assisted workflow and continue expanding measurement as adoption grows.
Applying SDD to our own real-world work allows us to identify the workflows where it delivers the greatest value and translate those practices into stronger customer outcomes.
Scaling the Entire Development Workflow
Faster implementation creates an opportunity to strengthen the entire development lifecycle. We have expanded verification capacity, enriched the product and UI guidance available to AI, and evolved documentation and release practices to support greater engineering throughput.
These improvements help ensure that requirements, implementation, testing, documentation, and release management advance together as AI-assisted development scales.
Coordinating improvements across the development lifecycle allows implementation gains to translate into greater product velocity.
As coding accelerates, teams can scale requirements, specifications, testing, integration, documentation, compliance, and release management together. This end-to-end approach converts implementation speed into sustainable product velocity.
AI coding speed creates the opportunity, but the surrounding engineering workflow determines how much of that speed becomes product velocity.
What’s Next for Our AI Engineering Workflow
We’ll continue expanding SDD across real work, measuring outcomes, and advancing the workflow as its use grows.
SDD provides a disciplined engineering practice for turning AI coding speed into product velocity while maintaining requirements quality, traceability, verification, and the human judgment complex product development demands.
Ready to explore what this could look like for your engineering organization? Get a demo of Jama Connect to see Spec-Driven Development in action.
Want to go deeper on the methodology first? Download the Spec-Driven Development Playbook for practical guidance on building an AI engineering workflow grounded in clear requirements, structured product context, traceability, verification, and human oversight.
