Custom Agentic AI Software Delivery for an IT Software Services Company

Overview

Stackular developed a custom agentic AI software-delivery solution around an IT software services company's established development process.

The company wanted to extend AI beyond individual coding assistance while preserving the way its teams defined requirements, reviewed work, tested changes, and approved software for delivery.

The engagement focused on connecting AI assistance to those existing responsibilities. Approved requirements and project decisions needed to remain consistent as work moved between analysis, implementation, review, and testing. Generated changes needed to meet the company's engineering expectations rather than simply appear complete.

Stackular tailored the solution to the company's delivery workflow, including requirements refinement, shared project context, implementation assistance, automated validation, human approval, and visibility into AI spending.

The result was an AI-assisted delivery process organized around how the company performs and accepts work, rather than treating code generation as an isolated productivity activity.

The Challenge: Make AI Support the Company's Delivery Process

The initiative addressed the gap between helping a developer produce code and helping the company deliver reviewed, tested software that satisfied an agreed requirement.

Code generation was only one part of that process. Requirements still needed to be clear. Design decisions needed to remain available during implementation. Proposed changes needed review, tests needed to run, and people needed to decide whether the work was ready to proceed.

Without coordination across those activities, faster implementation could simply move work into a queue for clarification, review, or correction.

The company also needed a consistent approach to oversight. AI-assisted work had to follow defined expectations about permitted activities, quality checks, approval, and spending. Those expectations could not depend solely on how an individual developer happened to use an assistant.

The objective was to make AI work within the company's delivery process while improving the continuity between its stages.

Stackular's Approach: Tailor Automation to the Company's Workflow

Stackular developed the solution around the company's requirements, engineering guidance, and review responsibilities.

Rather than imposing an unrelated delivery model, the work connected AI assistance to the activities the company needed to perform before software could be accepted.

At a business-process level, that included:

Delivery activity	How the solution supports the company's process
Requirements preparation	Identify ambiguity, propose clearer user stories, and develop acceptance criteria for review.
Planning and design	Use agreed requirements and project guidance to support implementation planning and security review.
Development	Produce proposed code and supporting tests within the approved scope of work.
Review and validation	Evaluate changes using review findings and the project's actual build and test results.
Delivery approval	Retain the human decisions required before software is accepted and released.
Cost oversight	Associate AI spending with the work performed and apply the company's spending controls.

These activities were connected through shared project context and reviewable work. The aim was to support the company's process from requirement to delivery, not merely accelerate the moment when code is written.

AI-assisted delivery workflow

AI assistance follows the company's path from requirements to delivery

1. Agreed scopeCompany's Approved Requirements
2. Assisted workAI-Assisted Planning and Development
3. EvidenceReview and Testing
4. AccountabilityHuman Approval
5. Accepted workDelivery

• Company StandardsEngineering expectations remain part of the work
• Shared ContextRequirements and project decisions remain available across stages
• Cost OversightAI spending remains connected to the work performed

Carrying Approved Requirements Through Delivery

One of the solution's central responsibilities is preserving the meaning of the approved work as it moves between people and stages.

The analyst, developer, reviewer, and tester need to work from the same requirement and relevant project decisions. Otherwise, each handoff creates an opportunity for the intended behavior to be interpreted differently.

The solution maintains that information as shared project context. AI-assisted planning, implementation, and review can therefore use a common understanding of what the company has agreed to deliver.

This makes continuity part of the delivery process rather than leaving it dependent on one person's notes or an individual AI conversation.

Improving Requirements Before Implementation

The solution helps identify incomplete or ambiguous user stories before development begins.

It can propose clearer descriptions, testable acceptance criteria, and scenarios that may have been missed. Where the intended behavior cannot be established from the available information, it asks for clarification rather than silently choosing an interpretation.

Proposed requirements changes remain subject to human acceptance. AI can help improve the input to development, but it does not become the authority for redefining the requested work.

This aligns requirements assistance with the company's approval process: clarify the work, review the proposed changes, and establish what will be tested before implementation proceeds.

Applying the Company's Engineering Guidance

The solution carries project guidance into the activities performed by AI.

Implementation, review, and testing are not treated as unrelated tasks. They are evaluated in the context of the approved requirement and the expectations established for the project.

This allows the company to use AI assistance without abandoning its own approach to engineering quality. Proposed work still needs to be understandable, reviewable, and supported by evidence that it behaves as intended.

The goal is not to make every project identical. It is to keep the AI-assisted work aligned with the requirements and standards that apply to the work being delivered.

Connecting Implementation to Review and Testing

Generated code is treated as proposed work, not as proof that a task is complete.

The solution uses actual builds and tests to evaluate changes. Review provides another opportunity to identify problems before the engineer's acceptance decision.

When validation fails, the findings can inform corrections and another evaluation. The feedback comes from the work itself rather than only from the AI's explanation of what it believes it completed.

That distinction is important to the company's delivery process. Engineers retain responsibility for accepting changes, supported by review findings and test results rather than a generated assurance that everything is ready.

Preserving Human Approval

Stackular incorporated human oversight into the AI-assisted workflow.

Requirements improvements are proposed for acceptance. Generated work is reviewed. Release remains subject to a person's approval.

The solution therefore distinguishes between performing a task and accepting responsibility for its result. AI can assist with analysis, implementation, corrections, and preparation, while the company's team retains the decisions about scope, quality, and readiness.

This keeps automation aligned with the company's delivery responsibilities instead of allowing the pace of generated work to bypass them.

Governing AI Activities Within the Workflow

The solution provides controls over what AI-assisted activities are permitted and records relevant actions and decisions.

This gives the company visibility into how AI participated in a work item and the evidence associated with that work. Oversight becomes part of the delivery record rather than something reconstructed after a concern arises.

The emphasis is on the company's own operating expectations: what work may be performed, what checks are required, and where a person must remain involved.

Connecting AI Spending to Project Work

AI spending is tracked in relation to the work performed.

Rather than relying only on an aggregate usage total, the company can examine the activity that generated a cost and relate that activity to a delivery item.

Spending controls provide another part of the operating model. The objective is to keep AI consumption connected to useful work and visible to the people responsible for delivery economics.

This adds cost oversight to the same process that already considers requirements, review, and quality, without assuming that more AI activity automatically produces more value.

Delivered Capabilities

AI Assistance Aligned with the Company's Process

The solution connects requirements preparation, planning, development, review, testing, and delivery approval around the company's way of working.

Continuity Across Delivery Handoffs

Approved requirements, decisions, and project guidance remain available as work moves between people and stages.

Reviewable Requirements Improvements

The team can evaluate proposed clarifications and acceptance criteria before they become the basis for implementation.

Evidence-Based Validation

Build results, test results, and review findings support the decision to accept a proposed change.

Human Oversight and Traceability

The company retains approval responsibility and visibility into the AI-assisted activities involved in delivery.

Work-Level Cost Visibility

AI spending can be associated with the work that generated it, providing a clearer basis for understanding usage.

Key Delivery Decisions

Fit AI to the Process, Not the Process to a Tool

The engagement centered on the company's delivery responsibilities and expectations. AI assistance was applied to those activities rather than treated as a separate workflow with its own definition of success.

Address More Than Code Generation

Requirements, planning, review, and testing all affect whether software can be delivered. The solution connects those stages so implementation is not accelerated in isolation.

Preserve Approval of Requirements

Clarifying a requirement can change what the team is expected to build. The solution therefore makes proposed improvements reviewable instead of silently changing the scope of work.

Separate Generation from Acceptance

Producing a change and deciding whether to accept it are different responsibilities. Automated validation and review support that distinction while people retain the approval decision.

Keep Quality and Cost in the Same Operating Model

AI activity needs to be useful, reviewable, and economically visible. Spending oversight is part of the delivery process alongside technical validation and human review.

Lessons from the Engagement

Start with How the Organization Accepts Work

The value of AI assistance depends on whether it helps produce work the organization can accept. Requirements, evidence, review, and approval matter alongside implementation speed.

Preserve Context at Handoffs

An agreed requirement should not become a different interpretation every time another person picks up the work. Shared context helps maintain continuity across analysis, development, review, and testing.

Improve the Input Before Accelerating the Output

Unclear requirements leave too much room for interpretation. Helping the team clarify them before development is as important as helping generate the implementation.

Keep Accountability Explicit

Automation can perform work without becoming responsible for accepting it. Human ownership remains important wherever decisions affect scope, quality, or release readiness.

Frequently Asked Questions

Discuss AI-Assisted Software Delivery

Stackular helps organizations apply AI to their development processes while preserving the requirements, engineering standards, and approval responsibilities that matter to their work.

Talk to Stackular about improving your software delivery process with purpose-built AI assistance.