Many teams can now build in days what used to take weeks. Their planning habits have not changed. Requirements still arrive as long documents, reviews still wait for meetings, and nobody is sure the finished feature solves the right problem.
This is the gap AWS tries to address with AI-DLC, the AI-Driven Development Life Cycle. It is worth understanding, and it is also worth questioning. This post covers what it is, where it helps, where it can fail, and what it changes for product consultants and fractional product managers.
What AI-DLC is, in plain terms
AI-DLC is a software engineering methodology from AWS. Its core idea is that AI should take part in the whole development process, not only in writing code. Adding an AI assistant to a traditional life cycle leaves the slow parts, such as long planning and fragmented collaboration, mostly untouched.
The model has three phases:
- Inception: the team states the intent, answers clarifying questions, and breaks the work into units of work.
- Construction: AI proposes designs, code and tests, and people review them.
- Operations: deployment and running the system, with approval for high-impact actions.
The vocabulary changes too. A cycle that takes hours to a few days is called a bolt instead of a sprint. Epics give way to units of work. AWS has released the workflows as an open-source project, and the approach is built into its Kiro IDE.
The thread running through all of it is human approval. AI drafts and asks questions. People decide.
Where the bottleneck moves
When building gets faster, the cost of a wrong decision goes up. A team can now ship the wrong feature in a week instead of a quarter.
I wrote about this in AI Made Shipping Fast. It Made Deciding Slower. Feedback still depends on manual steps that many users skip, so fast building does not automatically produce fast learning. AI-DLC puts more weight on the front of the process because that is where intent is set and where most expensive mistakes begin.

A real example: days of work, weeks of waiting
I saw this gap in my own work. In AI-DLC terms, I had just finished an inception phase: research and development for a new feature, a competitive analysis, a clear scope, and AI-generated mockups. That took a few days.
Then the approval gate at the end of inception took a couple of weeks. The work waited for leadership sign-off.
The feature could have shipped within days of those mockups. Instead, the gate took longer than inception itself, so the approval step set the pace, not the build. The tools were fast. The decision path was not.
This is where AI-DLC asks a hard question. If a cycle is measured in hours or days, a two-week approval queue is not a small delay. It is the whole cycle. That is not a workable pace in this industry now.
What it asks of an organization:
- Decision power inside the team. The people closest to the problem, usually the product owner, a designer and an engineer, should be able to approve work within agreed limits.
- Decision rights set by risk, not by habit. Low-risk, reversible changes go straight through. High-risk areas such as pricing, security, legal exposure and brand still go to leadership, with a short, time-boxed review.
- A team formed for speed. Small and cross-functional, with one clear owner for the intent, so that nothing waits for a handoff.
- Guardrails from leadership up front. Leaders set outcomes and boundaries in advance instead of reviewing every output after the fact.
There is a balance here. Leadership review is not always waste. Some decisions are costly to reverse, and some teams have not yet earned the trust to decide alone. The answer is not to remove approval. It is to decide in advance which decisions need it, who makes them, and how long they are allowed to take.
What a product consultant or fractional PM does in this model
This part is my own analysis, not something AWS prescribes. If the model works as described, four jobs become more valuable.
- Owning inception. Someone has to frame the problem, define the outcome, and decide what is in or out of scope before the AI starts generating. This is product discovery work.
- Supplying context. AI output is only as good as the context it gets. Customer insight, market knowledge and business constraints rarely sit in a repository. A PM brings them in.
- Validating, not just approving. An approval gate only helps if the reviewer can judge whether the result solves the right problem. Rubber-stamped gates are worse than no gates, because they look like control.
- Helping the team change its habits. Moving from weekly sprints to cycles of hours or days changes meetings, roles and metrics. An outside consultant can introduce that change without the history an internal lead carries.
For a freelance or fractional PM, this shifts the work. Less time goes into writing detailed tickets by hand. More goes into discovery, outcome metrics and decision quality. For small companies without a head of product, a part-time PM focused on inception and validation can be a practical fit.
How it fits with Kanban and Scrum
AI-DLC and Kanban solve different problems. Kanban manages the flow of work. AI-DLC defines how work is planned, built and approved with AI. They combine well.
- A Kanban board can show where each unit of work sits: intent, inception, construction, validation, operations.
- A WIP limit on the review column matters most. AI can generate faster than people can review, and the limit makes that bottleneck visible.
- Flow metrics such as cycle time and throughput give you measurement that AI-DLC does not prescribe.
Scrum is the harder fit. Fixed sprints sit awkwardly beside cycles measured in hours. That is the argument in Scrum Needs to Redefine Itself. Scrum does not have to disappear, though. Product Owner accountability, inspection and adaptation still hold up. What strains is sprint length, estimation and backlog refinement as they are usually practiced. For more on the flow side, see Why Kanban Builds a Stronger Agile Mindset and Transitioning from Scrum to Kanban.
Where AI-DLC can go wrong
A balanced view needs the failure modes.
- Review fatigue. If every unit of work needs human approval, reviewers can become the new bottleneck. Over time they approve faster and read less.
- Confident but wrong output. AI can produce clear, well-structured plans that rest on a wrong assumption. Clear writing is not the same as correct thinking.
- Process overhead for small work. The framework says it adapts to the size of the task, but teams can still apply the full ceremony to a simple fix.
- Weak context. If the team has no shared understanding of the customer, the AI will fill the gap with plausible guesses.
- A young method. It was introduced in 2025. There is little long-term evidence about how it performs across different teams and products, so treat vendor claims with the same care you would give any new methodology.
Who should not adopt it yet
- Teams without clear product ownership. Nobody will own inception, so the AI will set the direction by default.
- Teams with fragile engineering practices. Faster code generation on top of weak testing and deployment just produces problems sooner.
- Regulated environments where approval and audit rules are not yet mapped to the new gates.
- Teams that want a process to replace judgment. AI-DLC moves judgment earlier. It does not remove it.
For these teams, a smaller step often works better: use AI inside the current process, measure what changes, then decide whether to change the process itself.
A practical way to start
- Pick one low-risk product area and run a single cycle with a clear intent and a named owner.
- Define what each approval gate checks before the work starts.
- Limit work in progress at the review stage, not only at the build stage.
- Track cycle time from intent to production, and how often AI output needs rework.
- Review after a few cycles and change the process based on what you see.
The takeaway
AI-DLC reflects a real shift: building is getting cheaper, and deciding is not. That makes product judgment the scarce skill, and it is why consultants and fractional PMs have a clear role in it. The method itself is still young, and it suits some teams much better than others. Adapt it to your context instead of copying it whole.
If your team is trying to work out where AI fits in planning and delivery, I am happy to talk it through at tahirshahzad.com/contact/.
Sources: AWS DevOps and Developer Productivity Blog, “AI-Driven Development Life Cycle: Reimagining Software Engineering”; AWS Builder Center overview of AI-DLC; AWS Labs aidlc-workflows repository.
