Where Does AI Fit in Scrum & SAFe?

The moment teams first “feel” AI in Agile

AI, in a lot of organizations, comes without an over-the-top announcement. It just appears.

• When a Product Owner places an epic in a tool, they are given a structured, draft backlog.
• When a Scrum Master submits a retro theme, they will receive a summary with possible root causes.
• When an RTE puts in their objectives and dependencies, they will receive a risk heatmap for their PI Planning.

The hurdle that most teams encounter once they begin to use this tool is:

“If AI can suggest plans, risks, and priorities… what exactly is still our job?”

In each scenario, the experience is very similar: The team is able to move much faster than before; however, they encounter the more challenging question very quickly:

Agile can be “AI driven” but can only be done so through the use of augmenting the agile approach with the use of AI. Therefore, Agile can be enhanced and improved through the use of AI, as long as the parts of Scrum or SAFe that need to be human, remain human, and are not replaced by AI.

The moment teams first “feel” AI in Agile

AI, in a lot of organizations, comes without an over-the-top announcement. It just appears.

• When a Product Owner places an epic in a tool, they are given a structured, draft backlog.
• When a Scrum Master submits a retro theme, they will receive a summary with possible root causes.
• When an RTE puts in their objectives and dependencies, they will receive a risk heatmap for their PI Planning.

The hurdle that most teams encounter once they begin to use this tool is:

“If AI can suggest plans, risks, and priorities… what exactly is still our job?”

In each scenario, the experience is very similar: The team is able to move much faster than before; however, they encounter the more challenging question very quickly:

Agile can be “AI driven” but can only be done so through the use of augmenting the agile approach with the use of AI. Therefore, Agile can be enhanced and improved through the use of AI, as long as the parts of Scrum or SAFe that need to be human, remain human, and are not replaced by AI.
AI vs Automation vs Analytics (and why teams confuse them)

Before you can choose “where AI fits,” you have to have language that distinguishes among three very different concepts.

AI (Artificial Intelligence)

AI represents a system capable of generating, classifying, predicting, and suggesting through patterns found within the data. In today’s Agile development world, this means models of machine learning and generative AI capable of summarizing text, creating artifacts, and suggesting options.

Automation

Automation follows predetermined procedures accurately. Automation follows rules thus: If event X occurs, then action Y follows. Automation does not “reason.” Automation decreases manual tasks as well as waiting time.

Examples of Agile delivery:

• Autocreate a Jira ticket based on a production alert trigger.
• Auto-routing for pull requests for review based on file ownership.
• Auto-run tests on every commit.

You must differentiate between three different definitions of what is meant by “AI” in terms of its function and how it relates to your business. 

Analytics

Data transforms into information through analytics (metrics, trends, dashboards). Analytics provide metrics and trends to answer the questions of what happened and what is currently happening. Sometimes, it provides statistical forecasts; however, analytics do not create (i.e., ‘generate’) any content as does Generative AI (GenAI).

Examples of Analytics:

• Cycle time charts, burn-up and flow diagrams
• Trends in the number of defects by component
• The distribution over time of lead times
• The importance of distinguishing this difference
• The different types of output from AI can create confusion for the team members as to which output is correct.

Teams that do not understand that AI and Automation can produce different types of outputs will tend to have unrealistic expectations of reliability from AI if AI is considered as Automation.

Teams will put more trust into the information presented on the dashboard if they consider it to be akin to AI and stop asking the “why” question.

Teams will automatically use decisions based solely on the output of the system instead of incorporating judgment/advice from other team members if they consider Automation as being akin to Judgment.

A helpful guideline for team members is to think:

Automation = executes, Analytics = measures, AI = suggests, Humans = decide.

---

What stays the same: empiricism and accountability are not optional

Empiricism is the basis of Scrum, whereby decisions are made based on what has been observed. This observation is carried out by the ability to provide transparency, inspect and allow for a way to adapt. These three pillars of the Scrum framework are not “tools for making a choice” but are the basis why Scrum works with very complex tasks. The Scrum framework has been established to promote the existence of these pillars by providing the events.

Empiricism doesn’t get outsourced to AI

While AI can surface patterns quicker, it still requires empiricism by humans to:

• agree what “transparent” means for this product
• check the right evidence (not just the available evidence),
• adapt decisions with context and courage.

Additionally, Scrum also warns of predictability in complex systems. Forecasts and artifacts are useful, but they neither replace nor substitute empiricism. It only gets more dangerous when AI produces optimistic predictions.

Accountability stays human (Scrum) and role-based (SAFe)

Accountability in Scrum is defined in three categories in the Scrum Team: Product Owner, Scrum Master, and Developers. AI can help to assist these accountabilities but   is not entirely responsible for it. all the accountabilities of their team.

Accountability is also extended across teams within SAFe. Agile Release Train (ARTs) consists of multiple teams and PIs that work towards a common goal. PI Planning provides opportunities for teams, stakeholders, and leaders to gather and set goals.

In a sense, both SCRUM and SAFe allow for Artificial Intelligence (AI) to assist with fulfilling those obligations, but ultimately, the individual team member’s commitment is necessary for fulfilment.

Where AI supports empiricism in Scrum (without breaking it)

Think about AI as a lens, not a steering wheel. It can strengthen empiricism when used deliberately.

Use AI as a lens to clarify the science and implement it more effectively.

Transparency: making work and decisions easier to see

• Transparency is not achieved through:
• information scattered across multiple documents (5),
• inconsistent information (all with multiple versions),
• information that is so dense that no one will read it.

Artificial Intelligence assists in:

• summarizing long conversations into decisions and rationales;
• creating a standard backlog item template;
• extracting candidate acceptance criteria from user research notes; and
• translating jargon into plain language for stakeholders.

AI does not produce the decisions, but rather helps to make the present state more visible through transparency.
Inspection: noticing signals humans miss

The key is that inspection still happens in Scrum events, AI can prepare evidence, but the team inspects and decides.

• The reason inspections fail is because the team inspected:
• Something that doesn’t matter (vanity metrics),
• After the sprint and/or PI (too late),
• Without a shared understanding of what they were all inspecting together.

AI can help with the inspection process in many ways:

• Group problems by their likely cause(s)
• Identify common obstacles from the Daily Scrum meeting
• Find abnormal patterns in cycle time and correlate them back to dependencies
• Gather the key themes from the feedback received at the Sprint Review.

Ultimately, even though AI will prepare the data for you to inspect, the team is responsible for actually inspecting the data during the Scrum Events.

Adaptation: turning evidence into choices

The reason why adaptation fails is when teams have knowledge about the problem and either choose not to address it or address it in a manner that generates even more waste.

AI can quicken the adaptation process by:

• proposing experiment options (“If hypothesis X, propose a test with A/B or a spike”),
• creating a “next Sprint improvement story” from the retro activities,
• developing risk response strategies for PI goals.

Adaptation remains a human function that involves risk, ethics, impact on customers, and feasibility.

Scrum events: where AI helps (decisions, speed, insight) without replacing humans

Scrum events are scheduled so that they can be inspected and then adapted as needed based on the findings during the inspection cycle. Artificial Intelligence (AI) will provide the most value in helping to amplify the objectives of each event, rather than taking away from these objectives by turning them into a formal report.
1) Sprint Planning: faster clarity, better options

The Sprint Planning meeting provides answers to three basic questions: “why do we want to achieve this Sprint?”, “what can we achieve?”, and “how are we going to achieve this?”. AI has the ability to assist in Sprint Planning in the following ways:

• Transform a product goal into multiple potential sprint goals (rather than requiring one to choose the correct goal).
• Create Frameworks for Breaking Large Items into Smaller Items.
• Assist in drafting Task breakdowns that are based upon historical data from Similar PBIs.
• Identify Hidden Dependencies based on either Architecture Notes or Linked Tickets.

In Sprint Planning, the following decisions must be made by Human Participants:

• Choose an appropriate Sprint Goal.
• Determine Scope and Make Trade-off Decisions.
• Agree upon a Definition of ‘Done’ within this context.

A Healthy Pattern may consist of three candidates created by AI for a Sprint Goal, which the team selects from or edits to make their Sprint Goal. An Unhealthy Pattern may consist of AI creating a Sprint Planning Document, which then becomes the contract that management will follow during Sprint execution.

2) Daily Scrum: signal detection, not surveillance

Daily Scrum are mainly for developers to get an up-to-date view of their progress in relation to their goals and plans for the future.

Some of the ways AI can assist with (and ultimately replace) this task are:

• summarizing what the developers did yesterday and what the developer considers the biggest risks for today,
• showing which tasks are stuck and have not made any progress,
• providing an overarching question on which task blocks progress against the goal (i.e. what specifically is preventing the team from reaching its goal).

There are also some things that must continue to be performed by humans:

• cooperation and negotiation in real-time,
• choosing to collaborate, plan differently, or request assistance,
• psychological safety within partnerships.

In the past, there have been warnings from industry studies indicating that AI does provide increased productivity for developers; however, if leadership is unable to fix the friction and lack of clarity in collaboration, the benefits of using AI will be offset by the pain of inefficiency. This type of problem is exactly what the Scrum Events are designed to highlight, not hide.

3) Backlog refinement: better drafting, stronger shared understanding

While Scrum Guide does not provide a specific outline of a refinement activity, it is Source of Backup specification for achieving Product Backlog clarity through ongoing refinement.

AI can assist with this refinement activity, including:

• Creating user story narratives based on existing summary research
• Providing sample acceptance criteria for user stories
• Identifying ambiguity in acceptance criteria (e.g., “what is ‘fast’ in this context?”).
• Formulating a lightweight glossary of recurring terms and terminology.

However, human contribution will continue to play a critical role in:

• Product owner accountability for value ordering and customer trade-offs
• Determine the trustworthiness of any evidence produced by AI
• Validation of the actual user need for an item vs. an assumption based upon AI-generated artifacts.

4) Sprint Review: faster synthesis of feedback and outcomes

The Sprint Review examines the results and adjusts the Product Backlog to reflect the feedback from the stakeholders. In the future, AI will likely be able to help in the following ways:

• To identify common themes from the stakeholder feedback.
• To indicate which Backlog items the stakeholder’s feedback relates to.
• To provide prompt messages to remind stakeholders of the changes to their environment.
• To provide clear messaging of decisions made and any planned work (next steps).

However, there are certain critical activities that will always require a human element. These include:

• Negotiating with stakeholders.
• Deciding what work should be pursued next and why.
• Ensuring transparency in the Team’s communications (i.e. avoiding “AI-washed” communications).

5) Sprint Retrospective: better patterns, better experiments

Retro is looking to enhance the quality and efficiency of their process by utilizing artificial intelligence (AI).

Examples of how AI can assist Retros include:

• Grouping Retro notes into Themes
• Identifying repeated blockers experienced during Multiple Sprints
• Providing Suggestions for Experimentation Formats (Start/Stop/Continue, 5 Whys, Fishbone)
• Creating a single improvement backlog item that identifies a Successful Measure.

What needs to remain human led:

• Building Trust,
• Understanding Sensitive Team Dynamics,
• Determining what Changes are Realistic and Worth Trying.

You might consider creating a sample Creative Case Box with the title “Retro with AI.” The content in this box could present an example of how a team uses AI to cluster their notes and subsequently affirms the clustered themes manually, before selecting a single experiment to undertake.

Where AI fits in SAFe planning and delivery (without weakening alignment)

Agile(SAFe) helps organizations to plan their Agile Deliveries at many levels across teams & value streams. AI’s job at that point is to support organizations by reducing friction around aligning to/amongst Teams, managing dependencies & establishing flow, while still having people make commitments/prioritization decisions.

PI Planning: faster dependency and risk visibility

PI Planning creates alignment within Teams, stakeholders and creates PI Objectives that Teams commit to and deliver upon as directed by Leadership. AI can help with:

• Extraction of dependencies defined by Features; and past Integration Events
• Initiating Development of Risk Categories that can be used as checklists during ROAM to categorize Risks
• Summarizing Capacity Constraints based on historical Flow patterns
• Generating Draft PI Objectives using Feature Intent to assist Teams in arriving at a single set of objectives.

However, many of the tasks must still require human judgement, such as:

• Negotiating cross-Organizational Commitments
• Resolving dependency trade-offs
• Deciding what is a “no-go” for the organization.

[CREATIVE: Diagram – place here – “PI Planning with AI assist”: Inputs → AI suggests dependencies/risks → teams negotiate → committed PI objectives]

Continuous Delivery Pipeline: accelerating exploration-to-release learning

SAFe’s Continuous Delivery Pipeline integrates Exploration, Integration, Deployment and Release on Demand. By using AI as a means to accelerate the learning loops associated with both Exploration (identifying customer signals) and Delivery (supporting quality practices), teams will have a faster path to successful software development.

Below are examples of how AI aids teams in the various aspects of their Continuous Delivery Pipelines:

• Continuous Exploration: AI will analyze customer feedback and support tickets and identify common themes, which are opportunities for improvement.
• Continuous Integration: AI can serve as an assistant to help engineers create test cases and refactoring patterns.
• Release on Demand: AI can assist in drafting Release Notes and identifying potential Risk Areas in Change Logs.

Inspect & Adapt: turning large-scale evidence into better portfolio choices

As information grows, it can become overwhelming for teams to manage. With help from AI, teams will be able to:

• Distill goals into performance indicators.
• Identify issues across multiple departments (or service areas) that contribute to delays or inefficiencies.
• Develop ideas for the next Planning Increment.

However, for SAFe’s Learning Events to provide value, leaders must treat AI-generated insights like information that can help them make informed decisions, not as absolutes.

How to use AI in Agile without creating command-and-control behavior

The failure of most implementations of “AI in Agile” is due, not to the ineffectiveness of AI, but instead due to the use of AI’s outputs as justification for control over others by organizations.

The command-and-control trap looks like this

• Leadership treats an AI-generated forecast as a delivery commitment.
• Daily Scrum notes generated by AI are monitored by organizations to track team performance.
• AI’s analysis of cycle time is leveraged by organizations to put pressure on individuals rather than improve the cycle time of the system.

This approach is contrary to Scrum’s intent,Scrum makes work and problems visible so that teams can adapt their process not to control employees’ actions.

Guardrails that keep AI aligned with Agile values

1. Team-run Artificial Intelligence, Not Managerially Owned Monitoring

Wherever AI is implemented in the context of the team, the team is to command the data and the dissemination of summary content.

2. The output of the AI will contain evidence as well as uncertainty.

Ranges and feedback regarding confidence are to be promoted, rather than a simple numeric certainty.

3. Decision logs will be open and public to everyone.

In any situation in which AI has supplied a recommendation for a decision, a log should be kept which lists: inputs, AI Recommendation, Team Member Decision and Reason.

4. Measure results rather than actions.

The AI will produce many metrics. If we are to increase the value we are creating through Agile Methods, we will require fewer and more definitive ways of measuring value.

5. Preserve miles for Psychological Safety.

If AI-generated reports are used for monitoring, teams will not feel safe providing honest input.

Recent research conducted by computer software developers confirms that AI Time Saving does not guarantee improved results if the team is in an atmosphere of disorganization, a lack of clarity within the organization, and lack of support from the organization – the very climate where Command and Control are prevalent.

Agile decisions that should stay human-led even when AI is available

Options generated through AI can be excellent; however, they are not as dependable when it comes to accepting responsibility for their ramifications.

It has been documented by both academics and practitioners that AI should never be permitted to determine the outcome of a decision without human intervention. Human and ethical aspects of decision making must always be taken into consideration.

When using Scrum or SAFe, allow people to make these choices, including:

• The prioritization of competing interests through trade-offs (i.e., which items will receive attention now versus later)
• The commitments made to achieve sprint goals and Program Increment objectives (PI)
• The determination of whether a requirement was signed-off as “done” by the team or stakeholders
• The identification of acceptable risk levels (i.e., how much uncertainty can be accepted)
• The ethical nature of the decision and its potential impact on customers (i.e., equitable treatment, protecting customer privacy, avoiding harm)
• Identifying the health of the organisation (i.e., addressing conflict, maintaining equitable workloads, providing coaching)
• The overall strategic focus of the organisation (i.e., what we are developing as an organisation, and its purpose)

In summary, if the outcome of a decision impacts people’s lives, influences the long-term strategic direction of an organisation, or publicly commits funds/time to the success of a specific initiative, AI can advise on options that should be included. Ultimately, a human must make these decisions.

What “human-in-the-loop” means for AI in Scrum/SAFe

When using the Agile framework, the term “human-in-the-loop” has different meanings; for example, you may assume that it means that you just looked at the information provided.  As such, “Human-in-the-loop” refers to an AI system that can suggest or propose solutions, but a responsible person must validate and modify the proposed solutions before they can be included as part of the plan.

The importance of having a human involved in the process is shown through various studies indicating that in some design scenarios, positive results can be achieved by using a combination of the intelligence of people and artificial intelligence; however, there may be serious disadvantages to having humans rely too strongly upon the output of an AI.

How to operationalize human-in-the-loop in Agile work

• Mark the output from the AI system clearly: “Draft generated by AI – pending review.”
• Attribute ownership: backlog text to PO, technical design to Devs.
• Must require a validation step: fast checks (below) prior to its usage in planning.
• Add “evidence links” : What data or artifacts suggested this inference?

Validating AI outputs in Agile planning without slowing teams down

The fear exists, and it has been articulated this way: “If we validate everything, we lose the speed benefit.” The way to solve the challenge is to validate proportionally based on impact and risk.

The 3-level validation strategy: fast, medium, high

Level 1 (Fast check – 30-90 seconds): of low-risk drafting

When used for: storytelling text, meeting write-ups, content for templates

Guides:

• “Is it consistent with our actual statements/actions?”
• Are there any evident hallucinations
• Are there sensitive pieces of information within?

Level 2 (Evidence check – a few minutes): For planning-impact comments

Use when AI suggests: scope, dependencies, estimates patterns, risk flags

Checks:

• What artifacts is this based on (tickets, repo, metrics)?
• Are these inputs contemporary?
• At least one member of the team must confirm to reality.

Level 3 (Test/verify – timeboxed): for high stakes outputs

When AI impacts these aspects: release readiness, compliance, significant PI commitments

Checks:

• Data from real-world sources (system of record) should be validated
• Execute small spike or BackTest
• Verify with multiple stakeholders

This maintains the velocity, thereby covering Scrum’s focus on transparency and empiricism.

Preventing “false certainty” from AI forecasts and dashboards

AI communication can be very assertive. The agile mindset is the reverse. Forecasts must be considered hypotheses.

It is here that the Scrum warning concerning “complex work” becomes critical: in a complex setting, what will happen is essentially unknowable; forecasted decisions are based on what has already occurred, and inspection & adaptation are paramount.

Practical techniques teams use to avoid false certainty

• Forces are range values, not point values: Request best-case / likely / worst-case scenarios for AI.
• Explicitly Request for Assumptions: “What assumptions must be true for this forecast?”
• Backtest AI recommendations: Forecast for this month compared with the actual data for last month
• Correlation, Not Causation: It can detect patterns, but it is not capable of proving why there is such a pattern.
• The task is not about concluding answers using AI but asking questions

Example: ‘What should we inspect in the Sprint Review based on this trend?’

• Keep Dashboards Secondary to Conversations

Decisions in Agile occur at occurrences, and not only at dashboards.

HBR has identified the “dataism” fallacy, trusting too much in data and algorithms that magically deliver truth. It is a best anti-pattern for agile leaders to directly ignore.

Real-world examples (scenario-based) of AI used well in Scrum & SAFe

The pragmatic takeaway: AI is strongest when it reduces friction, not when it increases control

The most evident trend among credible sources and expert recommendations is this:

• AI can accelerate production work (such as drafting, summarizing, option generation).
• But the largest risk lies in using AI for organizational purposes such as decision centralization, certainty, or punishment for transparency.

Scrum and SAFe already have a mechanism for dealing with uncertainty, namely ‘inspect and adapt.’ The role of AI is to improve it, not supplant it.