Bayanihan Scrum
Collaboration-first, human-led, AI-assisted Agile for Filipino software teams
Joe Palala · Working paper · 2026
Abstract
Scrum assumes people will speak up, disagree and admit mistakes. Many Filipino teams find that costly. This paper proposes replacing status standups with collaboration sessions, and adding AI as an assistive tool without making it a gate. AI reads, analyzes and flags. People ask the questions, check the design and approve the spec. Collaboration stays human.
1. Problem
The Scrum Guide assumes team members are comfortable speaking in public, challenging estimates and questioning authority. In organizations with high power distance [4][9], silence is often read as agreement, QA softens feedback, and juniors defer to seniors. Feedback arrives late and quality suffers.
Adding AI raises a second problem. If one person must approve every AI output, the team has swapped one bottleneck for another.
2. Practices for team culture
Building on Lokalisadong Scrum [1], the daily standup becomes a collaboration session: where does someone need help, which assumptions should be challenged, which decisions deserve a second opinion. Hiya is redefined as shame over hiding problems, not over making mistakes. The principles below follow from this and from psychological safety research [3]. Remote teams have to cultivate this on purpose: power structures still exist at a distance, and they need to align so everyone works together (see [8], [9]).
Problems are gifts
Finding a defect helps the team.
Disagreement is respect
Silence is not respect.
Failure is data
Hidden mistakes are the real failure.
Ask early
A question today is cheaper than a bug next week.
Challenge ideas, never people
Critique requirements, designs and code.
Improve continuously
Do not wait for the retrospective.
Everyone teaches
No single point of failure.
Safety is everyone's job
Not only the Scrum Master's.
3. AI as an assistive tool
A note on terms. In classic machine learning, “human-in-the-loop” means people labeling and correcting data to train a model. This paper means something different, which it calls collaboration-first: people collaborate and decide first, and AI assists them. The test is whether the team still talks, questions and owns the outcome.
Three rules keep people in charge without slowing them down. First, there is no central approver: each role validates the work it owns. Second, review scales with risk, not volume. Formatting and documentation need none; architecture, security and releases need a human decision. Third, approval gates become short group sessions, so Scrum Masters and the whole team take part.
4. Structure
Work moves through three plain-text artifacts, following the requirements, design and tasks flow of Kiro [5], with roles borrowed from the BMAD Method [6]. Every role has a human owner and an AI assistant, and the human is accountable.
| Phase | Artifact | AI assists | Humans decide |
|---|---|---|---|
| Discover | requirements.md | Reads docs, flags gaps | Team questions assumptions |
| Design | design.md | Pattern pre-check | Senior developer decides; team challenges |
| Plan | tasks.md | Spots missing criteria | Scrum Master owns the stories |
| Build | code, tests | Suggests code and tests | Developers review; QA challenges |
5. The Scrum Master's role
AI does not remove the Scrum Master; it sharpens the job. The Scrum Master owns story preparation, facilitates the session at each phase boundary, limits work in progress so review queues do not pile up, and watches for silence, such as no one questioning an AI suggestion this week.
6. Limitations
This is a proposal. It has not been tested across teams, and there are no metrics yet. Risk tiers and role boundaries will need tuning in practice.
Appendix A. Supporting practices
Short descriptions of delivery practices that keep AI-assisted work traceable, tested and reversible instead of vibe-coded.
- GitOps.
- Git is the single source of truth. Every change, from a person or an AI, arrives as a reviewed pull request, so history shows who decided what and a bad change can be reverted.
- DevSecOps.
- Security checks run automatically with every change: dependency, secret and static analysis scans, alongside the tests.
- Continuous integration and delivery (CI/CD).
- Each change is built, tested and packaged by an automated pipeline, and released in small, frequent steps.
- Automated testing.
- Unit, integration and end-to-end tests, often written first (TDD) or from acceptance criteria (BDD), define what correct means independently of who wrote the code.
- Code review.
- A teammate reads each change before it merges, checking design, readability and risk. It is also how knowledge spreads.
- Infrastructure as code.
- Servers, networks and environments are described in versioned files, so they can be reviewed, reproduced and rebuilt.
- Observability.
- Logs, metrics and traces show what a running system is doing, so problems are found and explained quickly.
- Feature flags.
- New behavior ships switched off and is enabled gradually, so a release can be reversed without a redeploy.
- Architecture decision records.
- Short notes record what was decided and why, so later readers, human or AI, do not have to guess.
Protected branches and required approvers can also enforce the risk tiers in section 3.
Appendix B. AI, jobs and the economy
This appendix is the author's opinion. It is not supported by the research cited above.
- Companies can work alongside AI and keep their people. Staying competitive does not require eliminating humans; it requires pairing them with good tools.
- The pressure on jobs comes mainly from the underlying economic state, not from AI taking over. AI is often the stated reason rather than the cause.
- The current mood expects AI development to end in a singularity, as in scenarios like AI 2027 [10]. This paper is not about models or timelines. It is about keeping governance in human hands: who decides, who approves, and who is accountable. Those practices matter for the AI that exists today, which is useful, fallible, and in need of human judgment.
References
- Lokalisadong Scrum. LinkedIn post. linkedin.com/posts/joel-alcarez-89331321_lokalis
- Schwaber, K. & Sutherland, J. The Scrum Guide. scrumguides.org
- Edmondson, A. (1999). Psychological safety and learning behavior in work teams. Administrative Science Quarterly, 44(2), 350–383. doi.org/10.2307/2666999
- Hofstede, G. (1980). Culture's Consequences. Sage.
- AWS. Kiro: spec-driven development. kiro.dev
- Madison, B. The BMAD Method. github.com/bmad-code-org/BMAD-METHOD
- Böckeler, B. Understanding Spec-Driven-Development: Kiro, spec-kit, and Tessl. martinfowler.com. martinfowler.com/articles/exploring-gen-ai/sdd-3
- Fried, J. & Heinemeier Hansson, D. (2013). Remote: Office Not Required. Crown Business. en.wikipedia.org/wiki/37signals
- Meyer, E. (2014). The Culture Map: Breaking Through the Invisible Boundaries of Global Business. PublicAffairs. en.wikipedia.org/wiki/Erin_Meyer
- Kokotajlo, D. et al. AI 2027. ai-2027.com
Full human-in-the-loop document (the collaboration-first workflow) · Source on GitHub