Agile methodologies were designed for software development where requirements can be broken into user stories and progress is measured by working software delivered each sprint. AI projects break this model in specific ways: model performance is not predictable from the backlog, data quality issues surface mid-sprint, and “done” is a probability rather than a binary state. Agile AI Delivery adapts standard Agile practices to accommodate these differences while preserving the iterative, feedback-driven philosophy that makes Agile effective.

Where Standard Agile Breaks Down for AI

Estimation uncertainty - In software development, experienced teams estimate story points with reasonable accuracy. In AI projects, a task like “improve model accuracy by 5%” might take one sprint or ten. The relationship between effort and outcome is non-linear and often unpredictable.

Definition of done - A software feature is done when it passes tests and meets acceptance criteria. A model is “done” when it meets a performance threshold, but reaching that threshold is not guaranteed. Teams can do everything right and still not achieve the target accuracy.

Backlog stability - AI project backlogs shift frequently. An exploratory sprint may reveal that the planned approach is infeasible, requiring a pivot that invalidates the existing backlog. This is not scope creep; it is the nature of research-informed work.

Adapting Sprint Structure

Spike sprints - Dedicate the first 1-2 sprints of an AI project to pure exploration: data analysis, feasibility assessment, and baseline model creation. These sprints have learning objectives rather than deliverable objectives. The output is a go/no-go decision with evidence.

Dual-track sprints - Run data/model work and application/integration work in parallel tracks. The model track focuses on improving accuracy, while the application track builds the infrastructure that will serve the model. This prevents the common failure mode where months of model development produce a model with no deployment path.

Timeboxed experiments - Within a sprint, allocate specific timeboxes for experimental work (trying a new feature engineering approach, testing a different model architecture). If the experiment does not show promise within the timebox, abandon it and try the next approach. This prevents teams from spending an entire sprint on an unproductive path.

Metrics and Definition of Done

Replace binary done/not-done with threshold-based acceptance criteria. A model task is done when: accuracy exceeds X% on the holdout set, latency is below Y ms at the 95th percentile, and the results have been reviewed by a domain expert. Document these thresholds in the user story before the sprint begins.

Track both delivery metrics (velocity, burndown) and model metrics (accuracy, precision, recall, latency) on the sprint dashboard. The team should see both dimensions at every standup.

Sprint Ceremonies Adapted for AI

Standup - Add “blockers related to data” as an explicit category. Data access delays, quality issues, and labeling bottlenecks are the most common AI project blockers and need daily visibility.

Sprint review - Demo model outputs, not just UI features. Show stakeholders real examples of model predictions, including failures. This builds realistic expectations and surfaces domain knowledge that improves the model.

Retrospective - Include “what did we learn about the data or the problem” alongside process improvements. AI projects generate technical insights that should be captured and shared.

Team Composition

Effective Agile AI teams include data scientists (model development), data engineers (pipeline and infrastructure), software engineers (application integration), and a product owner who understands both the business problem and the constraints of ML. The product owner role is critical: they must be comfortable with probabilistic outcomes and able to translate model performance metrics into business value for stakeholders.

Common Anti-Patterns

Waterfall in Agile clothing - Planning all model work upfront, then executing sequentially. AI projects need genuine iteration: build, measure, learn, adjust.

Ignoring infrastructure - Spending all sprint capacity on model experimentation with no investment in deployment, monitoring, or data pipelines. By sprint 4, there should be a deployed model (even a simple one) that stakeholders can interact with.

Perfection before deployment - Waiting for 99% accuracy before deploying. Deploy a good-enough model early, monitor real-world performance, and iterate based on production data. The production feedback loop is the most valuable data source for improvement.

When to Use This Framework

Agile AI Delivery works best for projects with a defined business problem, access to relevant data, and a team that includes both ML and software engineering skills. For pure research projects with no near-term productionization, a lighter research management approach may be more appropriate.