AGENCYSCRIPT
CoursesEnterpriseBlog
👑FoundersSign inJoin Waitlist
AGENCYSCRIPT

Governed Certification Framework

The operating system for AI-enabled agency building. Certify judgment under constraint. Standards over scale. Governance over shortcuts.

Stay informed

Governance updates, certification insights, and industry standards.

Products

  • Platform
  • Certification
  • Launch Program
  • Vault
  • The Book

Certification

  • Foundation (AS-F)
  • Operator (AS-O)
  • Architect (AS-A)
  • Principal (AS-P)

Resources

  • Blog
  • Verify Credential
  • Enterprise
  • Partners
  • Pricing

Company

  • About
  • Contact
  • Careers
  • Press
© 2026 Agency Script, Inc.·
Privacy PolicyTerms of ServiceCertification AgreementSecurity

Standards over scale. Judgment over volume. Governance over shortcuts.

On This Page

The PDATA Framework: An OverviewStage 1 — Problem DefinitionWhy this is where most projects failWhat to produce at this stageGate criteriaStage 2 — Data ReadinessThe most underestimated stageKey assessments to runOutputStage 3 — Algorithm SelectionNot a free choiceA practical selection heuristicOutputStage 4 — Training and ValidationWhere intuition most often misleadsThree non-negotiable practicesOutputStage 5 — Assessment and DeploymentPerformance metrics are not optionalDeployment is not the finish lineThe business case loopApplying PDATA to Different Project TypesFrequently Asked QuestionsWhat is a machine learning basics framework and why does it matter?Can non-technical professionals use a framework like PDATA?How does PDATA differ from CRISP-DM?How long does each stage of PDATA take?What if my project is using a pre-trained model or API, not training from scratch?How do I know when to abandon a project rather than iterate?Key Takeaways
Home/Blog/Stop Picking Algorithms Before You Audit the Data
General

Stop Picking Algorithms Before You Audit the Data

A

Agency Script Editorial

Editorial Team

·March 22, 2026·11 min read
machine learning basicsmachine learning basics frameworkmachine learning basics guideai fundamentals

Machine learning gets misapplied constantly — not because practitioners lack intelligence, but because they lack a repeatable structure for thinking about it. They jump to tools before defining problems, pick algorithms before auditing data, and declare failure before measuring the right things. The result is wasted budget, eroded trust in AI initiatives, and a lot of "we tried that, it didn't work" sitting in the organizational memory.

What experienced ML practitioners actually use — even if they've never written it down — is a mental framework: a sequence of stages, each with its own questions, failure modes, and decision criteria. Making that framework explicit is what separates a team that ships useful models from one that endlessly experiments. This article introduces the PDATA Framework (Problem → Data → Algorithm → Training → Assessment), a named, reusable model you can apply to any ML project from day one.

Whether you're building an in-house capability or advising clients on AI adoption, having a shared vocabulary for machine learning basics reduces miscommunication, surfaces risks early, and makes projects auditable. Each stage of the PDATA Framework has a clear entry condition, a set of outputs it must produce, and a gate that determines whether you move forward. Learn to use it and you stop treating ML as a black box — you treat it as an engineering discipline with knowable constraints.


The PDATA Framework: An Overview

PDATA is a five-stage model for structuring any supervised or unsupervised machine learning project. It is deliberately technology-agnostic: the framework applies whether you're fine-tuning a language model, building a churn predictor, or training a computer vision classifier.

The five stages are:

  • P — Problem Definition: What decision does the model need to support, and how will success be measured?
  • D — Data Readiness: Do you have sufficient, clean, labeled data to train a model that generalizes?
  • A — Algorithm Selection: Which model family fits the data structure, problem type, and operational constraints?
  • T — Training and Validation: How do you build, tune, and validate without overfitting or leaking information?
  • A — Assessment and Deployment: How do you measure real-world performance and maintain the model over time?

Each stage produces artifacts — documents, datasets, benchmarks — that become inputs to the next stage. Skipping a stage doesn't eliminate it; it just means you'll discover that stage's problems later and more expensively.


Stage 1 — Problem Definition

Why this is where most projects fail

The most common machine learning failure mode is building the right model for the wrong problem. Teams often arrive at the ML stage having already decided on a solution — "we want a recommendation engine" — without articulating the underlying decision it needs to improve. Problem Definition forces you to work backwards from business outcome to model specification.

What to produce at this stage

  • Outcome statement: A single sentence of the form "We want to predict X so that decision-maker Y can take action Z." If you can't write this sentence, you're not ready to build anything.
  • ML problem type: Classify your problem — binary classification, multi-class classification, regression, clustering, ranking, anomaly detection. This determines your evaluation metrics before you touch data.
  • Baseline definition: What does current performance look like without ML? A model that improves on a coin flip by 3% is not a success if a simple rule-based system already achieves 80% accuracy.
  • Success threshold: The minimum performance level that justifies deployment. This is a business number, not a model number.

Gate criteria

Do not proceed to Data Readiness until the outcome statement is signed off by both the technical lead and the business stakeholder. Disagreement at this stage is cheap. Disagreement at Stage 4 is catastrophic.


Stage 2 — Data Readiness

The most underestimated stage

Data Readiness is where professional ML projects spend 40–70% of total project time, and where amateur projects spend almost none. The gap between these two approaches predicts outcomes more reliably than algorithm choice does.

Key assessments to run

  • Volume: For most supervised classification problems, you need a minimum of several hundred labeled examples per class to begin seeing meaningful generalization. Complex tasks (image recognition, NLP) typically require thousands to millions. Be honest about what you have.
  • Label quality: If your labels were generated by humans, measure inter-annotator agreement. Labels that two humans agree on only 70% of the time put a ceiling on model accuracy — no amount of architecture choice overcomes noisy ground truth.
  • Feature availability at inference: A feature that exists in your training data but won't be available when the model runs in production is a data leak waiting to happen. Audit every column for this.
  • Class imbalance: A fraud detection dataset where 0.1% of records are fraud needs deliberate handling — oversampling, undersampling, or adjusted loss functions — not default training.

Output

A Data Readiness Report: a short document cataloguing volume, label quality score, known gaps, and a go/no-go recommendation. This is also the right moment to look at the best tools for machine learning basics — your tooling choices for data versioning, labeling, and preprocessing pipelines belong here, not in Stage 3.


Stage 3 — Algorithm Selection

Not a free choice

Algorithm selection is constrained by Stages 1 and 2. Your problem type narrows the candidate set; your data volume and structure narrow it further. The residual choice is about operational trade-offs: interpretability versus accuracy, training speed versus inference latency, on-premise constraints versus cloud APIs.

A practical selection heuristic

Start simple and earn complexity. A logistic regression or gradient boosted tree often outperforms deep neural networks on tabular data with fewer than 100,000 rows. Deep learning earns its computational cost when:

  • Data is unstructured (images, text, audio)
  • Volume is large enough to exploit capacity
  • The feature engineering cost of classical methods would be prohibitive

For most business analytics problems — churn, lead scoring, demand forecasting — tree-based ensembles (Random Forest, XGBoost, LightGBM) deliver strong results with far less tuning. For language tasks in enterprise settings, fine-tuned foundation models have become the practical default, but they come with inference cost and explainability trade-offs worth examining. See Machine Learning Basics: Trade-offs, Options, and How to Decide for a detailed breakdown of those decision points.

Output

An Algorithm Selection Brief: chosen model family, two alternatives considered and rejected, and the specific reasons for rejection tied to constraints from Stages 1 and 2.


Stage 4 — Training and Validation

Where intuition most often misleads

Training a model is the part people imagine when they picture machine learning. It is also the stage most vulnerable to invisible errors — errors that won't surface until the model is in production.

Three non-negotiable practices

Train/validation/test split discipline: Your test set must be held out completely until final evaluation. Using it for any hyperparameter decision contaminates it. A common split is 70/15/15, but for small datasets, k-fold cross-validation on the training set is preferable.

Preventing data leakage: Leakage occurs when information from the future (or from the test set) influences training. Classic examples include scaling your entire dataset before splitting, or including a target-correlated feature that wouldn't exist at prediction time. Leakage produces inflated training metrics that collapse on deployment — one of the most demoralizing failure modes in applied ML.

Hyperparameter tuning strategy: Random search outperforms grid search in most practical settings because it covers the hyperparameter space more efficiently with the same compute budget. Bayesian optimization (e.g., Optuna, Ray Tune) outperforms both when tuning time is a meaningful constraint.

Output

A trained model artifact, training curve logs, and a validation performance report. These outputs feed directly into Stage 5 and become the baseline against which future model versions are compared.


Stage 5 — Assessment and Deployment

Performance metrics are not optional

Choosing the wrong metric in Stage 1 becomes most expensive here. Accuracy is almost always the wrong primary metric for imbalanced problems. Precision and recall trade off against each other in ways that are business-specific — a false positive in a medical screening context costs differently than a false positive in an email spam filter. How to measure machine learning basics: metrics that matter covers this in depth and should be read before you finalize Stage 1 outputs.

Deployment is not the finish line

A model deployed without a monitoring strategy will degrade silently. Data drift — the statistical distribution of inputs shifting away from the training distribution — is normal over time. Product changes, seasonal patterns, and user behavior shifts all cause it. Build the following into your deployment plan:

  • Performance dashboards updated on a defined cadence (daily for high-stakes models, weekly for lower-stakes)
  • Drift detection: Track feature distributions in production against training baselines. Significant divergence is an alert condition, not an observation.
  • Retraining triggers: Define the performance threshold below which a retraining cycle is automatically initiated.

The business case loop

A deployed model that isn't connected to a business outcome measurement is a science project. Close the loop by tracking the downstream metric from your Stage 1 outcome statement. If the model was built to reduce customer churn, measure churn. The ROI of machine learning basics framework gives you the vocabulary to make that case internally and to clients.


Applying PDATA to Different Project Types

The framework scales across project complexity. For a small classification project (a few thousand rows, binary output), you can move through all five stages in two to four weeks. For a large-scale NLP or vision system, individual stages can take months and involve multiple teams.

The most useful adaptation is knowing which stage is the current bottleneck. Agencies advising clients often find that the bottleneck is Stage 1 (the client hasn't defined the problem) or Stage 2 (the data doesn't exist in usable form). Technical skill can't fix either of those problems — only structured process can. That is why PDATA starts with Problem and Data before any algorithm is discussed. It is also worth noting how machine learning basics trends in 2026 are shifting the balance: foundation models are compressing Stage 3 for many use cases, but they are making Stage 2 (data quality, prompt design, evaluation) more consequential, not less.


Frequently Asked Questions

What is a machine learning basics framework and why does it matter?

A machine learning basics framework is a structured, repeatable model for planning and executing ML projects. It matters because ML projects fail most often due to process errors — wrong problem definition, poor data quality, misaligned metrics — not because of algorithm choice. A framework makes those errors visible early.

Can non-technical professionals use a framework like PDATA?

Yes, deliberately so. Stages 1, 2, and 5 of PDATA require business judgment more than technical expertise — defining outcomes, assessing data readiness, and measuring real-world impact are all decision-making tasks. Technical staff own Stages 3 and 4, but the framework creates shared checkpoints where both groups must align.

How does PDATA differ from CRISP-DM?

CRISP-DM (Cross Industry Standard Process for Data Mining) is a broader methodology covering business understanding, data understanding, modeling, evaluation, and deployment. PDATA is more opinionated and tightly sequenced, with explicit gate criteria between stages and a stronger emphasis on preventing data leakage and defining baselines before touching algorithms.

How long does each stage of PDATA take?

It depends heavily on project scale and data maturity. In typical business ML projects, Stage 2 (Data Readiness) consumes the most time — often 40–70% of total project duration. Stage 1 is often underinvested and should take longer than most teams allow. Stages 3 and 4 are faster when Stages 1 and 2 are done well.

What if my project is using a pre-trained model or API, not training from scratch?

The framework still applies, with Stage 3 and parts of Stage 4 modified. Algorithm selection becomes model or API selection; training becomes fine-tuning, prompt engineering, or retrieval-augmented generation setup. Stages 1, 2, and 5 are unchanged — you still need a defined problem, clean data for evaluation, and deployment monitoring.

How do I know when to abandon a project rather than iterate?

Return to your Stage 1 baseline and success threshold. If you've completed multiple training iterations and remain below the minimum viable performance level with no clear path to improvement, it is usually a data problem or a problem definition problem — not an algorithm problem. Restarting from Stage 1 or Stage 2 is a legitimate outcome, not a failure.


Key Takeaways

  • The PDATA Framework (Problem → Data → Algorithm → Training → Assessment) provides a reusable structure for any ML project, regardless of tools or team size.
  • Most ML project failures originate in Stage 1 (unclear problem definition) or Stage 2 (insufficient or low-quality data) — not in algorithm choice.
  • Each stage must produce documented artifacts and meet gate criteria before the next stage begins; skipping stages defers cost, it does not eliminate it.
  • Data leakage, class imbalance, and metric misalignment are the three most common technical failure modes and all have known, preventable causes.
  • Non-technical stakeholders own Stage 1 and are essential to Stage 5; the framework creates a structure where business and technical teams share accountability.
  • A deployed model without monitoring is a liability. Drift detection and retraining triggers are not optional infrastructure — they are part of the definition of a complete project.

Search Articles

Categories

OperationsSalesDeliveryGovernance

Popular Tags

prompt engineeringai fundamentalsai toolsthe difference between AIMLagency operationsagency growthenterprise sales

Share Article

A

Agency Script Editorial

Editorial Team

The Agency Script editorial team delivers operational insights on AI delivery, certification, and governance for modern agency operators.

Related Articles

General

Prompt Quality Decides Whether AI Earns Its Keep

Prompt quality is the single biggest variable in whether AI delivers real work or expensive noise. The model matters, the platform matters — but the prompt you write determines whether you get a first

A
Agency Script Editorial
June 1, 2026·10 min read
General

Counting the Real Cost of Every Token You Send

Tokens and context windows sit at the intersection of AI capability and operational cost—yet most business cases treat them as technical footnotes. That's a mistake that costs real money. Every time y

A
Agency Script Editorial
June 1, 2026·10 min read
General

Rolling Out AI Hallucinations Across a Team

Most teams discover AI hallucinations the hard way — a confident-sounding wrong answer makes it into a client deliverable, a legal brief, or a published report. The damage isn't just to the output; it

A
Agency Script Editorial
June 1, 2026·11 min read

Ready to certify your AI capability?

Join the professionals building governed, repeatable AI delivery systems.

Explore Certification