Pick the business problems you want to solve. Each one carries its key question, its technology, its team and its effort — so scope comes from business decisions, not from a list of tools.
Structure: Process → Activity → Use case → Key Business Question (KBQ) → Metric. Each case is scored 1-10 on difficulty and impact.
The structure (Process → Activity → Use case → KBQ → Metric) is DataRev's own methodology. The individual cases and their per-role day counts are authored here; they do not come from DataRev's delivery history. The base score too — but the adjustments by industry and by kind of work come from dated, published studies, listed at the foot of this page.
Test any workflow for agentic readiness and, separately, for the regimes that apply to it based on where it operates and what data it touches.
🤖
1. Agentic workflow feasibility
Readiness model50/100
Semi-Autonomous Agent (Human-in-the-Loop)
The workflow is best deployed as a semi-autonomous assistant that executes tasks and draft states, requiring human approval for final execution.
Technical prerequisites
• Partial REST APIs require adapter wrappers for agent integration.
• Human approval queue integrated prior to state mutations.
⚖️
2. Applicable regulatory regimes
Select every one that applies. Each brings its own regime.
This decides which privacy laws engage, regardless of how risky the AI itself is.
Autonomous decisioning
System acts without human override review
Regimes in scope1
1 regime applies to this workflow across 1 jurisdiction.
LFPDPPP
Mexico
Baseline obligations
Mexican Federal Law on Protection of Personal Data Held by Private Parties (2025)
Processing personal data in Mexico triggers the general regime whether or not AI is involved.
Obligations
✓Privacy notice (aviso de privacidad) available at the point of collection, with purposes itemised.
✓Documented consent basis for each stated purpose.
✓Procedure and contact point for ARCO rights (access, rectification, cancellation, objection).
✓Administrative, physical and technical security measures proportional to risk.
✓Third-party transfer rules, including model providers outside Mexico.
✓Breach notification where the incident materially affects data subjects' rights.
Timing: In force since 21 March 2025, when the new law fully replaced the 2010 statute.
Supervisory authority: Secretaría Anticorrupción y Buen Gobierno — took over the function after INAI was dissolved.
This tool is an indicative diagnostic, not legal advice. Actual obligations turn on facts a questionnaire cannot capture. Validate any decision with your legal counsel before acting.
Catalogue3 selected
Finance
Customer lifecycle
Operations and supply chain
Product and service lifecycle
Employee lifecycle (Hire to Retire)
Team required
What the selection demands, turned into people over a delivery window.
months
Data product manager1 person
Holds scope and sequencing. Without it the backlog is set by whoever shouts loudest.
Architect / tech lead1 person · 5 person-days
Designs storage, access and system boundaries. Triggered when where the data lives is still an open question.
Owns: Object storage and open tables · Query engine / lakehouse · Catalog and lineage — control plane
Data engineer1 person · 21 person-days
Builds the pipes: ingestion, integration and the loads that run unattended.
Owns: Ingestion and change data capture · Object storage and open tables · Query engine / lakehouse · Orchestration · Data quality and observability · Feature store
Analytics engineer1 person
Turns raw tables into tested, trustworthy models. The fastest-growing seat on modern data teams.
Owns: Transformation and modelling · Semantic and metrics layer
BI analyst1 person · 22 person-days
Turns the model into an answer the business uses on Monday.
Owns: Semantic and metrics layer · BI and visualization
Data scientist1 person
Models to extract patterns. Triggered when the case is predictive or prescriptive, not descriptive.
Owns: Training and experimentation
ML engineer1 person · 15 person-days
Takes the model into a production system with committed latency and availability.
Owns: Feature store · Serving and inference
MLOps / platform engineer1 person
Keeps the model alive after launch: deployment, monitoring, drift and rollback.
Owns: Serving and inference · Model monitoring and drift
Governance and quality lead1 person
Catalog, lineage, quality and access policy. Cisco 2025: only 51% have their data centralized.
Owns: Catalog and lineage — control plane · Data quality and observability · Governance, access and privacy
Change management and adoption1 person
Making people actually use it. Cisco 2025: only a third have a formal change-management plan.
Add a profile your organization needs that was not inferred:
Rounded up: half an engineer does not show up on Monday. A plan that implies half-people is a plan that slips.
Technology involved
14 layers
The platform capabilities the selection demands. If a case needs streaming or a vector store, the architecture has to plan for it from the start.
Data plane
Ingestion and change data capture
Move data out of source systems without breaking them or asking someone for a weekly extract.
FivetranAirbyteDebeziumAWS DMS
Built by: Data engineer
Object storage and open tables
The floor. MAD 2025 describes the shift from “warehouse vs lakehouse” to object storage with open table formats.
S3 / GCS / ADLSApache IcebergDelta Lake
Built by: Architect / tech lead · Data engineer
Query engine / lakehouse
Where queries run. Drives a large share of the platform's monthly cost.
BigQuerySnowflakeDatabricks SQLRedshiftClickHouse
Built by: Architect / tech lead · Data engineer
Transformation and modelling
Turn raw tables into tables a human can trust. This is where the analytics engineer lives.
dbtSQLMeshDataform
Built by: Analytics engineer
Orchestration
What runs, in what order, and what happens when something fails at 3am.
AirflowDagsterPrefect
Built by: Data engineer
Catalog and lineage — control plane
MAD 2025 places the neutral catalog as the stack's control plane: what exists, where it came from and who may see it.
Unity CatalogApache PolarisDataHubCollibra
Built by: Governance and quality lead · Architect / tech lead
Data quality and observability
Catch a broken table before the executive catches it in their dashboard.
Monte CarloSYNQGreat ExpectationsSoda
Built by: Governance and quality lead · Data engineer
Governance, access and privacy
Who can see what, on what legal basis, and how it is proven to an auditor.
ImmutaPrivacerapolíticas nativas del catálogo
Built by: Governance and quality lead
Analytics plane
Semantic and metrics layer
So “revenue” means the same thing in finance, in sales, and in the agent answering questions.
dbt Semantic LayerCubeLookML
Built by: Analytics engineer · BI analyst
BI and visualization
The surface the business consumes. Usually the most expensive licence line too.
Power BILookerTableauMetabase
Built by: BI analyst
Machine learning plane
Feature store
So the model sees in production exactly the features it was trained on.
FeastTectonVertex Feature Store
Built by: ML engineer · Data engineer
Training and experimentation
Train, compare, and be able to reproduce the model that got approved.
MLflowWeights & BiasesSageMakerVertex AI
Built by: Data scientist
Serving and inference
Publish the model as a service with committed latency and availability.
SageMaker EndpointsVertex EndpointsBentoMLKServe
Built by: MLOps / platform engineer · ML engineer
Model monitoring and drift
A model degrades silently. Without this layer nobody finds out until the business hurts.
EvidentlyArizeFiddler
Built by: MLOps / platform engineer
Add a layer your architecture needs:
Structure follows the 2025 MAD Landscape (FirstMark), the eleventh edition of the data and AI ecosystem map. Vendors illustrate each layer; they are neither recommendations nor an exhaustive list.
Process and responsibility mapping
Which activities your selection triggers, and who answers for each. Only actors that actually take part appear.
RResponsible (does)
AAccountable (answers for)
CConsulted
IInformed
Activity
Architect / tech lead (DataRev)
Data engineer (DataRev)
BI analyst (DataRev)
ML engineer (DataRev)
Process owner (client)
Data steward (client)
IT / Security (client)
End user (client)
1 · FramingDefine key business questions (KBQ) and success criteriaDeliverable: Prioritised KBQs with metric and target
R
·
C
·
A
·
·
C
Data architecture designDeliverable: Architecture diagram and decision records
A
R
·
·
I
·
C
·
2 · FoundationsAccess, security and complianceDeliverable: Roles, permissions and signed NDAs
R
I
·
·
·
C
A
·
Integrate sources and automate ingestionDeliverable: Pipelines running with monitoring
C
A
·
·
I
C
·
·
Data quality and shared definitionsDeliverable: Catalogue, lineage and quality rules
·
R
C
·
C
A
·
·
3 · BuildSemantic layer and metric modelDeliverable: Certified metrics, one version
·
R
A
·
C
C
·
·
Build dashboards and reportsDeliverable: Dashboards validated against the KBQ
·
·
A
·
I
·
·
C
Train and validate modelsDeliverable: Model with documented performance metrics
·
R
·
A
C
C
·
·
4 · AdoptionUser acceptance testingDeliverable: Sign-off from the process owner
I
·
C
·
A
·
·
R
Training and change managementDeliverable: Team trained and material handed over
·
·
R
·
A
·
·
C
Handover and ongoing operationDeliverable: Runbook, alerting and a named owner
C
R
·
·
·
C
A
·
Exactly one Accountable (A) per activity — the rule that makes RACI useful. Framing, acceptance and training decisions belong to the client, not DataRev: a project stalls when the process owner is merely informed of calls only they can make.
Why the plan carries governance, security and adoption
These are not filler lines. They are the gaps Cisco measured across 8,039 organizations, and the reason a technically sound project still fails to deliver value.
Organizations fully ready for AI (“Pacesetters”)13%
Have their in-house data fully centralized51%
Consider their infrastructure adaptable and scalable for AI34%
Have a process to measure the impact of their AI investments32%
Feel fully capable of securing agentic AI systems31%
Where these numbers come from
Every score adjustment comes from dated, published work. None of it is DataRev's own delivery history — that data does not exist yet — and nothing published before 2025 is used: the 2022-era maturity indices and the “87% of models never reach production” line predate generative AI in production and would misprice any agentic case.
8,039 leaders responsible for AI in organizations with 500+ employees, 30 markets, 26 industries. Fieldwork August 2025, analysed by an independent third party. Significant to ±1% at 95%.
Eleventh edition of the ecosystem map, ~1,150 vendors. Defines the data → infrastructure → ML/AI → agents/applications flow and adds an explicit agent layer for the first time.
Technical debt accounts for 21% to 40% of an organization's IT spend.
The per-role day counts are DataRev assumptions calibrated to industry norms, not measurements. Real scope is sized by measuring, not estimating — this exists to structure the conversation.
What now?
Let's review your result together and decide which use case to tackle first. The session is free and commits you to nothing.
