Enterprise AI consulting succeeds through a structured 7-stage framework—Discovery, Readiness, Roadmap, Proof of Concept (PoC), Production Build, Deployment, and Monitoring—combining CRISP-DM, Agile, and MLOps methodologies.
Ask ten AI consultants how an engagement works and you will hear ten versions of "we start with discovery." What follows discovery is where firms differ, and where buyers get hurt. A vague process hides unpriced work, unclear handoffs, and decisions nobody owns. A defined process makes cost, timeline, and risk visible before you sign.
This guide explains how AI consulting works from first conversation to production. It compares the main methodologies, walks through seven stages with their activities, deliverables, and decision gates, and shows what a realistic twelve-week path looks like. It is written for business sponsors who need to know what they are buying and for technical leads who need to know what will be asked of their teams.
The stakes justify the rigour. IBM's 2025 CEO study found that only 25% of AI initiatives have delivered their expected return and only 16% have scaled enterprise-wide. McKinsey's State of AI survey reports that fewer than 10% of organisations are scaling AI agents in any single function. In most failed projects the model was not the problem. The process around it was.
Why a Defined AI Consulting Process Matters
AI projects differ from conventional software projects in three ways that a good process must handle.
They are experimental: Model performance depends on data quality that is often only discovered after work begins. A rigid plan written before that discovery will be wrong.
They depend on data, not just code: Gartner has predicted that through 2026 organisations will abandon 60% of AI projects that are not supported by AI-ready data, and its survey found that 63% of organisations lack confidence in their data management practices for AI. Process must therefore test data early.
They change how people work: A technically successful system that staff route around has delivered nothing. Adoption has to be planned, not hoped for.
A defined process turns those three realities into stages, gates, and artefacts that sponsors can govern.
The Main Methodologies Behind AI Consulting
Most firms blend several approaches. Understanding them helps you read a proposal.
Methodology | Core idea | Strength | Limitation |
|---|---|---|---|
CRISP-DM (Cross-Industry Standard Process for Data Mining) | Cycle through business understanding, data understanding, data preparation, modelling, evaluation, deployment | Proven structure for analytics and machine learning projects | Light on governance, integration, and operations |
Agile and product-led delivery | Short iterations, working increments, continuous feedback | Handles uncertainty well, keeps users involved | Needs strong product ownership, can drift without a roadmap |
Stage-gate | Fixed phases with formal approval gates | Controls funding and risk, suits large enterprises | Can feel slow and rigid if gates are poorly designed |
MLOps lifecycle | Automate training, deployment, monitoring, and retraining | Keeps models reliable after launch | Covers operations, not strategy |
Design thinking | Start from user needs and workflows | Improves adoption and fit | Needs technical validation to be buildable |
A strong AI consulting methodology combines them: CRISP-DM thinking for data and modelling, agile iterations for build, stage-gates for funding and risk, MLOps for operations, and user-centred design for adoption.
The Seven Stages of an AI Consulting Engagement
The names vary between firms. The substance does not.
Stage 1: Discovery and Business Alignment
Purpose: agree what success means before any technology is chosen.
Typical duration: one to two weeks.
Key activities
Stakeholder interviews across business, IT, risk, and the teams that will use the system
Review of strategy, current processes, and pain points
Definition of target business metrics and their current baselines
Identification of constraints: budget, regulation, timeline, existing systems
Deliverables
Business objectives and success metrics document
Stakeholder map and decision-maker list
Preliminary risk and constraint register
Decision gate: sponsors confirm objectives and a named business owner.
Stage 2: AI Readiness Assessment
Purpose: establish whether the organisation can support the use cases being considered.
Typical duration: one to two weeks, often overlapping with discovery.
Key activities
Data audit: sources, quality, completeness, accessibility, lineage, and labelling
Infrastructure review: cloud, compute, integration capabilities, security posture
Skills and operating model review
Governance and compliance review: data protection, sector regulation, model risk
Deliverables
Scored AI readiness report
Data readiness findings with remediation steps
Gap-closure plan with effort estimates
Decision gate: proceed, proceed with remediation, or stop.
Stage 3: Use Case Prioritisation and Roadmap Design
Purpose: choose what to build first, and in what order.
Typical duration: one to two weeks.
Key activities
Longlist of use cases scored on value, feasibility, data readiness, risk, and strategic fit
Business case for the top candidates, including cost and expected benefit
Sequencing into phases, with the first pilot scoped tightly
Resourcing plan covering build, buy, and internal versus partner roles
Governance design, including risk tiers and approval steps
Deliverables
Prioritised use case backlog
Phased roadmap with milestones, owners, and budgets
Governance framework
Decision gate: sponsors approve the first pilot and its budget.
Stage 4: Proof of Concept
Purpose: test feasibility on real data before committing to production.
Typical duration: four to eight weeks.
Key activities
Data pipeline for the pilot scope
Model development or configuration, using short iterations
Evaluation against pre-agreed accuracy and business thresholds
Early user testing with the people who will use the system
Cost modelling for production, including run costs
Deliverables
Working prototype validated on your data
Evaluation report with performance, limitations, and failure modes
Updated production cost and timeline estimate
Decision gate: go, iterate, or no-go, based on thresholds set in Stage 1 and Stage 3. A well-run process treats a no-go as a useful result, not a failure.
Stage 5: Production Build and Integration
Purpose: turn the validated prototype into a reliable, secure, integrated system.
Typical duration: eight to twelve weeks.
Key activities
Production-grade data pipelines and feature engineering
Integration with core systems such as ERP, CRM, EMR, or core banking
Security engineering: access control, encryption, logging, consent management
Testing: functional, performance, security, and user acceptance
Documentation and operational runbooks
Deliverables
Production-ready system integrated into your environment
Security and compliance evidence pack
Test results and sign-off records
Decision gate: release approval from business, IT, security, and compliance.
Stage 6: Deployment and Change Management
Purpose: put the system into daily use and make sure people adopt it.
Typical duration: two to four weeks, with adoption tracking continuing afterwards.
Key activities
Phased rollout, often starting with a single team or site
Training, super-user programmes, and support channels
Workflow redesign so the AI is part of the process, not an add-on
Adoption and quality tracking against baselines
Deliverables
Training materials and rollout plan
Adoption dashboard
Early benefits report against baseline metrics
McKinsey's research identifies workflow redesign as one of the clearest differences between organisations that capture AI value and those that do not. This stage is where that redesign happens.
Stage 7: Monitoring, Optimisation, and Scale
Purpose: keep the system accurate and expand what works.
Typical duration: ongoing, with a structured three to six month optimisation window after launch.
Key activities
Monitoring for data drift, concept drift, performance, cost, and incidents
Scheduled retraining and version control with rollback capability
Periodic governance reviews and audits
Backlog review for the next use cases on the roadmap
Transfer of ownership to internal teams, if agreed
Deliverables
Monitoring dashboards and alert thresholds
Retraining plan and change log
Scale-up plan for additional teams, sites, or use cases
Deliverables at a Glance
Stage | Principal deliverables |
|---|---|
1. Discovery | Objectives and metrics, stakeholder map, risk register |
2. Readiness | Scored readiness report, data findings, gap-closure plan |
3. Prioritisation and roadmap | Use case backlog, phased roadmap, governance framework |
4. Proof of concept | Validated prototype, evaluation report, production estimate |
5. Production build | Integrated system, security and compliance evidence |
6. Deployment | Training, rollout plan, adoption dashboard |
7. Monitor and scale | Monitoring, retraining plan, scale-up plan |
What a Twelve-Week Path Looks Like
For a single well-scoped use case, a common structure is:
Weeks | Focus | Output |
|---|---|---|
1 to 2 | Discovery and readiness | Readiness report, agreed metrics |
3 to 4 | Prioritisation and roadmap | Backlog, first pilot scope |
5 to 9 | Proof of concept and build | Validated working system |
10 to 12 | Deployment and early monitoring | System live with adoption tracking |
Full production hardening, multi-site rollout, and scaling to further use cases usually continue for another three to six months. The point of the twelve-week path is that the first working system, and the first measurable result, arrive early enough to build confidence and funding for what follows.
Roles and Responsibilities
Clear ownership prevents most delivery friction.
Role | Responsibility |
|---|---|
Executive sponsor | Funding, priorities, removing blockers |
Business owner | Defines value, owns outcomes and adoption |
Consulting lead | Methodology, quality, stakeholder management |
Data and ML engineers | Pipelines, models, integration |
Security and compliance lead | Controls, approvals, audit evidence |
IT and platform owners | Environments, access, system changes |
End-user champions | Requirements, testing, peer training |
How to Judge a Consultant's Process
Ask these questions in the sales conversation:
What are your stages, and what artefact does each one produce?
Where are the decision gates, and who decides at each?
What happens if the proof of concept fails its thresholds?
How early do you test our data?
Who handles integration, and is it priced?
How is change management delivered, and how is adoption measured?
What do you do after go-live, and for how long?
What do we own at the end?
Red Flags in an AI Consulting Process
No gates. The engagement proceeds to build regardless of what discovery finds.
No data work before commitment. Data surprises appear in month three.
Integration absent from the plan. It will arrive later as a change request.
Change management as a footnote.
No post-launch phase. Drift and support are left to you.
Deliverables described as "insights." Nothing physical to inspect.
How Prognos Labs Runs the Process
Prognos Labs uses a four-part structure that maps onto the seven stages above: Discover (discovery and readiness), Design (prioritisation and roadmap), Build (proof of concept and production build), and Scale (deployment, monitoring, and optimisation). Every stage produces a named output, every gate has an owner, and the same team stays accountable from the first workshop to production tuning. Our twelve-week framework is designed so the first working system lands inside the first quarter, with governance and compliance designed in from week one.
Related Guides
What Is Enterprise AI Consulting? A Complete Guide to Strategy and Execution
AI Readiness Framework: Evaluating Data, Infrastructure, and Culture
AI ROI Framework: Measuring the Business Value of Artificial Intelligence.