Career roadmap
AI Solutions Architect
Decide what AI a business should build, how it should be built, and what it will cost to run.
Before you start AI Architect
- Architecture or senior engineering background
- Hands-on experience with LLM applications
- Ability to write and present design documents
Capability and use case selection
Most AI projects fail because the use case was wrong, not the technology.
The highest-value skill in this role: saying no to the wrong project early.
- Value, feasibility and risk scoring
- Where determinism is required
- Tolerance for error in the workflow
- Build, buy or do nothing
Calibrated expectations are what stakeholders are paying you for.
Ch — LLM Fundamentals- Current model capabilities honestly stated
- Reliability ceilings on open-ended tasks
- Where hallucination is disqualifying
- Capability change over time
A small catalogue of patterns covers most enterprise AI requirements.
- Retrieval-augmented question answering
- Extraction and document processing
- Classification and routing
- Agentic workflow automation
Architecture proposals that lack a cost model do not get approved.
- Cost modelling per transaction
- Value quantification
- Pilot versus full rollout economics
- Measuring realised benefit
The vendor landscape changes quarterly, and lock-in is a real risk.
- Vendor evaluation criteria
- Portability and abstraction layers
- Total cost of ownership
- Exit strategy
BuildAssess five candidate use cases for a business and produce a prioritised recommendation with reasoning.
Designing AI systems
Architecture that survives contact with real data and real users.
The most requested enterprise AI pattern, and the one most often built badly.
Ch — RAG- Ingestion, chunking and indexing design
- Hybrid search and reranking
- Permission-aware retrieval
- Freshness and reindexing strategy
AI programmes expose every existing data quality and governance weakness.
- Source data readiness assessment
- Access control inheritance
- PII handling in prompts and indexes
- Data residency constraints
AI features live inside existing systems, not beside them.
- API and event integration
- Human workflow integration
- Fallback to existing processes
- Legacy system constraints
Prompt, retrieve, fine-tune or train — with reasons and costs attached.
Ch — RAG vs Fine-Tuning- Decision framework for adaptation
- Hosted versus self-hosted models
- Open-weight model viability
- Model upgrade and deprecation planning
Latency, availability and cost budgets, set before building rather than discovered after.
- Latency budgets per component
- Availability and provider outage handling
- Throughput and rate limit planning
- Cost ceilings and controls
BuildA reference architecture for a retrieval system, with data flow, security and cost documented.
Evaluation and quality
Architects who cannot define quality cannot defend the system in production.
Acceptance criteria for a probabilistic system must be agreed up front.
Ch — Evaluation & Hallucination- Defining acceptable quality with the business
- Golden datasets and their maintenance
- Automated versus human evaluation
- Regression testing across model changes
What happens when the system is wrong, and who is accountable.
- Failure impact analysis
- Confidence thresholds and abstention
- Human review requirements
- Liability and accountability
Regulation is arriving, and enterprises need documented controls now.
- EU AI Act risk categories
- Model documentation and transparency
- Approval and review processes
- Inventory of AI systems
Prompt injection and data leakage are architecture concerns, not implementation details.
Ch — AI Security- Prompt injection at the architecture level
- Data leakage between tenants
- Output filtering and moderation
- Abuse and cost attacks
Quality drifts as models, data and users change.
- Quality sampling in production
- Cost and latency dashboards
- User feedback capture
- Incident response for AI failures
BuildAn evaluation framework for a described system, with acceptance criteria agreed with stakeholders.
Delivery and adoption
Most AI programmes stall between pilot and production. Architects unblock that.
A pilot that cannot fail teaches nothing and proves nothing.
- Scoping a meaningful pilot
- Success criteria before starting
- User selection and feedback loops
- Deciding to stop
The pilot-to-production gap is where most programmes die.
- Operational readiness requirements
- Support model for AI features
- Capacity and rate limit planning
- Phased rollout
Users who do not trust the system will not use it, regardless of accuracy.
- Setting user expectations
- Training and documentation
- Transparency about limitations
- Handling resistance
The second and third use case should be cheaper than the first.
- Shared components and reuse
- Central versus federated delivery
- Model access governance
- Internal enablement
The daily deliverable, and often the interview artefact.
- Architecture decision records
- Data flow and trust boundary diagrams
- Cost models
- Writing for executives and engineers
BuildTake a pilot to a production rollout plan with staged adoption and success metrics.
Interview preparation
Interviews are design conversations with heavy emphasis on judgement and cost.
Design an enterprise AI system live, including what you would not build.
- Requirements and constraints first
- Component choices with justification
- Cost and latency estimation aloud
- Failure and fallback design
Common in consultancies: a business brief and a presented recommendation.
- Assessing use case viability
- Phased delivery proposal
- Risk identification
- Presenting to a non-technical panel
Architects who cannot go deep lose credibility quickly.
- Retrieval quality debugging
- Fine-tuning versus retrieval reasoning
- Token economics arithmetic
- Evaluation methodology
Increasingly present, especially in regulated industries.
- Regulatory classification of a system
- Documentation requirements
- Handling a model deprecation
- Auditability of AI decisions
Saying no to an executive's favourite AI idea is part of the job.
- Talking a stakeholder out of a use case
- A project you recommended stopping
- Managing inflated expectations
- Handling a failed pilot
BuildTwo reference architectures with cost models and evaluation plans, written up publicly.
AI Architect tools on your CV
- Claude / OpenAI APIs
- Vector databases
- Cloud AI platforms
- Evaluation frameworks
- C4 / architecture tooling
- Cost modelling
What AI Architect employers ask to see
- Two published reference architectures with cost models
- A use case assessment framework applied to real candidates
- An evaluation plan with agreed acceptance criteria
- A pilot-to-production rollout you led
Consultancies, cloud partners and enterprises running AI programmes. Named repeatedly in demand surveys as organisations move from experiments to production systems.
Content last reviewed 2026-08-31. Guidance only — no institute or paid placement is endorsed anywhere in this book.