Path for architects
The diagram is the easy part now
Four stages from an accurate picture of LLMs to a design you can defend and govern.
- 4 stages
- 125 hours
- 17 weeks
Your pace, hours a week
4 stages, 125 hours
Stage 1, an accurate model of LLMs
Know what you are drawing
This week
Explain tokens and embeddings to a colleague in two minutes.
More free options
LLM module, Machine Learning Crash Course
Google, 45 minutes
No price or sign-in shown. 45 minutes.
Stanford CME295 Transformers & LLMs, Autumn 2026, lectures 1 to 3
Afshine and Shervine Amidi, Stanford Online, YouTube
Free on YouTube. Lectures 1 to 3, about 5 hours. Graduate level, so expect some maths.
Paid options
Ranked by fit, ratings and freshness. We earn nothing from these links.
You end up with
A half-page explainer you would give to a sponsor.
Answers these problems
More about this path
Where this gets you
Can write a reviewable LLM system design with failure handling, cost per request and an exit plan, and map it to NIST and OWASP.
Just lost your job?
If you have just lost your job, do this first
Start at stage 2. Take the 45-minute Google module, read the Fowler patterns, and write a one-page decision record comparing retrieval and fine-tuning for a real problem you have met, with how you would test it and what it would cost. About 10 to 15 hours. Spend nothing on exams in the first month.
Governance frameworks to know
- NIST AI Risk Management Framework and Playbook
NIST, United States
A voluntary method for managing AI risk through four functions: Govern, Map, Measure and Manage.
Version 1.0, January 2023, now under revision. Voluntary. The common vocabulary for AI risk worldwide.
- NIST AI 600-1, Generative AI Profile
NIST, United States
The companion to the framework above: the risks particular to generative AI, with suggested actions for each.
Version 1.0, July 2024. Voluntary. It says little about agents.
- OWASP Top 10 for LLM Applications 2026
OWASP GenAI Security Project
The ten most serious security risks in applications built on large language models, each with attack scenarios and fixes.
2026 edition, 3 August 2026. Voluntary and free. The usual checklist for a security review.
- OWASP Top 10 for Agentic Applications
OWASP GenAI Security Project
The ten most serious risks when an AI system can plan and act through tools.
First edition, 9 December 2025. Voluntary and free. Use it beside the LLM list whenever a model can take actions.
- MITRE ATLAS
MITRE
A catalogue of real attacker techniques against AI systems, laid out like MITRE ATT&CK.
A living knowledge base. Voluntary. For threat modelling and red teaming.
- ISO/IEC 42001, AI management system
ISO and IEC
The standard an organisation can be audited and certified against for how it governs AI, as ISO/IEC 27001 works for security.
Edition 1, December 2023. Voluntary and certifiable. The text costs CHF 225.
- EU AI Act
European Union
Law that bans some AI uses, puts duties on high-risk systems and general-purpose models, and requires people to be told when they deal with AI.
In force since August 2024, amended July 2026. Binding. It applies to anyone whose AI system is sold or used in the EU, wherever the team sits.
The dates moved in July 2026. High-risk duties now apply from 2 December 2027 for uses such as hiring and education, and from 2 August 2028 for AI inside regulated products. Systems already on the market must mark AI-made content by 2 December 2026.
- CSA AI Controls Matrix
Cloud Security Alliance
A spreadsheet of concrete controls for AI systems, each mapped to ISO/IEC 42001, the EU AI Act and the NIST framework.
Version 1.1, June 2026. Voluntary and free. Useful as the crosswalk between the frameworks on this list.
- Model AI Governance Framework for Agentic AI
IMDA, Singapore
Practical guidance on deploying AI agents: limit what they can touch, keep a named person accountable, set approval points and test before release.
Version 1.5, June 2026. Guidance, not law. The clearest free source on rules for agents and sign-off.
What you need
Your existing architecture experience. Neural network basics help for stage 1.
Worth the fee, or not
| Certificate | Fee | Valid | Verdict |
|---|---|---|---|
| AWS Solutions Architect Professional | $300 | 3 years | Defensible only if the target employer uses AWS. |
| Google Professional Cloud Architect | $200, renewal $100 | 2 years | Defensible for Google shops. |
| Microsoft AZ-305 | Varies by country, not shown | Free renewal | Needs the Azure Administrator Associate first. |
| TOGAF | $405 each, $610 combined | Not stated for new versions | Large process-heavy firms only. Says little about AI. No hiring evidence found. |
| AIGP | $649 or $799 | 2-year term | The one AI governance certificate an architect can take with no entry bar. Worth it if you are moving toward governance or face the EU AI Act. No hiring evidence found. |
| ISO/IEC 42001 Lead Implementer | Bundled with a 5-day course, price not shown | Not shown | Only if your employer or clients are going for ISO/IEC 42001 and will pay. The full title also needs experience you must already have. |
| CompTIA SecAI+ | Not shown | Not shown | A security exam with a governance section, not a governance certificate. For security architects only. |
Awareness-level AI badges are not worth paying for if you have years in IT. Before paying, count the name in 10 to 20 postings you want.
What the data says
Nothing we could open ties architect job losses to AI. Claims that ivory-tower roles are being cut have no named data behind them.
UK ad shares from IT Jobs Watch, 21 Nov 2026. Moderate for ad share. Weak as a claim of absence.
An earlier study found code complexity rose after one AI tool was adopted. The only direct measure of architecture found, across 151 open-source Java projects, saw barely any change in six months. Erosion is not proven.
Larsen and Moghaddam, citing He and others, 11 Jun 2026. Moderate: open-source Java only, over six months.
Loosely coupled teams see gains from AI, and tightly coupled ones see little or none.
DORA 2025, Google Cloud, 24 Sep 2025. Moderate, correlational. Run by Google.
What this path cannot do
- A decision log or a risk map may not protect a job.
- Finishing this path builds skill and gives you things to talk about. It cannot promise work.
Checked 21 November 2026.
