Engineering

AI Solutions Architect

Also hired as: GenAI Solutions Architect · AI Platform Architect · Pre-sales AI Architect

You design the AI systems enterprises actually ship — and talk them past the security review.

US salary · 2026
$170k – $320k+ OTE
Cloud-vendor architects earn base + bonus + RSUs; pre-sales flavors add variable comp. Consultancy bands run lower, hot AI-vendor bands higher. Estimates.
Typical background
Senior engineers or consultants with customer-facing miles — often existing cloud solutions architects adding AI depth. Rarely entry-level; five-plus years of experience is typical.
This roadmap
5 stages · 25 nodes
5 proof-of-work checkpoints
What is a AI Solutions Architect?

An AI solutions architect designs the AI systems enterprises actually deploy — and steers them past the places where those projects usually die. The job is decisions with evidence: build versus buy, RAG versus fine-tuning, which model, which cloud, what it costs at a million requests a day, and how it survives a security review. At cloud and AI vendors the role is customer-facing and attached to sales cycles; inside enterprises it is the platform architect deciding how a whole organization adopts AI without twenty teams reinventing the same stack.

The role matters now because enterprise AI moved from pilots to production budgets, and the failure pattern is well documented: promising pilots killed by compliance reviews, surprise inference bills, or architectures that could not scale past the demo. Vendors — AWS, Azure, Google Cloud, Databricks, NVIDIA, and the big consultancies — staff AI solutions architects specifically to prevent those deaths, because a deal that dies in procurement pays nobody. Someone who can design a compliant, cost-modeled system and explain it to a CISO is the difference between a pilot and a contract.

Distinct from its neighbors: a forward-deployed engineer embeds and builds hands-on with one customer; you design across many, and you own the trust conversation as much as the diagram. Versus an agent engineer, you decide the shape of systems more often than you implement them — though the credible architects have built the reference patterns with their own hands at least once. Breadth beats depth here, and communication is half the job: the best architecture that cannot be explained to a buying committee loses to a worse one that can.

What you'll actually do
  • Run discovery workshops that turn vague 'we need AI' mandates into concrete system requirements.
  • Produce reference architectures — RAG, agents, batch inference — with cost models attached.
  • Make build-vs-buy and RAG-vs-fine-tune calls, in writing, with the evidence that backs them.
  • Lead security and compliance conversations: data residency, PII handling, SOC 2, HIPAA.
  • Build proof-of-concepts sturdy enough to survive procurement, not just the demo call.
  • Advise on model selection and multi-model routing as pricing and capabilities shift.
  • Support sales cycles with demos, RFP responses, and honest objection handling.
  • Publish reference implementations and enablement material that scale you beyond one deal.
This role fits you if
  • You can explain hard technology to non-engineers without dumbing it down or showing off.
  • Breadth energizes you; going a year deep on one service does not.
  • Whiteboards, customers, and a bit of travel sound better than heads-down IC sprints.
  • You have been the engineer who 'talks to the client' — and liked it.
  • Trade-off analysis across cost, latency, quality, and risk is your comfort zone.
The toolbox
RAG architectureVector databasesAWS BedrockAzure AI FoundryGoogle Vertex AIDatabricksModel routing / gatewaysCost modelingPII & data governanceSOC 2 / HIPAA literacyTerraformArchitecture diagramsExecutive communication
The roadmap — 5 stages, 25 nodes
Stage 1

AI System Literacy

Weeks 1–4

You cannot architect what you do not understand. Get fluent in the patterns before you get fluent in the pitch.

LLM fundamentals for architectsCore
Context windows, token economics, latency profiles, and the failure modes that surface in production. This is the vocabulary every later decision is written in.
RAG — the default enterprise patternCore
Most enterprise AI is 'answer questions over our documents, safely.' Know the retrieval pipeline, where quality degrades, and what each stage costs.
When fine-tuning wins (and when it's a trap)Core
Fine-tune for behavior, format, and latency; retrieve for knowledge. Clients will ask for fine-tuning by name; you need the decision framework, not just the definition.
Agents 101Elective
Enough agentic literacy to scope 'can an agent do this?' honestly — autonomy levels, tool use, and where agents still fail in enterprises.
Checkpoint: three one-page architecture briefs✓ Checkpoint
A RAG support bot, a document-processing pipeline, and an agent workflow — each with a components diagram, a cost estimate, and a named top risk. Send them to two engineers for brutal feedback.
Stage 2

Build the Reference Patterns

Months 2–3

Architects who have built are believed; architects who have only diagrammed are tolerated. Build each pattern once, properly.

Build a RAG system yourselfCore
Chunking, embeddings, retrieval, reranking, generation — deployed and queryable. Every future security and cost conversation gets easier because you have touched each stage.
The model and hosting landscapeCore
Frontier APIs vs open weights; Bedrock vs Vertex vs Azure AI Foundry vs self-hosted. Build a comparison matrix you keep current — it becomes your most-reused asset.
Cost modelingCore
Tokens to dollars at 10k, 100k, and 1M requests per month, including caching, routing, and batch discounts. Surprise inference bills kill more projects than bad architecture does.
Fine-tune something smallElective
Run one LoRA job end to end so 'we could fine-tune here' comes from experience. An afternoon of GPU time buys years of credibility.
Checkpoint: publish a reference implementation✓ Checkpoint
A working RAG or agent system with an architecture diagram, a cost model at three traffic tiers, and deploy instructions someone else has successfully followed.
Stage 3

Security, Compliance, and Trust

Months 3–5

Enterprise deals die in the security review. The architect who can pass one is the architect who gets hired.

The AI threat surfaceCore
Injection, data exfiltration through tools, model supply chain. Learn the attack map well enough to draw it from memory in front of a security team.
PII and data governanceCore
Where customer data flows, what the model provider retains, redaction and residency options. 'Does the model train on our data?' is the first question every enterprise asks — have a precise answer.
Compliance mappingCore
SOC 2, HIPAA, and EU AI Act at working depth: which controls apply to an AI system and what evidence satisfies an auditor. Not a lawyer's depth — a review-passing depth.
Red teaming and governance frameworksElective
Know how organizations structure AI risk programs so your designs slot into their governance instead of fighting it.
Checkpoint: a security pack that unblocks a deal✓ Checkpoint
For your Stage 2 reference build: threat model, data-flow diagram, PII handling, and mitigations mapped to SOC 2-style controls — the document that gets a stalled enterprise deal moving.
Stage 4

The Architect Craft

Months 5–7

Discovery, decision frameworks, and communication — the consulting skills that turn technical breadth into closed deals.

Discovery that finds the real problemCore
Ask about data, users, volumes, and constraints before proposing anything. Half the value of a good architect is redirecting a bad idea in the first meeting.
Decision frameworks in writingCore
Build-vs-buy, RAG-vs-fine-tune, single-vs-multi-model — each as a written recommendation with the alternatives you rejected and why. Written trade-offs are the deliverable clients keep.
Communication artifactsCore
One diagram per audience: exec narrative, engineering detail, security data-flow. Practice the 20-minute architecture presentation until it feels like conversation.
Vendor certificationsElective
AWS Solutions Architect plus a GenAI specialty, or the Azure/GCP equivalents. Certs do not make you good, but they get vendor HR filters out of your way.
Checkpoint: two engagements end to end✓ Checkpoint
Real clients, OSS projects, or rigorous mocks: discovery notes, a written design recommendation with rejected alternatives, and a recorded 20-minute presentation for each.
Stage 5

Land the Role

Months 7–9

Four hiring tracks, each with a different interview. Pick deliberately and show up with a portfolio, not a promise.

Pick your trackCore
Cloud-vendor SA orgs, AI-vendor solutions teams, GSI AI practices, or internal platform architect. Vendor roles pay best; internal roles travel least; GSIs hire most.
The interview formatsCore
Whiteboard architecture with changing requirements, a customer roleplay where you handle pushback, and sometimes a POC take-home. Your reference builds and security pack are the prep.
Public presenceCore
Architecture teardowns and cost-model posts attract exactly the recruiters who staff these roles. One meetup talk on 'passing the AI security review' can carry an entire job search.
The FDE side doorElective
Forward-deployed engineering is the hands-on on-ramp: a year embedded with customers converts naturally into an architect role with better war stories.
Checkpoint: five applications, one panel✓ Checkpoint
Ship a portfolio site of your reference architectures, run five tailored applications across the four tracks, convert at least one to a panel round, and debrief every rejection in writing.
Build your portfolio

Nobody hires a AI Solutions Architect off a certificate. They hire off proof. Ship these and put them where people can click them:

01
Three one-page architecture briefs
A RAG support bot, a document-processing pipeline, and an agent workflow — each with a components diagram, a cost estimate, and a named top risk, hardened by brutal feedback from two engineers.
Proves: You can compress a system into a page a buying committee will actually read.
02
A deployed reference implementation
A working RAG or agent system with an architecture diagram, a cost model at three traffic tiers, and deploy instructions a stranger has successfully followed.
Proves: You architect from experience, not slideware — the test every customer engineer quietly runs on you.
03
A cost model and hosting matrix that stays current
Tokens-to-dollars at 10k, 100k, and 1M requests per month — caching, routing, and batch discounts included — plus a maintained comparison of Bedrock, Vertex, Azure AI Foundry, and self-hosting.
Proves: You can kill the surprise inference bill before it kills the project.
04
A security pack mapped to controls
For your reference build: threat model, data-flow diagram, PII handling, and mitigations mapped to SOC 2-style controls — the document that restarts a stalled enterprise deal.
Proves: You can pass the review where enterprise AI projects actually die.
05
Two engagements, end to end
Real clients, OSS projects, or rigorous mocks: discovery notes, a written build-vs-buy or RAG-vs-fine-tune recommendation with the alternatives you rejected and why, and a recorded 20-minute presentation for each.
Proves: You can run the full architect loop — discover, decide in writing, defend it to a room.
Position your profile
Headline formula

I design AI systems enterprises actually ship: [RAG / agent] reference architectures with cost models at [X]M requests/month and security packs that pass review.

Resume bullets to earn
  • Published [N] reference architectures with cost models at three traffic tiers; deploy guides followed successfully by engineers I've never met.
  • Wrote the build-vs-buy recommendation for [system]; the rejected-alternatives analysis held through procurement and the design shipped.
  • Produced the security pack — threat model, data-flow, SOC 2-mapped mitigations — that unblocked [deal or deployment] after [X] weeks stalled in review.
  • Cut projected inference spend [X]% at [Y] requests/month with routing, caching, and batch strategy at design time.
  • Delivered [N] discovery-to-recommendation engagements with recorded architecture presentations; [N] designs adopted.
Where to be visible
  • A portfolio site of reference architectures — diagrams, cost models, deploy guides. It is the artifact interviews open with, so build it early.
  • Architecture teardowns and cost-model posts on LinkedIn — the recruiters who staff SA orgs live there and search these exact terms.
  • One recorded talk — a meetup session on 'passing the AI security review' can carry an entire job search. Pin it everywhere.
  • Vendor community programs: AWS Community Builders, Microsoft MVP, Google Cloud champions — they feed directly into vendor SA hiring pipelines.
  • Publish and maintain your comparison matrices — models, hosting, vector stores. A matrix people cite makes you the reference, literally.
What interviews actually test
  • Whiteboard architecture with moving requirements — mid-design they add HIPAA or 10x the traffic. Practice re-planning out loud without losing the thread.
  • Customer roleplay: a skeptical CISO or a cost-anxious VP pushes back while the panel watches. Rehearse objection handling as deliberately as you rehearse diagrams.
  • Trade-off interrogation: RAG vs fine-tune, build vs buy, single vs multi-model — with costs, reasons, and the case where you'd flip. Your written recommendations are the prep.
  • Cost-model depth: 'what does this cost at a million requests a day' with follow-ups. Bring numbers you computed beforehand, not numbers you improvise.
  • A POC take-home or presentation round: your reference implementation and recorded 20-minute talks are direct rehearsal — reuse them shamelessly.
Who's hiring
AWSMicrosoft AzureGoogle CloudDatabricksNVIDIAAccentureDeloitteSnowflakeAnthropic / OpenAI solutions teams
Free courses that map to this role
Related roles
FAQ

What does an AI solutions architect do day to day?

A mix of customer conversations, design work, and selective building: discovery calls, architecture diagrams and cost models, security questionnaires, POC development, and internal enablement. At vendors, expect sales-cycle rhythm — demos, RFPs, and quarter-end pressure alongside the technical work.

Is this a pre-sales job or an engineering job?

Both flavors exist. Vendor SA roles sit beside sales with variable comp and customer travel; internal platform-architect roles are pure engineering leadership. Read the org the role reports into — sales means pre-sales, CTO means engineering — and pick the rhythm you actually want.

Do I need to code as a solutions architect?

Yes, though less than an IC. Credibility comes from working POCs and reference implementations you built yourself; architects who cannot build get quietly routed around by customer engineers. You will code in bursts — a demo this week, none the next.

Should I recommend RAG or fine-tuning?

Default heuristic: RAG for knowledge that changes, fine-tuning for behavior, format, or latency — and they combine. The honest professional answer is 'run an eval on your actual task before committing,' which is exactly the answer that builds client trust.

Can this be an entry-level job?

Rarely. The role trades on judgment and credibility, which need reps — typically five-plus years in engineering, consulting, or cloud architecture. Realistic on-ramps: forward-deployed engineering, vendor support engineering, or building AI systems as an IC and moving over internally.

Start building, today

Every stage above maps to free lessons on this site. No signup, no paywall — open the first course and ship your first checkpoint this week.

Browse the courses →