Log inNext →
Field reading14 min

What an Applied AI Architect Does

By the end of this lesson you'll be able to describe the applied AI architect's job from first call to production, explain the trusted-advisor posture in concrete behaviours, tell the role apart from Applied AI Engineer and Forward Deployed Engineer, and say what the interview loop is likely to test (and how little is publicly known about it).

Try itFrom first call to productionstage 1 of 7 · Architect leads
Discovery
You do

Find the business problem, the workflow, the data, the constraints and the decision process

You leave behind

A discovery summary with metrics and open questions

Also in the room

Account executive, business owner

Watch for · Ask for the data owner and security now, not later.

The architect leads from discovery through the proof of concept and the security review, then hands off to whoever builds production. Who leads each stage is the course's framing of the three applied roles, not an official org chart.

The job in one sentence

An applied AI architect helps a large customer decide whether, where and how to use a model like Claude, designs the solution, proves it works on the customer's own data, and hands it to the people who will build and run it. You write prototypes and evals, but most of your output is decisions, designs, numbers and conversations.

Your coach for this track is Grace Liu, a fictional principal architect. Every customer she hands you is fictional.

What the posting says

Anthropic has hired for this work under the title Solutions Architect, Applied AI. The posting describes (Official, as mirrored on job boards in 2026):

  • A pre-sales architect for large enterprises, and a trusted technical advisor who helps customers see what Claude can do and how it fits their technology stack.
  • Someone who builds evals, designs scalable architectures, and works with Sales, Product and Engineering to take customers from discovery to deployment.

On 2026-10-08 Anthropic's careers page listed these jobs under Applied AI as Applied AI Architect, split by segment and region (Official, anthropic.com/jobs). Read the current posting for your segment yourself: a Startups architect and a Public Sector architect spend their weeks very differently.

Reference: segments listed on the careers page, and what this course couldn't check

Segments listed on 2026-10-08: Industries, Commercial, Enterprise Tech, Startups, Digital Natives Business, Partnerships, Public Sector and Cyber, plus regional roles in Asia-Pacific and Europe and manager roles for verticals such as Financial Services and Healthcare & Life Sciences. The full job descriptions sit on a separate hiring site this course couldn't load, so this lesson doesn't quote years of experience, travel or pay.

From first call to production

The work follows the customer through seven stages: discovery, qualification, solution design, proof of concept, security and procurement, deployment handoff, and expansion. Your role changes at each one, but one thing doesn't: you make sure what gets sold can actually be built and will actually work. Step through the diagram above to see what you leave behind at each stage.

Two things are easy to miss. Qualification is part of your job: telling Sales early that a use case is a poor fit saves months. And the security review can easily become the longest stage, so start it during discovery (module A2), not after the proof of concept (A4).

Reference: the seven stages as a table
Stage What you do What you leave behind Who else is in the room
1. Discovery Find the business problem, the workflow, the data, the constraints and the decision process A discovery summary with metrics and open questions Account executive, customer business owner
2. Qualification Decide with Sales whether this is a good fit and worth a proof of concept A go/no-go and a scoped use case Account executive, sales leadership
3. Solution design Pick the approach, design the architecture, estimate cost and latency A solution design doc (lesson 04) Customer architects and engineers
4. Proof of concept Agree success criteria up front, build evals on real data, run the bake-off honestly Eval results and a recommendation Customer engineers, sometimes Applied AI engineers
5. Security and procurement Answer the security review, explain data handling, support legal Completed questionnaires, an architecture diagram for the CISO CISO's team, legal, procurement
6. Deployment handoff Hand the design and evals to whoever builds production: the customer, a partner, or a forward deployed team A handoff pack: design, evals, risks, open items Customer engineering, partners, FDEs
7. Expansion Check the metrics landed, find the next use case, feed field lessons back to Product A results review and the next opportunity Account team, Product

The trusted-advisor posture

"Trusted advisor" means the customer believes you'll tell them the truth even when the truth costs your company a deal. A vendor says "Claude can do that"; an advisor says "Claude can do the classification well. The final payout decision should stay with a person, and here's why." Three habits sit behind it:

  1. Quantify. Every claim has a number and a source: an eval result, a cost estimate with its arithmetic, a baseline the customer measured.
  2. Say what you don't know. "I'll confirm that with our security team by Thursday" builds more trust than a confident guess that fails in the security review.
  3. Recommend against yourself when it's right. If a rules engine solves 80% of the problem for a tenth of the cost, say so. Customers remember the architect who saved them money.
Reference: vendor lines and advisor lines
A vendor says An advisor says
"Claude can do that." "Claude can do the classification well. The final payout decision should stay with a person, and here's why."
"Let's start a pilot." "Before a pilot, we need a baseline. How long does this take today, and how often is it wrong?"
"Accuracy will be very high." "On your 300 test cases we got 91% right. Here are the 27 we got wrong and what they have in common."
"We support every requirement." "Two of your requirements conflict. Here are the options and what each one costs."
"Fine-tuning will fix it." "Let's try better prompts and retrieval first. They're cheaper to change, and we can measure whether they're enough."
A customer asks whether Claude can fully automate their claims payouts. What does the advisor's answer contain?

Three things: a specific yes for the part the model does well (triage, drafting, extraction), a clear "payouts stay with a person" for the high-risk decision with the reason, and a question that anchors everything in a baseline: "What's today's error rate?" A plain "yes, end to end" is the vendor answer; "let's pilot on live claims next week" skips the baseline.

Working with Sales, Product and Engineering

Sales. The account executive owns the relationship, the deal and the forecast. You own the technical win: the customer's technical team believes the solution will work and will say so. Tell your account executive bad news (a blocked data source, a sceptical CISO) the day you learn it.

Product. Field feedback counts only when it's specific: "Four financial-services customers blocked on X this quarter; here's each workaround and the revenue at stake."

Engineering and FDEs. Write the handoff so specialists don't repeat discovery: goals, constraints, evals, decisions and why, and what's open.

How it differs from the neighbouring roles

The difference is the output: the architect produces decisions, designs, eval plans and cost models across many accounts; the Applied AI Engineer builds eval frameworks and reference implementations; the FDE ships production code inside one or a few customers. (The course's framing from module C1, not an official org chart.)

Reference: the three applied roles side by side
Applied AI Architect Applied AI Engineer Forward Deployed Engineer
Main question Should we, and how, at what cost? How do we build it so it works? How do we ship it inside this customer's systems?
Typical stage Pre-sales through handoff Discovery through deployment, deeper on the build Deployment and production
Main outputs Designs, recommendations, eval plans, cost models Eval frameworks, prototypes, reference implementations Production code, integrations, MCP servers, agents
Code Prototypes, evals, scripts A lot Most of the week
Accounts at once Many Several One or a few
Who you persuade Executives, CISOs, customer architects Customer engineers Customer engineers and operators

Choose this track if you enjoy whiteboards, tradeoffs and rooms with executives more than long stretches of implementation. The FDE track (modules 01–10) and the Engineer track (E1–E7) cover building in depth.

How this shows up in interviews

Public data on architect loops is sparse: confirm the format with your recruiter.

  • Official (candidate AI guidance): write your own application answers; no AI in take-homes or live interviews unless told otherwise; preparing with Claude is encouraged.
  • Reported (module C5): a culture or values interview, as for every role at Anthropic.
  • Anecdotal: a recruiter screen, a technical screen on LLM application architecture, and a customer scenario (a case presentation or a live role-play).

Either way you'll think like an architect out loud: ask before designing, trade off with numbers, stay honest under pressure. A5 trains the pressure; the A7 mock loop puts it together.

An interviewer asks, "A customer wants to fine-tune a model on their support tickets. What do you say?" What's the first thing out of your mouth?

A question, not an answer: "What problem are they trying to fix, and what does it cost today?" Then the baseline (how good is the current process, how good is a plain prompt plus retrieval on 100 real tickets), and only then whether fine-tuning closes a gap you can measure. Answering "yes, here's how" first is the vendor reflex the interviewer is listening for.

Practise each prompt below in under three minutes, with one real example and one number.

PRACTICE PROMPTS (original)
1. "Walk me through taking a large insurer from a first meeting to Claude
   in production. Where does it usually stall?"
2. "Tell me about a time you told a customer, or your own sales team,
   that something wasn't a good fit."
3. "A customer wants to fine-tune a model on their support tickets.
   What do you ask before answering?"

ANSWER SHAPE
  Situation (1 line) -> what I asked or measured -> the number
  -> the decision and who it upset -> what happened next

Cheat card

  1. The architect decides whether, where and how to use the model, proves it on the customer's data, and hands off a design others can build.
  2. Seven stages: discovery, qualification, design, proof of concept, security and procurement, handoff, expansion. Qualification is yours, and start security in discovery.
  3. You own the technical win; the account executive owns the deal. The role differs from Engineer and FDE by output: decisions and designs, not production code.
Advisor answer = the part the model does well (with a number)
               + what stays with a person, and why
               + the baseline question: "What does it cost / how wrong is it today?"

Interview line "I'd rather lose a deal early by saying it's a poor fit than lose the customer later. So I quantify every claim, say what I don't know with a date to find out, and recommend the cheaper option when it's enough."