Skip to content
Ameen Abozaid
Englishالعربية
ARC-045

AI solution architecture

  • 6 Lessons
  • 4 h practice (estimate)
Start module
On this page 9 topics

Overview

Topics

  • Reference architectures for assistants
  • RAG
  • and agents over enterprise data
  • deployment options
  • cost and latency

Learn first

4 h of practice (estimate)

Learning outcomes

  1. Select a reference architecture for an AI capability: assistant, retrieval over enterprise content, or agent with tools.
  2. Integrate AI with enterprise systems (ERP, CRM, knowledge bases) through governed APIs and events rather than direct database access.
  3. Choose deployment options (hosted API, private endpoint, self-hosted) against data, latency, and cost requirements.
  4. Estimate cost and latency of an AI capability and design fallbacks to human or non-AI paths.

Lessons

  1. 01AI reference architectures
  2. 02Grounding AI in enterprise data
  3. 03Agents and tools over enterprise systems
  4. 04Deployment options and their trade-offs
  5. 05Demo: architecture of the Larkspur customer-service assistant
  6. 06Demo: costing the assistant per conversation

Lab and exercises

Lab

ARC-045-LAB1 · Larkspur AI assistant

Outcomes: 1–4 · Time: ~4 h

  1. Write the use cases and choose the reference architecture for each.
  2. Draw the C4 container view and the data-flow diagram, with the model reaching ERP and CRM only through governed APIs.
  3. List the tools with read or write type and approval rules.
  4. Choose the deployment option and justify it against data, latency, and cost.
  5. Estimate cost and latency per conversation with stated assumptions.
  6. Design the human handover and the fallback when the AI path is down.
  7. Defend the architecture to role-played customer-service and security leads.

Acceptance: C4 container view, data flows, deployment choice justified, cost per conversation estimated, human handover designed.

Exercises (case-based; six exercises)

IDOutcomeTask
ARC-045-EX011Choose assistant, retrieval, or agent for: store opening hours, returns-policy answers, "where is my parcel", "refund my order". Justify.
ARC-045-EX022A developer proposes giving the model a read-only database account on the ERP. Write the objection and the alternative.
ARC-045-EX033Choose hosted API, private endpoint, or self-hosted for Larkspur given customer names and order data in prompts and a three-person platform team.
ARC-045-EX044Estimate cost for 40,000 conversations per month, 6 turns each, with stated token assumptions and a placeholder price.
ARC-045-EX054Define five handover triggers and what the human agent receives.
ARC-045-EX062, 3Review a design with no audit of tool calls and a single provider with no fallback; give three corrections.
Show solutions
  • EX01 Opening hours: retrieval or static page; returns policy: retrieval; parcel: agent with read tool; refund: agent proposes, human or customer-confirmed workflow executes, no direct refund tool.
  • EX02 Bypasses access rules, audit, and schema isolation, and exposes more data than needed; use a user-scoped order API returning only required fields.
  • EX03 Private endpoint or hosted API with approved terms, since a small team cannot run models; keep identifiers minimal; verify provider terms on retention and training use.
  • EX04 Per turn: input tokens (prompt + retrieved text + history) plus output tokens; sum across six turns, with history growing; multiply by price per token (placeholder) and by 40,000; add retrieval, hosting, and monitoring; show cost per conversation, not per call.
  • EX05 Low confidence, customer asks for human, refund or complaint, repeated failure, detected distress; agent gets transcript, summary, identified customer, and actions already taken.
  • EX06 Log each tool call with user, arguments, and result summary; add timeout and fallback to the human queue; add a second provider or a non-AI path.

Outcome 4 on latency is also covered by lab step 5.

Workplace evidence: an AI solution architecture you designed or reviewed.

Quiz

Outcomes: O1 reference architectures · O2 integrating with enterprise systems · O3 deployment options · O4 cost, latency, and fallbacks.

A self-check for you: nothing is graded, saved or sent anywhere. Choose an answer, then check it. You can try again as often as you like.

  1. 1A customer asks questions answered by the current returns policy documents. The best fit is:O1
  2. 2What is the main added risk when moving from retrieval to an agent with tools?O1
  3. 3How should the assistant read order data?O2
  4. 4Who runs a tool the model requests?O2
  5. 5For a write action such as a refund, the safest design is:O2
  6. 6Which deployment option places a provider model inside your cloud network boundary?O3
  7. 7Self-hosting mainly shifts to you:O3
  8. 8The cost of the assistant should be estimated:O4
  9. 9When the model provider is down, a good fallback is:O4
  10. 10Which is a common mistake listed for AI solution design?O1

Practical part

Architecture defense for the Larkspur assistant: C4 container view, data flows, deployment choice, cost per conversation, and human handover. Pass: quiz at least 8/10, at least five of six exercises correct, and a defense in which no path lets the model reach core systems directly.

Common mistakes

Letting the model query core systems directly; no human handover; costs estimated per request without conversation volume.

Assessment

Architecture defense.

Workplace evidence

An AI solution architecture you designed or reviewed.

Career relevance

AI is now a standard part of enterprise solution designs.

Type to search levels, phases, modules and lessons.