AnthropicClaude Certified Architect – Professional2026-101.1 Translate business problems into Claude-based AI solutions
Assign Claude the right workflow jobs
Completed by the knowledge check
Learning objective
Translate workflows and pain points into the decisions and outputs Claude should own, justifying AI against deterministic automation and non-AI alternatives.
Core idea
An architect converts a request such as “automate complaints with Claude” into explicit allocations of work. Decide which output needs language understanding, which decision follows an exact rule, and which consequence needs a person. The unit of allocation is a job within the workflow, not the department or the entire process.
Anthropic’s ticket-routing guidance starts by investigating how people route work, what automated rules exist, and where they fail. Its agent-design guidance recommends the simplest solution that works. Together these give you a defensible starting point: identify the bottleneck before selecting a model, an agent, or a framework. An AI budget does not make every job an AI problem.
The allocation should explain both the benefit and the authority Claude receives. Producing a suggested complaint category differs from deciding whether a customer qualifies for compensation. A fluent explanation does not grant authority over the payment system.
How it works
Locate the pain in the actual workflow
Walk through completed cases with the people who handle them. Record the input, output, current owner, elapsed time, active handling time, and exception for each stage. Distinguish a queue caused by missing information from time spent reading information already available. Summarization can address the latter; it cannot supply an absent source record.
Use the questions in Anthropic’s ticket-routing guide to expose the existing decision structure: how service levels affect priority, which specialists receive cases, how automated rules behave, and what happens to ambiguous inputs. Your artifact is a workflow allocation table that associates each proposed Claude job with a specific observed failure of the existing approach.
Do not infer automation value from case volume alone. A high-volume exact calculation can belong in code. A lower-volume task with lengthy, multilingual correspondence can warrant Claude if reading that correspondence dominates handling time. Compare the work removed with the verification work introduced.
Compare alternatives for each job
Keep a functioning system of record as the owner of the rules it already implements. A policy lookup or arithmetic calculation does not become more valuable because Claude repeats it. Where a missing form field causes most delays, changing the form can remove the problem before inference is needed.
Anthropic’s ticket-routing guide identifies conditions that can favor Claude over traditional machine learning: limited labeled examples, changing categories, unstructured text, semantic rules, ambiguity, and multilingual inputs. Treat these as reasons to test suitability, not promises of sufficient accuracy. An existing trained classifier with satisfactory performance remains a candidate baseline.
Use this allocation to compare alternatives:
| Job | Candidate owner | Evidence that would justify the choice |
|---|---|---|
| Enforce a fixed entitlement rule | Existing service or deterministic code | The rule is explicit and the service already returns the authoritative result |
| Remove repeated manual data entry | Form or integration change | The required value already exists in another trusted system |
| Score a stable category | Existing classifier | Labeled history and measured performance meet the need |
| Interpret correspondence into a category | Claude | Semantic ambiguity defeats existing rules on representative cases |
| Approve an expensive exception | Human decision-maker | The cost of an incorrect action exceeds the benefit of automatic approval |
The last row is a business authority decision in your allocation, not a claim that Claude cannot generate an approval recommendation. State whether the person checks every case or only defined exceptions, and include that workload in the comparison.
Specify Claude’s owned output
Name the output as extraction, classification, summarization, or generation. Extraction produces fields; classification selects from defined categories; summarization prepares evidence for a named reader; generation produces a draft. Each output needs its own test. “Handle the complaint” hides several distinct contracts.
For a classification job, define the categories before prompting. Anthropic’s guide emphasizes clear user-intent categories and also accounts for urgency, customer type, service level, and language. Keep an explanation short and tied to supplied evidence so a reviewer can check it. An explanation is useful review material, not independent proof that the label is correct.
Define the downstream authority separately. For example, Claude proposes a category and extracts an order identifier; the order service verifies the identifier and returns entitlement; an employee approves compensation. This separates meaning interpretation from the authoritative business decision without asking the model to repeat exact rules.
If every stage is known in advance, a predefined workflow can orchestrate those stages. Anthropic distinguishes such workflows from agents that dynamically direct their own process and tool use. Pattern selection belongs to 1.3; here you establish whether the job even needs that flexibility before allocating an agent’s authority.
Key distinctions and trade-offs
Language difficulty versus process difficulty: A complex approval process can still consist of explicit rules. Conversely, a short message can require substantial semantic interpretation. Select Claude for the latter capability, rather than the apparent complexity of the organizational chart.
Recommendation versus execution: An output can be useful without owning the final action. A draft or suggested category creates a review cost, but preserves the existing decision owner while suitability is tested.
Automation versus process improvement: Better intake forms, queue rules, integrations, and existing software are alternatives to inference. Compare them against the same bottleneck. If a form improvement eliminates the delay, a model call has no remaining value to defend.
Single call versus agent: Agentic systems can trade latency and cost for better performance. Those costs need a task-specific benefit. A fixed reading-and-classification job is not enough evidence to require autonomous planning.
Architecture / decision scenario
A retail returns team of eight must acknowledge requests within one working day. Staff spend twelve minutes reading long email threads, but only one minute applying return-policy rules. The sponsor asks for an agent that issues refunds automatically. These timings and constraints are illustrative business assumptions.
You split the workflow into thread interpretation, order identification, entitlement calculation, refund approval, and acknowledgment drafting. A form improvement captures the order identifier for web requests. The order service retains date and entitlement checks. An existing classifier remains the baseline for simple email categories. Claude is tested on disputed, multilingual threads where that classifier sends too many cases to a general queue.
Claude owns a summary, a proposed intent category, and an acknowledgment draft. It does not own the refund decision. The employee approves the draft and any payment; the workflow records that decision before the payment service acts. A summary that saves six minutes but requires seven minutes of correction fails the business justification.
A second request asks Claude to identify fabricated product photographs. Anthropic’s vision documentation states that Claude cannot determine whether an image is AI-generated and should not be relied on to detect fake images. You exclude that job from Claude’s allocation and retain the established fraud-review process. The accepted scope is a language-processing workflow, supported by existing systems and people, with a measurable handling-time hypothesis for 1.1.4.
Common exam traps
- Treating the sponsor’s request for an agent as evidence that autonomous planning is necessary.
- Replacing an authoritative policy service with a model’s interpretation of the same rule.
- Comparing Claude only with manual work while ignoring a form, integration, or existing classifier.
- Treating a generated explanation as proof of a correct decision.
- Assigning Claude an action when the validated use case supports only a recommendation.
- Assuming image understanding includes reliable detection of synthetic images.
What to remember
Allocate jobs before selecting technology. Tie each Claude-owned output to a bottleneck requiring semantic interpretation, compare it with non-AI alternatives, and retain explicit owners for rules and consequential actions. The resulting scope should explain what Claude produces, what it cannot establish, and why that allocation is worth testing.
Knowledge check
An insurer asks Claude to “speed up claims intake.” Walking completed cases shows that most delay comes from claimants leaving the policy-number field blank, although every claimant submits through a logged-in portal whose account record already holds that number. Staff read the short claim descriptions in under a minute. Which change removes the observed bottleneck with the least added verification work?
A telecom support team routes email with an existing trained classifier whose measured accuracy meets the target on simple English billing and outage emails. Disputed threads in several languages are mostly misrouted to a general queue; the team has only a few dozen labeled disputes, and labeling more would take months, but dispute routing must improve this quarter. How should you allocate the categorization job?
A home-insurance workflow already uses Claude to summarize claimant correspondence. The fraud lead now wants Claude to flag uploaded damage photographs that might be AI-generated, so that the established fraud review can be skipped for unflagged claims. What allocation should you make?
References
- Anthropic, Ticket routing
- Anthropic, Building effective agents
- Anthropic, Vision