Insights

Architectures that work, by industry.

Reference designs for enterprise AI, grounded in Anthropic's published guidance. Each one shows the architecture, the human decision, and how to measure it.

Method

Method Workflow or agent: choosing the architecture for enterprise AI work 4 min Anthropic documents five workflow patterns and a high bar for agents. How to choose, and the context rules that keep either reliable. Method Write the eval before the system 4 min Define success and build the evaluation first. Criteria, graders, and an example spec for document extraction. Cost Routing across Haiku, Sonnet, and Opus: where the cost goes 3 min Current Claude prices, a worked example you can recompute, and what routing, batching, and caching change. Governance What zero data retention and HIPAA cover, feature by feature 3 min Which Claude API features zero data retention and HIPAA cover, and what that does to a regulated architecture.

By industry

Financial services Financial services: assembling the evidence packet for an AML alert 4 min An orchestrator gathers evidence for an alert in parallel, cites every fact, and leaves the decision to the analyst. Insurance Insurance: coverage review for incoming claims 3 min Map loss facts to policy clauses with citations, compute limits in code, and hand the adjuster a reviewable file. Healthcare Healthcare: assembling prior authorization packets 3 min Assemble the evidence, check it against payer criteria, list the gaps, and stay inside HIPAA-eligible features. Life sciences Life sciences: cross-document consistency checks for regulatory submissions 3 min Find inconsistencies across protocols, plans, and study reports before submission, each tied to a source location. Legal Legal: first-pass contract review against your playbook 3 min Check each clause against the playbook in parallel, cite the text, and leave the call to the attorney. Manufacturing Manufacturing: from a nonconformance report to a root-cause brief 3 min Gather batch, equipment, and procedure evidence for a deviation and rank hypotheses for the quality engineer. Retail and consumer Retail and consumer: reading all of your customer feedback 3 min Classify all feedback cheaply, detect issues the taxonomy does not know, and route them with evidence. Logistics Logistics: handling shipment exceptions, from alert to approved action 3 min Classify exceptions, cost the options in code, auto-approve the small ones, and route the rest to a planner. Data and analytics Data and analytics: agents that test your data end to end 3 min Triage data-quality failures into routed tickets, and run analytics test journeys from a plain-language instruction.