The Svolta blog. Written from inside the work.
Focused answers on workflow measurement, automation, AI agents, integration, evaluation, governance, and operating production systems.
Focused questions. Production answers.
Each article answers one buyer or operator question and links back to the relevant guide, service, and proof. Technical depth stays when it helps the decision.
Evals before agents: the order that saves the project
Write the eval set on your own data before you pick the framework, the model, or the vector database. Skip this and the project burns a quarter relitigating quality questions you could have answered in a week.
Mac Sweeny · 4 June 2026 · 8 min readProcess automation consultant: what to expect before a build
A process automation consultant should leave you with a mapped workflow, a defensible baseline, named owners, integration and access boundaries, acceptance evidence, and a clear build, buy, redesign, or stop decision.
Mac Sweeny · 19 Aug 2026 · 9 min readAI agent implementation checklist for production
A production AI agent needs a bounded job, accountable owner, controlled data and tool access, test evidence, human stops, monitoring, and a rollback path. Use this checklist before increasing its autonomy.
Mac Sweeny · 18 Aug 2026 · 9 min readChatGPT, Grok, Claude, Gemini or Copilot? A practical review for service businesses
A use-case review of five general AI assistants for Australian service businesses, with fit labels for research, customer communication, proposals, spreadsheets, SEO and advertising preparation.
Mac Sweeny · 18 Aug 2026 · 12 min readHow we turn inboxes into actionables
Svolta pulls the inboxes together. One agent identifies what the email is. A second fills the object schema that starts the workflow. One client-reported desk: over 1,000 emails a day, 2–3 hours a day back across departments.
Mac Sweeny · 14 Aug 2026 · 7 min readAI agent infrastructure: the layer that makes agents safe to run
Agents fail in production for infrastructure reasons, not model reasons: no grounding, no evals, no audit trail, no data layer. The five components underneath every agent system that survives contact with a real queue.
Mac Sweeny · 4 July 2026Custom AI agent development services: what 'custom' should mean
Every vendor sells 'custom' AI agents. The word should mean your workflow, your systems, your data, your evals, and your ownership at handover. A checklist for reading proposals, and the parts that should never be custom.
Mac Sweeny · 4 July 2026Agentic workflows that survive production
A demo agent and a production agent share a vocabulary. They do not share an infrastructure. Caching, deterministic fallbacks, observability, eval gates on every change. The boring discipline most teams skip.
Mac Sweeny · 21 May 2026 · 7 min readProcess mapping is the first AI artifact
Map the process before you write a prompt. The map is where you find the work an agent can actually do, and the implicit context that will kill the project if it stays hidden until integration.
Mac Sweeny · 3 Apr 2026 · 6 min readBook a free consultation.
Bring one workflow, its weekly volume, and the systems it touches. A free 30-minute fit call tells you whether Workflow Scoping is the right next step.