- The automation contract checklist most vendors never bring up An automation contract checklist covering credentials, documentation, the support window and IP, so handover doesn't quietly become a six-month retainer.
- From pilot to production: the gap is ownership, not technology Most pilots fail the pilot to production jump on missing owners and no rollback plan, not on model quality. Here is what actually closes the gap.
- Automation pitch red flags you should never ignore Automation pitch red flags are easier to catch than they look: no auditable number, no error handling, no exit plan, and "AI-powered" standing in for a feature.
- Build buy or commission: the decision that actually matters The build buy or commission question has a real answer: it depends on team size, budget, and how core the process is, not on comfort.
- Twelve automation vendor questions to ask before you buy Twelve automation vendor questions that separate a real system from a demo. Ask them before you buy, and expect a straight answer each time.
- Agent observability: if it runs unattended, it must log Agent observability isn't optional once no one is watching the screen. What to log, why silent failures are the default, and what counts as enough.
- Keeping token spend predictable with caps and a downgrade ladder Keeping token spend predictable means budgets per task, hard caps, caching, and a model-downgrade ladder, not hoping usage stays flat.
- AI agent error handling: retry, repair, degrade, escalate AI agent error handling needs a fixed order of fallbacks: retry, repair, degrade, escalate. Here is what each stage actually does and when to stop.
- Idempotent cron jobs: the scheduling trigger most teams forget Idempotent cron jobs stop missed runs and overlapping executions from corrupting agent workflows. What to build before you trust a schedule.
- Structured output contract: treat the schema as an interface A structured output contract is the real interface between a model and your system. Schema design decisions here outlast every prompt you write.
- Prompting for reliability, not cleverness Prompting for reliability means boring, explicit prompts that name failure modes instead of chasing a clever persona or tone.
- Building an agent workflow: a build log A week-by-week log of building an agent workflow for a real queue: what worked, what broke, and the mistake that cost a day of cleanup.
- Running a content operation with agents Agents can draft, research and format content at scale, but judgment and the final review gate still belong to a person who owns the outcome.
- Measuring quality, not just speed A faster AI workflow that produces wrong output is worse than a slow one that doesn't. Here's how to measure quality so speed stops lying to you.
- Automating a broken process makes it break faster Automation speeds up whatever you feed it, including dysfunction. Fix the process before you script it, or you industrialise the problem.
- The maintenance bill nobody budgets for Automations don't stay finished. Models change, APIs deprecate, schemas drift. Budget 20-30% of build cost per year to keep things running.
- What to Automate First: Frequency, Cost, and Stability The loudest complaint is rarely the right target. A frequency times cost times stability score tells you what to automate first, and what to leave alone.
- Time saved is not money saved Automation reports hours saved, but hours only become money through one of three conversions. Most projects never make that step explicit.
- The honest maths of automation ROI A worked example of automation payback that includes the maintenance hours, the exception rate, and the review time vendors leave out of the pitch deck.
- You cannot ship an agent you cannot measure Evaluation belongs before launch, not after. What a minimum agent eval set looks like, and why teams that skip it ship blind.
- Multi-agent systems: when more agents make it worse Adding agents adds coordination cost before it adds capability. Most working systems are one agent, good tools, and a queue.
- "Memory" is the wrong word for what agents need Agent "memory" is marketing. What agents actually need is state management and retrieval, and conflating the three produces brittle systems.
- Designing the human handoff in agentic workflows Most human-in-the-loop checkpoints are theater. Here's where a person must stay in an agentic workflow, and where the checkpoint is pure overhead.
- Agent or Workflow? A Ten-Minute Test Most products sold as AI agents are fixed pipelines with a model bolted on. Run this ten-minute test before you buy or build one.
- The four ways agent pilots die in production Most agent pilots don't fail loudly. They stall on unattributed errors, missing owners, and cost overruns nobody tracked from day one.
- What "agentic" actually means, and where the word does no work Most products sold as agentic are deterministic pipelines with a model inside. A ten-minute test for telling the two apart, and why it changes your budget.
- The five workflows I rebuild for almost every client Not the interesting automations — the boring ones. These five appear in nearly every engagement because they address the same handful of failure points.