Long-form on shipping AI.
Cornerstone articles on AI agents, automation, RAG patterns, voice AI, and the production reality behind the demos.
The AI Development playbook: how we ship agents in 6 weeks
We ship production AI agents in 6 weeks by being opinionated about tools, refusing to skip discovery, building evals from day one, and treating code agents as a force multiplier. This is the playbook — what we use, what we refuse, and why it lands consistently.
Microsoft 365 + AI: what's actually useful in 2026
Microsoft 365 AI in 2026 has matured. Copilot earns its keep for specific roles (sales, finance, ops). Power Platform AI Builder is fine for low-volume. Custom integrations (your own AI on top of Microsoft Graph + Dataverse) is where the real wins are. Pick per use case; ignore the marketing.
Voice AI for service businesses: a buyer's guide
Voice AI works for service businesses with predictable call patterns and meaningful inbound volume. Booking, qualification, status, FAQ. Real cost ~€0.10-0.40/call. Real build cost €15-50k for a single-line deployment. Evaluate vendors on recording, escalation paths, and CRM integration — not on the demo.
AI agents for accounts payable: a deployment guide
AI agents in AP automate the high-volume, low-margin work of invoice keying and PO matching. Honest savings: €3-5 per invoice in loaded cost, 70-90% reduction in human handling time, payback typically 4-8 months on €25-50k builds. The agent isn't the hard part — the reviewer UI and the ERP integration are.
Why your AI chatbot fails (and what to fix)
Most chatbots that fail in production fail for one of six reasons: no retrieval, bad retrieval, no evals, no escalation, no observability, no scope. Tuning the prompt won't fix any of them. The fix is engineering — and the engineering is well-understood by now.
ChatGPT API vs Claude API vs Gemini: which to pick (2026)
Claude Sonnet 4.6/4.7 is our default for production agents — most reliable tool calling, best structured output, strong reasoning. GPT-4o wins for voice (Realtime is best-in-class) and the largest ecosystem. Gemini 2.5/2.0 wins for long-context, vision-heavy document work, and cost-sensitive volume workloads. Pick per task; abstract behind a provider interface.
Hiring an AI development agency: 12 questions to ask
Twelve questions that separate serious AI dev shops from demoware vendors. Asks about evals, observability, code ownership, provider lock-in, references, and what they'll refuse to do. If a vendor can't answer cleanly, walk.
How much does an AI agent cost? Real numbers from real builds
AI agent builds in 2026 typically cost €4-8k for discovery, €15-30k for a working prototype, €25-80k for production, €2-5k/month for retainer. Per-call infrastructure cost runs €0.01-€0.40 depending on shape. Honest numbers from real builds, with the trade-offs explained.
From PowerApps to production: when low-code stops scaling
PowerApps is excellent for many Microsoft 365 organisations until it isn't. It hits ceilings around concurrent users, complex business logic, performance, and developer ergonomics. The migration path is rarely 'rewrite everything in Next.js' — it's hybrid: keep PowerApps for what it's good at, move the parts where it's failing to TypeScript on Cloud Run.
RAG done right: the patterns that survive production
Production RAG is engineering, not magic. The patterns that survive: hybrid retrieval (vector + BM25), rerank top-k with a cross-encoder, metadata filtering, source dating, citation rendering, sampled human review. Without these, your retrieval is good in the demo and broken in production.
Building a phone agent with Twilio + GPT-4o: a complete walkthrough
Build a phone agent: Twilio provisions the number and streams audio, a Node.js bridge on Cloud Run pipes the audio to GPT-4o Realtime, function-calling tools execute real actions (book appointment, log lead, transfer). Recording, transcript, and observability on every call. Production deployment in 3-6 weeks.
How AI invoice processing actually works (and where it breaks)
Modern AI invoice processing uses vision LLMs (Claude, GPT-4o, Gemini) to extract structured data from PDFs and images, then validates against business rules and routes by confidence — auto-post, review queue, or reject. The model is not the hard part; the schema, the reviewer UI, and the eval suite are.
The state of AI development in 2026
In 2026, AI development is shipping production agents that earn their keep — document processing, voice, workflow orchestration — backed by Claude / GPT / Gemini and engineered with evals, observability, and guardrails. What's underrated: well-engineered automation with one or two LLM-judgment steps. What's overrated: 'autonomous AGI' marketing.
How AI agents actually work (under the hood)
An AI agent is a reasoning loop: the model plans, calls a tool, observes the result, replans. Underneath: function-calling APIs, retrieval-augmented context, typed tool schemas, guardrails, evals, and observability. This is the technical breakdown — what each layer does and how they fit together.
AI agents vs automation: which one do you actually need?
Use plain automation when the rules are deterministic — same inputs, same outputs, no judgment required. Use AI agents when inputs are unstructured (PDFs, emails, voice) or each instance needs a decision. Most production systems mix both: automation moves the predictable steps, an agent handles the messy ones.
What is an AI agent? The full breakdown
An AI agent is a system that turns a goal into a sequence of tool calls. Where a chatbot answers questions, an agent completes jobs. It plans steps, picks tools, executes them, recovers from failures, and either finishes the task or hands off to a human. The defining ingredients are a goal, retrieval, tools, guardrails, evals, and observability.