Recipes for AI: From Pilots to Production
Álvaro de Nicolás ·
What follows are the practical lessons, methodologies and recipes I have distilled from the Revamp, HBX and Zapier playbooks, cross-checked against the latest industry reports. They are organized to be applied immediately: model management first, then segmentation, then the cultural traits and the timeline-by-timeline recipes that separate the companies extracting value from the companies still stuck in the pilot graveyard.
1. Managing the model: hallucinations, reasoning and “thinking”
The core lesson. Do not mistake fluency (sounding smart) for reliability (being right). Large language models act as a vast index of memorized patterns, not a logical brain.
Practical recipes to avoid hallucinations
- Use “thinking slow” architectures. Do not ask the LLM to solve complex constraint problems directly. Use it as a universal UI to translate human language into code or commands for reliable tools (calculators, databases).
- The “data structure” fix. Simple RAG implementations often fail because models struggle to link data silos. Invest in offline processing where you use LLMs to pre-build indices, summaries and links between your documents before a user asks a question.
- Force “thinking” mode. For complex tasks, use chain-of-thought reasoning models (OpenAI o-series, DeepSeek R1). They plan before answering and raise accuracy on logic tasks — though they are not immune to errors.
- The golden rule of prompting. If the information isn’t in the prompt, the model generally doesn’t know it. Treat the model as stateless and predictive: it guesses the next word, it does not check facts.
Questions to ask ourselves
- Am I asking the AI to reason (high risk) or to transform information I provided (lower risk)?
- If the data changes (a new CEO, a new policy), will the model fight its training data? RAG systems often hallucinate when new facts conflict with pre-training.
2. The Revamp Matrix and LLM cross-checking
The Revamp principles map cleanly onto current LLM capabilities — and reveal where the analogy strains.
- “Data must remain in-house for high-sensitivity tasks.” Convergent. Open-weight models (DeepSeek R1, Llama) allow local hosting and align with the privacy mandate. Divergent: proprietary models still lead on multi-step agentic behaviour.
- “Human-in-the-loop is mandatory for judgment-based tasks.” Convergent. Current AI achieves only ~60–70% reliability on basic reasoning. Using AI for drafting (contracts, HR letters) is safe; using it for decisions (hiring) without oversight is dangerous.
- “Focus on task automation, not job replacement.” Convergent. The PwC Jobs Barometer shows AI-affected industries see ~3× higher productivity growth, not job collapse. The focus is augmenting human judgment, not replacing it.
3. Methodology: segmentation and triage
Do not ask “can AI do this?” Ask “should we do this now?”
Step 1 — the feasibility filter
- Data readiness. Is the data clean and accessible? If you need to scrape from three incompatible systems, the use case is Tier 3 or 4 (park / later).
- Process maturity. If the process varies every time a human does it, AI cannot automate it.
- GDPR & privacy.
- Green (low risk): no personal data (policy summarization, coding). Deploy immediately.
- Yellow (medium risk): pseudonymized or internal data. Enterprise licence + DPIA required.
- Red (high risk): automated decision-making on personal data (firing/hiring). Do not deploy.
Step 2 — the value filter
High value solves a strategic bottleneck (skills shortage) or a massive time-sink (reporting). Low value is “cool tech” with little business impact (generative avatars for internal meetings).
The resulting matrix
| Tier | Profile | Example | When |
|---|---|---|---|
| Tier 1 | High value + high feasibility | Standardized HR reply generator, code documentation | Year 1 |
| Tier 2 | High value + medium feasibility | Burnout detection (sensitive data) | Year 1–2 after governance |
| Tier 4 | Low value + low feasibility | Full autonomous org design | Ignore |
4. Corporate traits and pitfalls
Traits of successful adopters
- Executive clarity. The CEO or CFO must explicitly state “AI is a baseline capability.” A cultural signal, not just an IT project.
- “Code red” urgency. Treating AI adoption as immediate strategic necessity, not a “nice to have.” In the early stages, speed matters more than perfection.
- Distributed innovation. They don’t wait for a central AI team to build everything. They empower AI champions in every department to build their own workflows.
Common pitfalls
- The pilot graveyard. Many pilots, no plan to scale. Fix: define scaling metrics (ROI) before the pilot starts.
- Legacy mindsets. Bolting AI onto broken processes. Fix: redesign the workflow first. Don’t automate a bad process — eliminate the friction.
- Data paralysis. Waiting for “perfect data.” Fix: start with use cases that rely on unstructured data (text, documents) which don’t require perfect schemas.
5. Recipes by timeline and impact
Year 1 — the quick wins (cost & efficiency)
Minimal integration, immediate time savings.
- Recipe 1 — the “first draft” engine. AI drafts standard responses (HR queries, customer support), contracts and policy documents. Upload policy PDFs to a secure RAG system; prompt: “Answer this employee query based only on the attached policy.” ROI: 250–300 hours saved per team per year.
- Recipe 2 — meeting & knowledge distillation. Auto-summarize meetings and extract action items via Microsoft Copilot or specialized plugins. ROI: 150+ hours per manager per year.
- Recipe 3 — coding & technical acceleration. GitHub Copilot or Cursor for code generation, documentation and legacy refactoring. ROI: 20–50% developer productivity gain.
Year 2 — strategic integration (value & revenue)
Requires data integration and governance maturity.
- Recipe 4 — the skills intelligence layer. Shift from job titles to a skills-based organization. Use AI to infer skills from CVs, project history and reviews to build a dynamic skills graph. ROI: enables internal mobility, reduces hiring cost.
- Recipe 5 — customer/talent signal detection. Sentiment analysis on sales calls or employee surveys flags churn and flight risk early. ROI: 10–15% reduction in unwanted attrition.
- Recipe 6 — testing & calibration. Use AI to check performance reviews for bias and inconsistency across managers. Anonymize the data and score for clarity, bias and actionability. ROI: higher fairness and defensibility in HR decisions.
The board-level takeaway
The competitive edge in 2026 is not adoption of AI — it is the discipline of AI. Everyone has access to the same frontier models. The moat is organizational: governance, security, workflow redesign, and an honest tiering of use cases that respects feasibility and value at the same time. Less hype, more plumbing.