Insights
Page 3 of 6
A development partner for your startup: future, not legacy
Choosing the software development partner for your startup is one of the most expensive decisions you make before you have revenue, and it's almost always made on the wrong criterion: who delivers fastest and cheapest. Speed has two invoices: you pay the first now; you pay the second when you scale, when code built "by feel" —no spec, no tests, understood only by whoever wrote it— shows up as technical debt in your Series A due diligence. Legacy isn't old code: it's code nobody can change safely, and in a startup it appears within six months. How to choose a partner that builds to scale (Spec-Driven Development, context engineering, human verification) and leaves the method and control in your team — with five questions to bring before you sign.
Copilot/Cursor ROI: why you're measuring what doesn't matter
The ROI of Copilot, Cursor or any AI copilot is almost never measured well: it's counted in active licenses and in a "sense of speed", two metrics that age badly. This week the capital markets gave a hint of why —Indian IT loses more than 46% of its value since the Aug-2024 peak and TCS opens earnings with margin compressed to 24.0%—: they're putting a price on the difference between renting capacity and retaining method. Change the numerator and the denominator of your ROI calculation: measure by the capability your team retains (method, context, criteria), not by installed licenses. That's the asset nobody can reprice on you.
AI-accelerated dual-track Scrum: how to implement it in your team
You bought AI licenses for your team and the velocity is still the same. It's the number-one complaint from CTOs in 2026, and it almost always has the same root: AI accelerates writing code, but a team's bottleneck is rarely writing code — it's deciding what to build and verifying that it works. A dual-track Scrum separates the lane that decides (discovery) from the one that builds (delivery via Spec-Driven Development), with AI doing the heavy lifting of both and the human signing off at the gates. How to implement it in 8 weeks without stopping delivery, what to measure and why in the teams onext transforms it's worth ×7 in velocity with 0 sprints lost.
Where a CEO starts transforming their company with AI: the 5 axes of the first quarter
A mid-market CEO knows they have to get started with AI, but doesn't know where to begin. The mistake isn't theirs: it's whoever answered "buy a tool". Before the first euro, your company needs an honest inventory of five axes in ninety days — real processes, corporate intelligence, tools already active with no policy, governance and EU AI Act, and internalized transfer. It isn't run by your CTO (technical perimeter) or your consultancy (which sells dependency). It's delivered by a new role with a mandate from the CEO: the AI Officer, with a time-to-independence metric and an exit date from day one.
The 6 technical changes to your Claude Code setup before June 15: agent budgets, rate limits and observability
On June 15 Anthropic starts counting Claude agent consumption separately from human chat consumption. Without technical governance before the switch, the agent bucket empties 2-3 times faster than the chat bucket. The 6 changes: separate budget in CLAUDE.md, per-workflow rate limits in CI, OpenTelemetry traces per agent, no-loop rules in SDD specs, human approval above a tokens/hour threshold, and a weekly audit of input vs output tokens.
Your Claude bill changes on June 15: the FinOps conversation your CFO is about to start (and how to arrive prepared)
From June 15, 2026, Anthropic separates programmatic usage from interactive usage across all its subscriptions. Pro $20 → $20 in credits; Max 20x $200 → $200 in credits; Team $30/seat → $20 credits/seat. When they run out, the agent doesn't stop — but the cost now comes out of the API billing pool with no discount. The six FinOps controls before the switch: per-workflow tagging, per-workflow cap, priority by business line, budget alerts, multi-model hedge and evaluation with cost built in. The InfoWorld phrase says it clearly: "the direction will not vary." OpenAI, Google and Microsoft will replicate it in 12-24 months.
HTML beats Markdown for artifacts that live: the Spec-Driven Development decision table
Of the 14 Spec-Driven Development artifacts, 8 migrate to HTML and 5 stay in Markdown. The decision isn't ideological — it's operational: the larger the artifact and the longer its lifespan, the more sense HTML makes. Includes the full table, the reframe of the Tech Lead as a "compute allocator" per Thariq Shihipar, and the update to onext's AI Engine method as of May 2026.
Month 6: why your dev team's AI transformation breaks at that point — and the sustaining system that prevents it
The first 3 months of AI transformation are enthusiasm with method. By month 6, most teams return to baseline without knowing why. The collapse always follows the same order: stale workflows → false positives in quality gates → silent bypassing → back to square one. The sustaining system that prevents it has three pieces: a rotating owner, a single health metric and a monthly 30-minute ritual.
The 70% gap: your AI adoption in the development team isn't a technical problem — it's an organizational problem
More than 40% of agent-based projects will be cancelled before 2027 (Gartner). After 12 transformations in development teams, the pattern is consistent: 70% of AI success isn't technical. The 5 organizational dimensions no tool solves on its own. Four questions to tell a technical problem from an organizational one in your next committee. Includes the METR paradox and a Karpathy quote.
The model isn't the source of truth. The chat isn't the system. Code isn't the only artifact.
The METR paradox: +20% subjective productivity, −19% real. The reason is the absence of three disciplines almost nobody applies. Claim 1: the model isn't the source of truth → project constitution. Claim 2: the chat isn't the system → workflows with quality gates. Claim 3: code isn't the only artifact → Spec-Driven Development. Three installable antidotes. An honest Karpathy quote to close.
Start with AI now or wait 6 months? The right question is a different one
The AI decision is the most deferred one in Spanish leadership committees in 2026. This piece separates the three legitimate reasons to wait (high regulatory risk with no legal analysis, data with no operational conditions, nobody with the judgment to lead) from the three reasons to start now (organizational learning curve, competitive window, talent). It brings the empirical record from Bain & Company (41% in production report positive results vs 25% in pilot) and the public Q1 2026 cases from European and American integrators. It closes with four contexts on each side, the two traps that invalidate the decision and the single question that frees the committee from the cycle of deferrals. Includes a downloadable one-page matrix to take to your next committee.
Context engineering: why it will survive the agent Marketplace
A Q2 2026 pillar piece on the one layer the Marketplace can't sell. It separates the three layers of the enterprise AI stack (cognitive, control, context), explains why only the context layer resists commoditization, and breaks the discipline down into five operational layers: curated instructions, working memory, episodic memory, retrieval policy, and invalidation/expiration. It closes with the cost-per-request contract, the three most recurrent anti-patterns (RAG as a band-aid, prompt-as-database, memory with no TTL), and a brownfield migration plan executable in four weeks without stopping production.
Sovereign AI is no longer data residency: the split of control in 8 axes
Q2 2026 pillar piece that closes the sovereign AI cluster. It separates the three concepts that get mixed in the Spanish and European market (infrastructure sovereignty, vendor-stack governance, operational sovereignty) and develops the split of control across eight operational axes: identity, data, model, context, agent, logging, human-review gates and exit. It includes the comparison matrix of onext against three market archetypes (industrial sovereign-AI brand, EU sovereignty integrator, Microsoft Agent 365 GA May 1 with the E7 bundle at $99/user/month). It closes with when you need each proposal and the exercise to take to your next RFP.
Context engineering: 7 questions for your next AI RFP when headcount is no longer the criterion
Analysis for CTOs/CIOs/Heads of Procurement preparing an AI RFP in 2026: the seven questions that filter "real context engineering" from "context engineering as a label". It covers documented methodology, verifiable flagship cases, auditable internal certification, ownership of the context layer, operational governance, portability at contract end, and a reproducible eval framework. Includes concrete red flags, a preferable alternative for each question, a real-engineer-vs-label summary table, and a one-page downloadable PDF ready to take to the committee.
Google Cloud Next 2026 for the enterprise CTO: 5 announcements that matter and 5 that are noise
Google Cloud Next 2026 closed with more than 200 announcements and a new vocabulary (agentic cloud, agent control plane, agent marketplace) that's going to be in your next committee before May. The five announcements that shape your roadmap: Agent Control Plane as a category, Gemini Enterprise native in Workspace, Partner Fund $750M + Faculty embedded in Accenture, Deloitte's catalogue of 1,000+ agents and PwC $400M in compliance AI Act. The five that are noise for your seat. Includes three questions to add to the RFP, a map of the AI stack in three layers (cognitive, context, control) and what to do before July 1.