Sebastián Ocampo's lab
AI systems: agents, teams and results
Documented cases of AI systems and agents applied to real business functions (marketing today, finance and operations on the way): what was built, what it cost, what it produced and how it is governed. No borrowed theory: real operations.
Most content about AI in business is written by people selling tools or who have never operated one. This lab is the opposite: every case published here comes from a system I design and operate myself, with its real numbers. The format is fixed and non-negotiable: what problem existed, what system was built, what it cost (money and hours), what it produced and what rules govern it.
The lab advances by vertical. The first is marketing, with this house's complete system documented as a case study: y8y, a trilingual publication operated by one person with AI agents. The next verticals (finance and operations) will be published here in the same format once their cases clear the same bar: real numbers or nothing.
I work in Spanish, French and English, with a focus on Switzerland and the European markets where a well-governed system justifies a high ticket: there, governance is not an appendix, it is the entry requirement. If you want to discuss your case, the short path is the profile.
Case studies
- 2026-07-25 Case: a market study with six AI agents in parallel, one capture day and a datum-by-datum audit Six AI agents working in parallel, each on one market segment, captured 83 real AI leadership job postings across Switzerland and Europe in a single day and turned them into a structured dataset of 22 fields per posting, with the source URL on every row. What made it reliable was not the model but the contract: text fields copied verbatim, cells left empty when the data was not visible, every statistic's definition frozen in a file, and an independent audit that recounted each figure before use and found two publication-blocking errors. This case documents the full system, its decisions, its honest accounting and its failures.
- 2026-07-25 Case: governance that signs. How an AI-agent operation decides what gets published Governing AI agents is not writing a manual: it is turning every rule into a mechanism that can reject work. In this real operation, the editorial rules live as 34 automated tests that block publication (validated structure for every piece, one query per page, zero orphans, verifiable style), the agents' work passes those gates before a person reviews it and signs it by name, and failures are published in a visible log that feeds back into the rules. The principle transfers to any company: if your AI policy cannot reject a piece of work on its own, it is not governance, it is a suggestion.
- 2026-07-20 One person, three languages, 300 URLs: y8y's editorial system with AI agents y8y.ai is a trilingual robotics publication operated by one person with AI agents: more than 300 indexable URLs in Spanish, French and English, with source verification, structured data and automated distribution, on infrastructure with zero monthly cost (static hosting). The complete system (JSON content graph, automatic validation, editorial playbook and human governance of the final verdict) is documented in this case.
Operational blueprints
- 2026-07-26 Blueprint: identifying and prioritizing AI use cases (inventory, decision matrix and pilots that get measured) To identify AI use cases in a company, you inventory tasks (not technologies) by asking each function which work is repetitive, expensive or bottlenecked, you score each candidate on a three-criteria matrix (annual value in hours or euros, data and error risk, and feasibility with current tools), and you launch at most two pilots with a measured baseline and a decision date, where scale or kill are the only two valid outcomes. This blueprint brings the matrix with its scoring rubric, the one-page pilot charter, three worked examples and the five decision rules, exactly as we use them in our own operation.
- 2026-07-25 Blueprint: hiring a Chief AI Officer (what the market asks for, what it leaves unsaid and what that person must build) Before hiring a Chief AI Officer, decide which figure you actually need: in our own corpus of 79 recent AI leadership postings across Switzerland and Europe, none uses that title; the market hires Heads of AI and transformation leads, 43% are handed governance duties and 13 of 79 are asked to build the team from scratch. This blueprint turns those postings into decisions: which figure to hire for your size and sector, what to demand in writing (and what gets demanded without being written), the AI operating system that person must build in 90 days, the team that comes after and realistic year-one KPIs.
- 2026-07-24 Blueprint: rebuilding your content operation with AI agents (team, costs, risks and adoption) A content team operated with AI agents has three real line items: model subscriptions and API consumption (tens to several hundred euros a month depending on volume and model), the platform (from near zero on a static site up to CMS fees, plugins and publishing APIs), and the dominant one, the human supervision that reviews and signs every piece. Production per piece drops to cents or a few euros in tokens versus tens or hundreds per human piece, but the honest total only comes from adding all three: this blueprint includes an interactive calculator with published Claude and OpenAI prices so you can run your number, not ours.
Quick answers
- How do you measure the ROI of AI in marketing? AI ROI in marketing is measured as (incremental benefit minus total cost) divided by total cost, where total cost includes what almost everyone forgets: not just subscriptions, but the human hours of supervision, review and correction. Incremental benefit is measured against a prior baseline (what the task cost or produced before), never against zero. Without a baseline and a review-hours counter, any ROI presented is an opinion.
- How much does implementing AI cost in an SMB? Implementing AI in an SMB has three real line items: model subscriptions and API consumption (from tens to several hundred euros a month depending on volume), the platform or integrations (from near zero using general-purpose tools up to specific software fees), and the dominant line almost nobody budgets, the human supervision hours that review what the AI produces. A serious two-week pilot on a single task can cost only the subscription you already pay plus one person's hours; what costs thousands a month is not starting, it is scaling without having measured.
- How much does a Head of AI earn in Switzerland? The only salary bands published by Swiss employers in our corpus of real AI leadership postings (July 2026) are CHF 130,000-150,000 gross per year (base plus bonus) and CHF 160,000-200,000 per year, the second seen on a job-board mirror rather than the employer's page. These are reference points from specific postings, not market medians: only 2 of the 21 Swiss postings analyzed publish a band, so the real market is almost always negotiated behind closed doors.
- How long does it take to implement AI in a company? The honest timelines, measured on real operations: a serious pilot on a single task takes 2 to 8 weeks (two to build with general-purpose tools and the rest to measure against a baseline); a company's full AI operating system (inventory, usage policy, two measured pilots, committee and KPIs) takes about 90 days; and turning that into a lasting advantage is a permanent quarterly rhythm of rescoring cases and deciding, not a project with an end date. Distrust both extremes: the provider promising transformation in a week and the one budgeting a year before showing anything measurable.
- Is my company ready for AI? Your company is ready for AI if it can answer yes to four questions: do you have high-volume repetitive tasks written down with three numbers attached (hours they cost today, monthly volume and an owner)? Is the data for those tasks digital and accessible, not paper or heads? Is there a person with a mandate and real time to carry it (half a week for two quarters, not Fridays)? And is there an executive sponsor willing to let a pilot be killed if the numbers do not work? If the first fails, you do not have an AI problem but a method problem, fixable with two weeks of inventory; if the others fail, AI will wait until they are fixed.
- Why do AI projects fail? AI projects fail almost always for five causes that are not technical: the wrong task gets chosen (a flashy demo instead of a high-volume task with an owner), no baseline is measured before starting (so nobody can prove improvement), the pilot has no decision date (and ages in a drawer), nobody supervises the system after launch (quality degrades silently when data or the environment changes), and failures get hidden instead of logged (so they repeat). Independent research points the same way: failures come from misunderstanding the problem and the data far more than from the model.
- Why does AI that worked stop working? Because an AI system is not rules-based software: traditional software does today exactly what it did yesterday, and when it breaks, it breaks in plain sight. A system that learns from data depends on three things that change without notice (the data coming in, the model serving it and the environment it operates in), and when one changes, quality does not collapse: it slides. That is called drift, and it is invisible unless you are measuring it. The antidote has four pieces: a quality metric measured from day one, a threshold that triggers review when the metric drops, versioning so you can roll back to the previous state in minutes, and a quarterly date to rescore whether the system still deserves to be in production.
- What is the CAIM certification? CAIM (Certified Artificial Intelligence Manager) is the artificial intelligence management certification issued by PECB, an ISO-aligned certification body. It attests to the competence to lead AI projects in an organization: aligning them with business goals, interpreting their results and ensuring compliance with regulation and internal policies. It is earned by passing an exam after AI-management training, and its focus is direction and governance, not programming.
- What is AI governance in marketing? AI governance in marketing is the set of written rules defining what an AI system may do without human approval, what requires review, what is forbidden and who answers for every published output. The minimal practice is four documents: a list of delegable tasks by risk level, a pre-publication review flow, a register of sources and data the AI may use, and a named owner per system.
- What is an AI agent in marketing? An AI agent in marketing is a system that executes a complete task end to end (researching a market, producing a report, preparing a campaign) by using tools and taking intermediate steps on its own, unlike a chatbot, which only replies to messages. The practical definition has three pieces: a measurable goal, access to tools (search, read, write, publish) and written limits on what it may decide without a person.
- What is an AI use case? An AI use case is a concrete business task, with an owner, a monthly volume and a measurable current cost, that artificial intelligence can do or accelerate: for example, "classifying the month's 900 support emails, which today consume 45 hours of the service team". It is not a technology ("add a chatbot") nor an intention ("use AI in marketing"): if it cannot be written as verb, object, volume and owner in one line, it is not a use case yet, it is an idea.
- What is a fractional Chief AI Officer? A fractional Chief AI Officer is an external AI executive who works for your company a few hours or days a week instead of occupying a full salary: they build the system (usage policy, prioritized use cases, pilots with a baseline, KPIs) and train someone inside to run it. It makes sense for SMBs without the volume for a full role; in our corpus of 79 European AI leadership postings, 21 are consultancy or advisory-side, a sign this market already works. The clause to demand: operating documentation and one trained insider, so the knowledge does not leave with the advisor.
- What is an AI pilot? An AI pilot is a bounded trial (4 to 8 weeks, one single task) that measures whether an artificial intelligence system improves that task against a baseline measured before starting, with an agreed metric, a success threshold and a fixed date on which you decide to scale or kill. Whatever lacks a baseline, a metric and a decision date is not a pilot: it is a prolonged demo, and prolonged demos are the most expensive way of not deciding.
- What does a Chief AI Officer do? A Chief AI Officer is the person accountable for the company using AI with measurable value and controlled risk, and the real job has four layers: governance (usage policy, risk register, compliance with frameworks like the EU AI Act or the Swiss nFADP), use cases (inventorying, prioritizing, piloting with a baseline and deciding what scales), platform and data (models, tools, access and security) and adoption (training, internal champions and KPIs reported to leadership). In practice the European market almost never uses that title: it hires the same function as Head of AI or AI transformation lead.
How a case gets documented in this lab
Every case answers five questions in this order. The problem: which marketing task was expensive, slow or impossible by hand. The system: which agents and workflows were assembled, with which tools and what human role. The account: total cost (subscriptions, infrastructure and supervision hours) against measurable output. The result: what changed in the indicators that matter, with a date. The governance: what the AI may decide alone, what requires human review and what is forbidden.
The order matters because it is the inverse of tool marketing: there, you start with the demo and never reach the P&L. Here the P&L is question two, and if a system doesn't pass it, the case says so.
The verticals: same discipline, different P&L
Marketing is the entry vertical because its output is measurable in public (content, rankings, distribution) and its risk is bounded. Finance comes next: automating reporting, reconciliation and analysis, where the cost of an error rises and governance goes from advisable to mandatory. Operations closes the loop: the internal flows (support, documentation, quality control) where agents work the most hours with the least glory.
The order is not casual: it is the order of increasing risk. Each vertical inherits the previous one's governance rules and adds its own; in regulated markets like Switzerland and the European Union, that traceability (what the AI decided, what a person reviewed, with which data) is exactly what a buying committee demands before signing.
Free skills for AI agents
This lab's methods, packaged as installable skills (SKILL.md format, MIT license) for Claude Code, Claude.ai and any agent that reads the standard. Each is useful on its own and links to its long-form blueprint with examples and failures. Install instructions in the skills index.
| Skill | What it does |
|---|---|
| honest-case-study | Writes case studies in this house's six beats: problem, system, honest accounting, measured result, what failed and governance. |
| ai-use-case-matrix | Turns an AI wish list into two measured pilots: inventory, three-criteria matrix and a charter with a decision date. |
| research-integrity-contract | Makes agent research citable: verbatim-or-empty, URL per row, frozen definitions and an independent recount. |
| honest-ai-cost-account | Computes an AI workflow's real cost with the three lines: tokens, platform and the supervision almost everyone omits. |
Frequently asked
- Do I need a technical team to set up AI agents in marketing?
- To start, no: the first profitable systems are usually research, drafting and quality-control workflows on top of existing tools, assembled by the marketing team itself. You do need one person who thinks in systems (inputs, outputs, review, metrics) rather than prompt tricks. The jump to agents with data access and automated publishing does require technical judgment and written governance rules.
- Which marketing tasks should you automate first with AI?
- The ones meeting three conditions: they repeat weekly, they have a clear output format, and their errors are cheap to detect. In practice: competitive and keyword research, first drafts of content, quality control against a style guide, localization into other languages and reporting. What you automate last is whatever touches money or reputation without a net: direct publishing, customer replies and media buying.
- Will AI replace marketing teams?
- It will change their shape before their existence. What disappears is pure production work (versions, formats, translations, reports); what gains value is judgment: which story to tell, what to measure, what never to delegate. The cases in this lab show smaller teams operating bigger systems, with the person in the role of editor and P&L owner, not typist.