---

Ribeirão Preto, an in-person course on installing solar panels. A clay-tile roof built on the ground — it wasn’t even a real roof. And there I was, looking at it, thinking: I studied 250 hours for this?

I had just founded the first solar energy company in João Monlevade. Six certification modules at Blue Sol. Scores between 92% and 100%. And all it took was a training roof sitting flat on the ground for me to realize that heights were not for me. Not one meter up. Not even a fake one.

That question — asked in front of a training roof that wasn’t even real — created everything that came after. If I wasn’t going up on roofs, what was I going to do with 250 hours of solar energy, a technical degree and 12 years of experience in web and marketing?

The answer came months later, when the partners joined: a marketplace. Not selling and installing panels — building the platform where other people sell and install.

Every pivot in my career followed that pattern: a problem that looked like a loss turned into the best investment I ever made. Fear of heights turned into a marketplace. A dead startup turned into a steady client base. Three universities with no degree turned into 800 hours of problems solved. Not because I’m resilient — because my operating model gains from error.

If you’re a technical founder running a lean team, this story is probably yours too.

---

The AMD K6 and the world before Stack Overflow

In 1998 I was 11 years old and had the only computer in the neighborhood. AMD K6, 4GB hard drive, Windows 98. People paid me to format their theses to ABNT standards (Brazil’s national standards body). I charged R$ 1.00 per page. Typing and formatting. At 11.

It wasn’t a calling. It was what I knew how to do with what I had.

At 15, a technical course in computing. At 17, freelancing in design and web. In 2004 it was Internet Explorer 6, where everything broke. There was no Stack Overflow in Portuguese. When I got stuck on an error, I read the English documentation — as best I could — and tested until it worked.

I got into UFES for Graphic Design and transferred to UEMG in 2007. In 2010, I joined Zavala Propaganda — where I built Aneethun’s website and the English and Spanish sites for Bio Extratus. Brand internationalization for a cosmetics company from Minas Gerais, back when “responsive website” was still a novelty.

---

DENATRAN, the police officers and the hacked devices

In 2012, GCT in Belo Horizonte. Traffic Management and Control. My job title: IT assistant. My actual work: QA on the certification of the first electronic ticketing device approved by DENATRAN (Brazil’s national traffic department) in the country.

An electronic device that replaces the paper pad an officer uses to write tickets. Every field, every flow, every validation had to work perfectly — because if it didn’t, DENATRAN wouldn’t approve it. And if they didn’t approve it, the whole project was dead.

I was there in person for the federal audit. There’s always something that needs adjusting in the heat of an audit like that. The programmer was good. We passed.

After that, 16 cities across Brazil training teams. Salvador, Uberlândia, Governador Valadares, Feira de Santana, Teófilo Otoni. Each city with different hardware and people who didn’t want to use the system. Military police, municipal guards and traffic officers who sometimes preferred the paper pad to the technology.

And the problem nobody saw coming: the officers hacked the devices. They installed WhatsApp, games and personal apps on equipment meant exclusively for issuing traffic citations. We had to implement blocking tools.

I implemented remote support — and that changed the operation. Dozens of devices a week, between maintenance and new ones to configure. What used to take days started taking hours. What sometimes left a professional in Bahia without equipment for a week was now resolved the same day.

What I learned at GCT doesn’t fit on a certificate: debugging under pressure with the federal government in the room, and training people who don’t want to learn.

---

Oak: 200 partners, 50 power plants, and a law that changed everything

After the roof in Ribeirão Preto, the pivot to a marketplace created Oak Energia — the first renewable energy marketplace in Brazil. Not just solar: photovoltaics, micro hydroelectric plants, biodigesters, vertical-axis wind turbines.

200 installer companies registered as partners. 50 power plants sold and approved through the platform in cities all over Brazil. A YouTube channel with weekly content. 3 editions of the Feira de Sustentabilidade do Médio Piracicaba (a regional sustainability trade fair). The idea was very well received wherever we talked about it — in Brazil or abroad. We never saw anything like it.

I took Fórmula de Lançamento (a Brazilian product-launch course) because I needed traction. SEBRAE Fast Motion (SEBRAE is Brazil’s small-business support agency) invited us into their accelerator. Conversations with investors. In parallel, I got into UFOP for Computer Engineering — first place overall. Five semesters juggling university and the startup. I stopped during the pandemic. I never went back.

The partners and I had agreed on a 5-year horizon. And, like so many startups, we didn’t become a unicorn.

---

The meeting

Long sales cycles. Slow approvals. And ANEEL (Brazil’s electricity regulator) revised the rules on self-generated energy. The legal uncertainty froze the entire sector.

It wasn’t a dramatic day. It was a meeting between partners. We looked at the numbers, at the 5-year horizon we had agreed on, at the regulatory picture. And we decided each of us would go his own way.

Six years. 200 partners. 50 power plants. 3 trade fairs. A YouTube channel that still exists today as a memento. And the company idle, ready to be restarted if someone wants to buy it.

Most people look at that paragraph and see failure. I look at it and see this: I lost the company, but I didn’t lose a single capability. I knew how to build a platform. I knew how to close R$ 300,000 deals over WhatsApp, without ever meeting the client in person — and with a client who was never in a good mood. I knew how to organize 200 partners who had never met each other. And I knew how to start over.

The cost of the mistake was 6 years. What I learned from the mistake built everything that came after in weeks. That isn’t resilience — resilience is taking the hit. This is gaining from the hit.

---

The cemetery that arrived before the name

The transition was immediate. The first client came through a referral — a cemetery — before the company even had a new name.

Think about what that means. A cemetery trusts its digital operation to someone who has just shut down a startup. They didn’t know me personally. Someone referred me. And the system I built had to work when a grieving family decided to pay an installment on a Sunday night. No margin. No “let’s see if it works.” If the portal failed, the family went home without resolving what they came to resolve.

That’s the level of trust that holds up what we do. And the level of consequence that has shaped the whole operation ever since.

Juliana — an international relations specialist with a graduate degree in Economic Engineering — gradually took over management, content and strategy. She isn’t a designer who became a CEO. She’s someone who understood that strategic content is what sustains a client base over the long term.

By the time we chose the name “Evolutiva Negócios Digitais,” the agency had already been running for months. The clients came before the name.

---

What we do and why it works

Two founders. A steady client base for 4 years. Our clients are cemeteries, currency exchange houses and e-commerce stores.

The cemetery’s 24/7 payment portal, with omnichannel support and on-call teams, has to work all the time. If it fails, the grieving family isn’t served when they need it, and the people who need these services urgently can’t use them.

The exchange house’s chatbot has to quote the right rate. If it fails, the client gets a made-up number and loses money.

The e-commerce fraud screening has to get it right. If it gets it wrong, a legitimate customer is blocked and never comes back.

If the system fails, we lose the client. We have no layer of protection, no legal department, no investor covering the loss. That’s what makes us reliable — the skin in the game is real.

Three layers hold up the operation:

Scrum adapted for a portfolio. A weekly sprint covering every client. A binary target of 5 items — either it passed or it didn’t. One single big thing per day. Three hands per card: whoever did it doesn’t verify it, whoever verified it doesn’t move it. 7 completed sprints, a retrospective every Friday.

Sovereign AI. Language models running on our own infrastructure — zero cost per query. This isn’t “we use ChatGPT.” It’s: we have a server with models that know each client’s business. The cemetery’s chatbot schedules visits. The exchange house’s answers rate questions with real data. The salon’s knows the history of 82,000 appointments over the last 15 years. When the AI vendor tripled its price — and that has already happened three times since 2023 — our clients didn’t feel it.

260 problems solved and documented. Every problem becomes a reusable procedure. Every error becomes a documented trap. Production deploys went from 45 min to 8. The monthly report, from 4h to 40 min. The base grows every week — compound interest applied to operations.

I’m publishing these numbers because it took me 6 years to find them — and nobody should need that long.

---

The numbers that don’t fit on a diploma

3 universities: UFES, UEMG, UFOP. No degree. First place overall in Computer Engineering.

800 hours of courses and training with certification from 12 issuers. From Google Ads to photovoltaic solar energy. From UX Research to advanced negotiation.

Formal credentials are fragile — they lose value when the market changes. The ability to solve new problems is antifragile — it gains value when the market changes.

The 250h of solar looked useless. Until I needed to model a R$ 50,000 investment. The 444h at Alura looked like a “beginner course.” Until I needed to build an entire system on my own.

---

For anyone who recognizes themselves here

If you run a lean team and the bottleneck is time, not talent — the way out isn’t hiring more people. It’s multiplying what each person resolves.

Four years with a steady client base. Two founders. Zero clients lost to technical failure.

Full list of every course: evolutivanegociosdigitais.com.br/certificados-joao-vitor/

I serve 4 steady clients. If you want to talk about sovereign AI, adapted Scrum or how to operate with a lean team before my calendar fills up — send me a message on LinkedIn.

---

I spent months trying to identify a bird in my backyard. I built an Alexa skill so I’d never have to do it again.

A pair of green birds showed up every morning on the same wire, near my house. I was sure they weren’t maritacas (plain parakeets) — the size was different, the noise was different. But I didn’t know their name.

It took months. I searched WikiAves, which is the reference in Brazil and excellent work — but for a beginner like me, the sheer volume of information was more intimidating than useful. I didn’t know the right terms to filter by. I didn’t know whether “green with a red patch on the wing” was an exotic sabiá (thrush) or a common parrot. I found out later: they were Maracanã parakeets.

And then came the irony that planted the whole thing: I went to check what their call sounded like so I could confirm it, and found out that parrots don’t have a standardized call. The song I wanted to use as proof simply doesn’t exist in any classifiable form. The challenge was already in my head.

The first idea was wrong (and unethical)

My original plan was to use Alexa to listen to the bird sounds around me and identify them by audio. Two problems killed that idea the same day:

Alexa only picks up human voice. It doesn’t listen to the environment — the microphone is optimized for spoken commands. Background sounds are noise, not signal. There is no “listening mode” I could turn on.

Attracting birds with playback is a questionable practice. Playing a species’ song to attract it causes territorial stress, can drive birds away from their nests and interferes with breeding. IN ICMBio 14/2018 (a regulation from Brazil’s federal biodiversity agency) and CEMAVE (its bird research and conservation center) advise against it. When I found that out, I dropped the idea without a second thought.

So the skill became something else: instead of “record what you heard”, it became “describe what you saw”. You talk to Alexa in natural language — color, size, where it was, bill shape — and it returns the most likely species.

The stack: Python, Bayes and four open databases that don’t talk to each other

Alexa isn’t exactly a smart device. It recognizes voice and calls a function in the cloud. All the intelligence sits in a Lambda (Python) on AWS, and the real work was making four databases that were never built to talk to each other produce a coherent answer.

AVONET (Tobias et al. 2022, CC BY 4.0) has standardized morphological traits for more than 11,000 species — that’s where the 7 attributes the skill extracts from your description come from: primary color, secondary color, size by anchor, habitat, posture, bill and tail.

GBIF has occurrence records for birds in Brazil. I used that data as a Bayesian prior: if you’re in São Paulo and describe “a big green bird”, the scorer gives more weight to the species that actually show up in the region (30 km radius, with at least 20 records). Bem-te-vi (great kiskadee) before a rare species with the same description.

xeno-canto has the real song recordings. And here comes a decision that wasn’t optional for me: I couldn’t just hook into the audio without crediting whoever recorded it. Every playback shows the recordist’s name, the license (CC BY-NC-SA 4.0) and the recording number in the archive. Python turned out to be great for pulling those audio files and organizing both species and authors programmatically.

Han et al. 2025 (CC0) fills in with plumage color data that AVONET doesn’t cover.

The AMAZON.SearchQuery slot type in pt-BR captures free speech — “I saw a yellow little bird with black wings on the ground in the backyard” comes in whole as text and the parser extracts the attributes. There’s no menu, no “say 1 for color”. You speak the way you’d speak to a friend.

The numbers, with nothing made up

1,627 species from all of Brazil. 1,014 with a real xeno-canto recording. The database started with 390 species from the São Paulo region and was expanded to cover every Brazilian biome.

Accuracy simulated with realistic noise (incomplete descriptions, wrong attributes, regional synonyms): the right bird shows up in the top 3 suggestions in ~54% of cases. It’s not 90%. It’s what the model delivers today with open data and a parser that accepts free speech. Being transparent about that number matters more than rounding it up.

There’s also a song quiz: the skill plays a real recording and you try to guess the species. Get it right and it keeps score. Get it wrong and it tells you what it was and plays it again so you can make the association. It’s a way to train your ear without having to be out in the field.

And the biggest engineering bottleneck wasn’t the scorer — it was the Alexa Developer Console. Testing an Alexa skill is a slow process, with deploys that drag, logs that lag and an interface few people have mastered. Anyone who has developed for Alexa knows. Anyone who hasn’t will find out.

What the skill does NOT do (on purpose)

It doesn’t record the bird’s sound — Alexa can’t. It doesn’t loop songs to attract birds — that’s unethical. It has no usage metrics because it was just published. It has no user testimonials because I’m not going to make them up. WikiAves, xeno-canto, GBIF and AVONET are credited sources, not partners.

Every song playback comes with a responsible-playback notice. Single playback, controlled volume, credit to the recordist, guidance about the breeding season. That isn’t a feature — it’s an obligation.

The invitation: test it, break it, improve it

The code is MIT and the repository is public: github.com/jvitorcarvalho/aves-do-brasil-alexa.

I want people to have fun with this. To discover the call of the urutau (the common potoo), which is one of the most curious things you’ll ever hear. To identify a bird in the backyard and tell a friend over lunch. For a developer in Manaus to grab the repository and calibrate the priors for the Amazon. For an ornithologist to fix the common names that came from GBIF and probably have errors.

Today the skill covers all of Brazil — 1,627 species, 1,014 with a recording. There’s room.

To try it: “Alexa, abre Aves do Brasil”. It’s free, no ads, no data collection.

To contribute: open an issue, send a PR, or tell me that the common name for the tico-tico (rufous-collared sparrow) in your town is something else. It all helps.

---

Developed by Evolutiva Negócios Digitais (João Monlevade/MG, Brazil). Sources: AVONET (Tobias et al. 2022, CC BY 4.0), GBIF (CC0), xeno-canto (CC BY-NC-SA 4.0), Han et al. 2025 (CC0), WikiAves (public search). Playback notice follows IN ICMBio 14/2018 and CEMAVE guidelines.

Frequently asked questions

How does the Aves do Brasil skill identify a bird from a spoken description?

The skill uses a Bayesian scorer that cross-references 7 attributes extracted from your speech — color, size, habitat, posture, bill, tail and secondary color — with morphological data for 1,627 Brazilian species. The data comes from AVONET (Tobias et al. 2022), complemented by GBIF occurrence data to weight the species most likely in your region. You describe what you saw in natural language and get back the species that best match the description.

Does the skill play songs to attract birds?

No. The skill plays real xeno-canto recordings only once per query, with controlled volume and credit to the original recordist. Looping songs to attract birds causes territorial stress and can interfere with breeding — a practice discouraged by IN ICMBio 14/2018 and by CEMAVE. Every playback comes with a responsible-playback notice.

What is the hit rate of the voice-based identification?

In simulations with realistic noise (incomplete descriptions, wrong attributes, regional synonyms), the right bird shows up in the top 3 suggestions in roughly 54% of cases. That number reflects the current performance with open data and a free-speech parser. The skill doesn’t round its metrics up — being transparent about the limitations is intentional.

Is the Aves do Brasil skill free, and does it collect personal data?

The skill is free, with no ads and no personal data collection. The code is open source under the MIT license, available in the public repository. The databases used (AVONET, GBIF, xeno-canto, Han et al. 2025) are all open and properly credited.

Can I contribute to the skill or correct common bird names?

Yes. The repository accepts issues and pull requests — contributions from ornithologists, developers and birdwatchers are welcome. Regional common names that differ from the GBIF records can be corrected directly. The current database covers 1,627 species from Brazil and 1,014 with a real recording, but there’s room for regional calibration and expansion.

In three years I put an AI agent into production, and then I took it offline with my own hands. It held a good conversation. It answered accurately. And it solved nobody’s problem.

That’s the part almost nobody tells. The market is full of launch stories and empty of shutdown stories — and the second kind teaches more.

This article has two halves. In the first, my honest inventory: what’s running, what’s still in testing, what died, and why the thing that increased my output the most wasn’t artificial intelligence. In the second, the market numbers that explain why this story matters to anyone working in tech today — including why Scrum Master openings fell 62% while 85% of Product Manager openings started asking for AI.

Who’s talking

I started in graphic design. I moved to IT, spent years in WordPress and web development — more than two decades working on sites, servers and things broken in production. Today I work in MarTech: the layer where marketing, systems and money meet.

I’m not an AI researcher. I’m the guy who answers the phone when checkout stops processing on a Friday night. My clients are real companies that, combined, bill more than one million reais per month, and what they buy from me isn’t technology — it’s predictability.

That changes what I call success. A model that impresses in a demo and fails at the edge of the real flow is worth nothing to the people who pay me. That yardstick explains every decision that follows.

Why I stayed quiet

I founded Evolutiva in 2022 and published little after that. It wasn’t a content strategy. It was lack of time and, to be straight about it, some reluctance to talk before I had something to show.

Three years later I do. Including what didn’t work.

The AI agent that went into production and I took offline

It worked. And keeping it live was still a mistake.

A client was getting dozens of calls a day with the same questions: contract status, duplicate invoice, transfer of ownership. The information existed — scattered across thousands of documents, exports and spreadsheets piled up over years.

I built a RAG agent (Retrieval-Augmented Generation — in plain terms: the AI looks the answer up in the client’s real data before responding, instead of making it up). It shipped to production in partnership with another technology company.

The agent understood the question. It found the right document. It phrased the answer clearly. By any conversational metric, it passed.

And I took it offline.

Why I shut down something that worked

Because answering well isn’t the same as resolving.

The client’s system exposed no integration points. The agent could say “your invoice was due on March 12”, but it couldn’t issue the duplicate invoice. It knew a transfer was possible, but it couldn’t start the process.

It was a brilliant employee locked in a room with no phone.

The customer calling in didn’t want information. They wanted the problem solved. The agent delivered half the way — and in customer service, half the way is sometimes worse than nothing, because it creates an expectation and sends the customer back to the queue.

The failure wasn’t the model’s. It was the infrastructure underneath it. And keeping live a system that makes a good impression without delivering results is technical debt dressed up as a showcase.

That was the most expensive lesson I’ve learned, and it became my rule: AI without the ability to act is a demo, not a product.

Why a legacy system became a SaaS before reaching production

Another client ran on a legacy system that could no longer handle the volume: recurring billing, manual reconciliation, spreadsheets passed around by email. I started the rebuild — database migration, containers, data pipeline, automated billing.

Halfway through I noticed something that changed the project: the problem wasn’t that client’s. It was the whole sector’s.

Every business in that market had the same pain, the same manual process, the same spreadsheet. Building a custom solution for one client would solve one case. Building a multi-tenant platform — a SaaS where each company gets its own isolated environment on the same base — would solve the market.

I rebuilt the architecture as SaaS.

Honest caveat: that system hasn’t reached production yet. It’s in pilot. I could write “SaaS platform for the sector” and let you assume dozens of clients are running on it. They aren’t. There’s a finished architecture, a pilot underway and a product decision I stand behind — but production is a word you only use after a real user depends on the system to do their job.

Why the FX desk is still in testing (and will stay there until it’s right)

I built a digital FX trading desk (foreign exchange): quotes, contracts, compliance documentation, all in real time. In it, an AI agent acts as the system’s interface — the operator talks, and the agent executes inside the real flow.

That’s the exact opposite of the first case: here the AI has hands.

And that’s precisely why it’s still in testing.

The system moves real money, from real clients, across different countries. The operation has to be fast, because FX has a quote window. It has to be error-proof, because a wrong digit isn’t a bug — it’s somebody’s loss. And it has to be auditable, because there’s regulation involved.

When the cost of an error is financial and cross-border, “it seems to be working” isn’t a release criterion. Rushing here isn’t agility: it’s recklessness.

What I learned spending more than R$ 25,000 on AI

I’ve invested more than R$ 25,000 in artificial intelligence tools over the past few years — ChatGPT, Gemini, Claude, Perplexity, OpenRouter, OpenCode, in practically every combination of subscription and model that existed in that period.

It wasn’t a course. It was daily use, on work that had to ship.

The most uncomfortable lesson: what I knew changed almost every week. The model that was best for a task in January wasn’t in March. The prompt trick that worked stopped being necessary once the model improved. A tool that looked essential became redundant with one release.

Anyone who treated AI as something you learn once ended up with dated knowledge. What accumulates isn’t the shortcut — it’s the judgment: knowing when a model is good enough, when the problem isn’t the model, and when the right answer is to use no AI at all.

What connects agents to systems: MCP

What ties all of this together today is MCP — Model Context Protocol, a standard adopted by Anthropic, Google and Microsoft to give language models “hands”.

In practice: with MCP, the model doesn’t just talk. It queries databases, triggers processes, records information and returns results inside the system.

Remember the agent stuck in the room with no phone? MCP is the phone. If it had existed with today’s maturity back when I built that project, I probably wouldn’t have shut anything down.

What actually doubled my output — and it wasn’t AI

It was process.

That usually disappoints people expecting the name of a tool, but it’s the honest answer. AI increased my execution speed. Process increased my consistency — and consistency is what clients buy.

I work alone most of the time. Scrum, as written in the manual, assumes a team: a daily with several people, separate roles, an alignment ceremony. Alone, that turns into theater — a meeting with myself.

I adapted it. I kept what solves a real problem for someone working solo:

Scrum practice	How I adapted it for solo operation
Sprint	Closed weekly cycle, with scope frozen at the start
Prioritized backlog	Single queue by client impact, not by whatever I feel like that day
Daily	Short morning review: what’s blocked and what ships today
Definition of Done	Criteria written before starting — without it, “almost done” lasts forever
Velocity	Cards closed per week — the only metric I track
Retrospective	Half an hour on Friday: what slipped and why

The gain didn’t come from any single practice. It came from stopping deciding priorities every day. Open scope and a shifting queue burn energy nobody accounts for — and working alone, that energy is the scarcest resource there is.

Result: I’ve been closing roughly twice as many cards per week. Not because I work more hours. Because I stop less, redo less and finish what I start.

The combination that works is this: AI to execute faster, process to execute the right thing. The first half alone produces a lot of disposable work at high speed.

The honest job-market analysis: what rose, what collapsed

While I was building all this, the market reorganized itself in a way almost nobody predicted. I went after the numbers — and some of them are uncomfortable.

Why Scrum Master openings collapsed

Because the practice got absorbed and the separate job title stopped justifying itself.

In January 2023, Capital One cut more than a thousand positions of Scrum Master, agile coach and delivery lead in a single decision. Royal London let go of 90% of its agile coaches a few months later. In the UK, permanent Scrum Master openings fell from 79 to 30 in two years — a 62% drop.

Capital One’s own reasoning is the interesting part: the role was “critical in the early stages of the transformation”, but as the organization matured the natural step was to “integrate agile delivery processes directly into engineering”.

Translated: the method won, the job title lost. Scrum is still running in 87% of teams. What disappeared was the person hired exclusively to facilitate ceremonies.

There’s a second factor, and it’s harder. There are more than 4 million agile certificates issued worldwide. At Scrum.org, fewer than 1% of certificates reached the highest level. When the supply of credentials grows much faster than the supply of proven skill, the market stops being able to tell the two apart — and starts treating everyone as interchangeable.

And the Product Manager? That’s a different story

The product market recovered — it’s the highest volume of openings since 2022, with more than 7,300 positions open globally. But the recovery has an owner.

In an analysis of 654 real Product Manager job descriptions, 85% already mention AI or machine learning and **32% require experience with *evals*** — the tests that measure whether an AI system actually works, instead of just appearing to.

And here’s the number that caught my attention the most: about 30% of open PM roles are AI PM roles, but fewer than 5% of senior PMs have ever put an AI agent into working order.

That gap is the entire opportunity summed up in one line.

What this means in practice

The salary premium for AI skills jumped from 25% to 56% in one year — more than doubling. Demand for those skills grew 142% in twelve months, according to LinkedIn’s own data.

For anyone based in Brazil, the effect is direct: product professionals hired by international companies have been earning in the range of US$ 4,000 to US$ 12,000 per month, typically 2 to 4 times the local equivalent. Payments Engineer roles involving AI pay between US$ 5,000 and US$ 9,500 a month.

A caveat I have to make: these surveys are mostly from the UK and the United States. The direction carries over to Brazil, the exact magnitude doesn’t. Treat it as an order of magnitude, never as a guaranteed salary floor.

How I read all this

It isn’t “Scrum is dead” or “product is over”. It’s something else, and more specific:

The market stopped paying for process in isolation and started paying for process applied to a system that works.

A Scrum Master who only facilitates ceremonies became a cost. A Product Owner who only writes user stories and prioritizes a queue became a cost. What still holds value — and more of it — is whoever has method and understands the system being built.

That’s exactly why the earlier section of this article matters. I don’t use Scrum because I like frameworks. I use it because, without process, AI would just make me produce more garbage in less time. And I don’t sell AI because it’s fashionable. I sell it because I know what breaks underneath it.

The rare combination isn’t knowing AI or knowing process. It’s knowing both and carrying the scars of having done it in production.

Who this article is for

If you run an operation that needs payments infrastructure, automation or AI agents that work in production — and you’re tired of pretty demos that don’t survive the real flow — the conversation is straightforward.

If you build technology teams and you’re looking for someone who has built, broken and shut down their own project and can explain why, same.

I’d rather say “that’s still in testing” than sell certainty I don’t have. In systems that handle money, that’s the difference between a supplier and a problem.

Frequently asked questions

What is an AI agent in production, in practice?

It’s an AI system someone depends on to do their job — not a demo. The difference isn’t in the model, it’s in the ability to act: query real data, execute operations in the system, handle errors and leave an auditable trail. An agent that only talks is a prototype, even if it’s published at a public address.

Why do AI projects fail inside companies?

Most of the time, because of infrastructure, not the model. Disorganized data, legacy systems with no integration points, nonexistent authentication and unmapped processes will take down any agent, however good it is. That’s exactly how my first agent in production failed: the AI was right, the foundation underneath it wasn’t ready.

What is MCP (Model Context Protocol)?

It’s an open protocol that lets language models execute actions in external systems, rather than only generating text. With MCP, an agent queries databases, triggers processes and records information inside the company’s systems. It’s the standard being adopted by Anthropic, Google and Microsoft to connect AI to real tools.

Is Scrum worth using when you work alone?

Yes, as long as it’s adapted — applying the manual without a team turns into empty ceremony. What sustains the gain in a solo operation is the weekly cycle with closed scope, the single queue prioritized by impact and the definition of done set before starting. What you drop are the ceremonies that exist to align several people.

Are Scrum Master jobs disappearing?

Dedicated openings shrank a lot, but the practice didn’t. In the UK, permanent Scrum Master positions fell 62% in two years, and companies like Capital One and Royal London eliminated hundreds of agile roles in one go. At the same time, Scrum is still running in about 87% of teams: organizations absorbed the method into engineering instead of keeping the separate job title. The market stopped paying for facilitation in isolation and started paying for method applied to systems.

Does a Product Manager need to know AI in 2026?

In practice, yes. In an analysis of 654 real Product Manager openings, 85% mention AI or machine learning and 32% require experience with evals. About 30% of open PM roles are specifically AI PM roles, while fewer than 5% of senior PMs have shipped a working AI agent — and that gap is exactly what sustains the 56% salary premium for those who have the skill.

How long does it take to put an AI agent into production?

It depends almost entirely on the state of the systems it will operate, not on the agent. On a base with integrations ready and data organized, weeks. On a legacy system with no integration points, the real work is the foundation — and that usually takes months. Any timeline given before auditing the infrastructure is a guess.

---

Want to talk about your case? Reach me on WhatsApp or at [email protected].