25 years in build, one conclusion: value is moving from building to flying

Your AI is live. Who is flying it?

I spent twenty-five years selling build.

I co-founded and grew a digital consulting and services group: more than 1,000 people, platforms delivered on four continents, then a sale to a peer group. The model behind that growth fit in one simple equation: companies need to build more software than they know how to produce, and that scarcity is billed by the day.

That equation has just broken. And I would rather write it myself than pretend not to see it: AI is going to absorb most of the build capacity of companies. Not a marginal share, not just the thankless tasks: most of it. Whether we like it or not.

The shift is already measured

This is no longer a forecast. As early as April 2025, Satya Nadella said that 20 to 30% of the code in Microsoft's repositories was now written by AI, and the curve will not come back down. Autonomous development agents now chain together tasks that used to keep whole teams busy: specification, generation, testing, refactoring, documentation. Nearly nine organizations out of ten now use AI in at least one function.

Every CIO sees it inside their own walls: one augmented developer produces what a small team used to produce. A well-equipped business unit assembles in a few days what used to require a project. Software production, the bottleneck that justified forty years of IT services firms, fixed-price projects and time-and-materials contracts, is becoming abundant.

And economics never forgives abundance: what becomes abundant stops being expensive, and stops being strategic. Build is going the way compute and storage went before it. It does not disappear. It becomes a commodity.

The paradox nobody owns

You might think this abundance mechanically produces value. The numbers say the opposite, with rare brutality.

Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025; the same firm expects more than 40% of agentic AI projects to be cancelled by the end of 2027. MIT, across more than 300 initiatives analyzed, finds that only about 5% of pilots deliver rapid value acceleration. And McKinsey sums it up in one sentence: nine organizations out of ten use AI, 37% see an impact on their bottom line.

Companies have never built so much. Never have so few of those builds produced value.

This paradox has an explanation the market still looks at too little: we have industrialized production, not operation. Everything, budgets, skills, partners, metrics, is still organized around the act of building, at the precise moment when building stops being the problem. The platforms are there. Nobody knows how to fly them.

The rush, then the piloting

Look at what has happened in your company over the past three years. An AI feature switched itself on in the office suite with an update, without anyone asking for it. An agent was put together in three days by a business unit. An assistant slipped into the CRM, another into the ticketing tool, a third into the ERP. As early as 2024, three knowledge workers out of four were using AI at work, most of them with their own tools. Each of these tools does something. Nobody answers for it.

This is not a fault, it is a rush, and every technology that mattered has had its own. But the rush is coming to an end, and what comes next is inevitable: AI under human oversight, inventoried, measured, with someone accountable for it. This is no longer just my conviction, it is the law: since 2 August 2026, telling people they are dealing with a machine is no longer a courtesy, it is an obligation (Article 50 of the AI Act), and human oversight of risky systems falls on the company that deploys them, not only on the vendor that sells them (Articles 14 and 26).

What aviation understood before us

When automation entered cockpits, some predicted the end of pilots. Exactly the opposite happened: the job moved. The modern pilot barely touches the controls; they design a flight plan, monitor instruments, run procedures, arbitrate exceptions. And their responsibility has never been greater, precisely because the system has become too complex to be flown "by hand".

Aviation turned automation into safety because, alongside the autopilot, it invented the discipline of supervising it: standardized instruments, checklists, alert thresholds, systematic incident review, a culture of measured reliability.

That is exactly what the augmented information system is missing. AI has given us the autopilot of build. Nobody has built the cockpit yet.

The CIO, architect and pilot

So here is my conviction, the one that twenty-five years on the builders' side allow me to state: the CIO who wins the decade will no longer be a commissioner of build. They will be the architect and the pilot of a cockpit.

Architect, first: when production becomes abundant, what becomes rare is the overall design. Deciding what should exist, what should connect, what should remain human. Architecture stops being a project phase and becomes the permanent function that keeps abundance from producing chaos, because a system where anyone can build is a system that drifts by default.

Pilot, second. A cockpit worthy of the name continuously displays five dials, and I know few executive committees that can read them today:

1. Reliability. Which pipelines, which agents, which reports can be trusted, measured continuously, not discovered after an incident by users.
2. Costs. The real consumption of capacity and models, managed like fuel, not discovered on the invoice.
3. AI Act evidence. The register of systems, their risk class, the applicable obligations, the named owners and the dated evidence, documented as you go, not reconstructed under the pressure of an inspection.
4. Usage. The real use of what has been built, by whom, and what it brings in, because an unused asset is a pure loss that abundance makes painless, and therefore invisible.
5. Reference sources. The company knowledge that feeds the systems, maintained as an asset, because a model fed with outdated knowledge produces errors with perfect confidence.

And as in a cockpit, the human no longer steps in everywhere: they step in where the system's confidence is insufficient. It is the reversal of the service delivery model I practiced my whole career: no longer people assisted by tools, but systems supervised by experts, triggered by thresholds. Experts who know the platform from the inside, and who have sold you nothing else.

What this shifts, what this creates

Let's be honest about the consequences, including for my former trade.

The integrator model, selling person-days to build, will not survive abundance as it is. But it is not the trade that becomes a commodity, it is one of its tasks. Expertise, on the other hand, becomes rarer and more valuable: when anyone can have an agent produced in three days, value moves to those who can say whether it is trustworthy, put it to the test, keep it running in production, and answer for it. The senior consultant who has delivered ten years of Salesforce, SAP or Microsoft does not disappear; they change seats, from the construction site to the cockpit, and their value goes up, like that of everyone in the field who knows how to fly what others have built. That is what we are building 56North around: the cockpit itself, already delivered; partners for human testing and operations, being brought together; and a network of senior experts drawn from the practitioner communities of each platform. They are the ones who will fly.

One condition only, which aviation set before us: the manufacturer does not certify its own aircraft. The integrator's conversion to "AI managed services" runs into this conflict: you cannot be judge and party, calmly measuring the reliability of what you built and billed yourself. Building and controlling must remain two separate trades, even when the same experts know how to do both.

Three functions will therefore concentrate value: those who design (the architects of the augmented system), those who fly (the cockpit operators, with their instruments and thresholds), and those who measure (the independent third parties able to say, figures in hand, whether a platform flies or only pretends to). Electricity created electrical engineering, the cloud created cloud operations, cybersecurity created the SOC. The discipline of operating intelligence does not yet have an established name. It will have its standards, its scores, its operations centers, and its pioneers.

The danger, the advice

The danger	The advice
AI features switched themselves on in your tools with updates. Nobody knows how many there are, or who answers for them.	Build the register of every AI system in service, including those shipped by your vendors, each with a named owner.
An agent's reliability is discovered when a user complains.	Measure answer quality continuously, on real conversations, with alert thresholds written down in advance.
Model costs are discovered on the invoice, at the end of the month.	Track consumption per system and per use, and set a budget per agent, the way you manage fuel.
The evidence the AI Act requires will be pieced together in a hurry on the day of an inspection.	Document as you go: risk class, applicable obligations, dated evidence, recorded along the way rather than after the fact.
Whoever built your AI systems is also the one telling you they work.	Keep building and controlling apart: hand the measurement to a third party that has sold you nothing else.

The call

To the CIOs reading this, and to the executives who will one day sign the register of their AI systems, one question only, the one I would ask my own executive committee: your AI is deployed. Who operates it? Not who built it. Who flies it, with which instruments, which thresholds, and which dated evidence an auditor will accept. If the answer is silence, or "our integrator, I suppose", you now know your project for the decade.

The augmented build is not a threat to CIOs. It is their promotion, provided they agree to change seats: leaving that of the client who buys construction, taking that of the captain who answers for the trajectory.

I spent twenty-five years selling construction. I will spend the next ones building the cockpit.

Pascal Mennesson, co-founder of Maltem Consulting Group, founder of 56North. If you are an executive and would like to discuss this, request a first conversation.