We are All Platform Engineers Now

02 Sep, 2026

The pitch you already heard

Somewhere around 2019, platform engineering arrived with a slide deck and a promise. Stop making every team reinvent deployment. Build an internal product. Golden paths. Reduce cognitive load. Treat your developers as customers, because customers are people you are contractually obliged to be nice to.

The promise was real. The delivery was often a Backstage instance nobody updated, a wiki page titled "Getting Started" last edited by someone who left in 2022, and three engineers who became the permanent on-call rota for everyone else's YAML. The cognitive load did not disappear. It was moved into a queue and given a service level objective.

Hold that thought, because it is about to happen to your entire company.

The new math

Two forces are colliding, and neither of them cares about your roadmap.

The first is the business pressure that has been building since interest rates stopped being free. Demand for internal services keeps rising. Headcount does not. The only remaining lever that leadership can pull without a difficult conversation is the phrase "let's make that self-service." Every function is now asked to serve more requesters with the same or fewer people, which mathematically requires the requesters to do more of the work themselves.

The second force is that building the thing that lets them do it themselves just got absurdly cheap. A finance analyst who has never opened a terminal can now produce a working intake portal, a natural language query layer over the expense data, and an agent that answers 80 percent of the questions she used to answer by hand. In a weekend. For roughly the price of a team lunch.

The old objection to self-service was "we cannot afford to build that." That objection is dead. What arrives in its place is not paradise. It is a bill that shows up later, in a different currency.

Self-service is a cost transfer, not a cost cut

This is the part that never makes the deck. Self-service does not eliminate work. It relocates it, and it splits it into two piles.

The first pile goes to the requester. They now fill in the form, pick the correct option from a dropdown they do not understand, and discover that their situation is the one edge case the form does not cover. Their time is not on your budget, so it is free. This is the same accounting trick that made self-checkout profitable.

The second pile goes to you, forever. You have stopped doing the task and started operating the system that does the task. Doing is bounded. Operating is not. Nobody has ever been paged at two in the morning by a spreadsheet.

What the org chart looks like in eighteen months

The functions that were previously defined by their expertise start to be defined by their platform. The work drifts, meeting by meeting, from the thing to the thing that does the thing.

• HR ships an onboarding agent, then discovers it needs a content lifecycle, a policy owner per jurisdiction, and an answer to why it told a new hire in Germany about a US benefit.
• Finance ships a self-service reporting layer, then spends the next year explaining that two numbers differ because two people defined "active customer" differently in 2021.
• Legal ships a contract review assistant, then invents an approval tier system, because the difference between "advice" and "output" is the entire profession.
• Procurement ships a vendor intake flow that becomes, unavoidably, the company's actual governance mechanism, since it is the only door with a lock on it.
• Marketing ops ships six things, four of which are the same thing, none of which know about each other.

Each of these teams now has, whether or not anyone says the words out loud, a product backlog, an adoption problem, a support channel, a documentation debt, a permissions model, an audit obligation, and a deprecation dilemma. Congratulations. That is a platform team. The badge says Compensation Specialist but the job is Site Reliability Engineering for the compensation platform.

The maintenance iceberg

Creation cost fell off a cliff. Maintenance cost did not move at all, and in some places it went up.

Building an internal tool with AI is now a weekend. Owning it is a career. The parts that were always expensive are still expensive and are now more numerous, because there are ten times as many tools.

• Somebody has to decide what the correct answer is, which means somebody has to have an opinion, in writing, that they can defend.
• Somebody has to notice when the correct answer changes and update the thing that was confidently giving the old one.
• Somebody has to handle the requests the platform refuses, which are always the requests that matter most, because easy things do not escalate.
• Somebody has to write the evaluations, because "it seemed fine when I tried it" is not a quality process, it is a vibe.
• Somebody has to say no. This is the single most important platform skill and the one no function has hired for.

The abstraction leaks, as abstractions do. The difference now is that when it leaks, the person holding the bucket is a benefits administrator debugging a retrieval pipeline she did not build, in a language she does not speak, for a colleague who is copying her manager.

Everything platform engineering learned, now on a loop

The lovely thing about a lesson is that it can be learned repeatedly by different people at full price. Here is the discount version, offered in the certain knowledge that it will be ignored.

1. A platform nobody adopts is not a platform, it is a hobby with a login page.
2. If adoption is mandatory, you have not built a golden path, you have built a toll booth, and people will find the side road.
3. Internal demand is infinite when the price is zero. You will need some form of prioritization that is not "whoever is loudest in Slack."
4. Ship with a deprecation policy or inherit an estate. Version one of anything internal outlives the person who built it and the reason it existed.
5. Dedicated funding, or nothing. A platform maintained in someone's spare 20 percent will be reliable roughly 20 percent of the time.
6. Documentation is the product. The tool is a delivery mechanism for a decision somebody made once.
7. Measure the thing you promised. If you claimed to cut cycle time, produce the cycle time, not a screenshot of the dashboard.

The re-centralization speedrun

Here is the arc, compressed. Central function is a bottleneck, so work is decentralized to the edges where AI has made building cheap. The edges produce forty overlapping platforms with inconsistent data definitions and nine different opinions about who is allowed to see salary bands. This is discovered during an audit, or worse, during a board meeting.

A Center of Excellence is announced. It exists to standardize, consolidate, and provide shared building blocks so that teams do not have to reinvent deployment.

You have arrived back at platform engineering, having taken the scenic route, with more tooling and less institutional memory. The main difference is that the sprawl was produced in six months rather than six years, which is what everyone means when they say AI accelerates the business.

Who actually comes out ahead

Not the people who are fastest at building. That skill is now a commodity and its price is falling in public.

The winners are the people who know what the right answer is and can encode it. Judgment does not automate, and it turns out judgment was most of the job all along, wrapped in enough manual labor that nobody had to look at it directly. When the manual labor is stripped out, what remains is exposed and measurable. Some people will look magnificent in that light. Others were the manual labor.

The other winners are the ones who take ownership seriously. Every organization is about to be full of unclaimed automated systems making consequential decisions on the authority of whoever ran the prompt. Being the person who put their name on one, wrote down its limits, and monitored whether it still worked will be worth an unreasonable amount, mostly because so few people will do it.

The part nobody puts on the slide

Self-service was never about service. It was about who absorbs the friction. AI has not removed the friction, it has made it cheap enough to redistribute at scale, and the direction of travel is outward and downward, toward whoever has the least leverage to refuse.

So yes. We are all platform engineers now. Not because we chose the discipline, but because the discipline is what is left when your function stops doing tasks and starts running a service with users, an uptime expectation, and no budget line.

The on-call rota is over there. Bring a runbook. Nobody has written one.