In the data teams that have made the switch, almost all the code is now written by AI. And yet a delivery, a dashboard, a metric or a simple feature, still takes weeks, often months, to reach the people who asked for it.
The problem isn't the tools. A data platform is an industrial line: ingest, model, analyse, report. AI has been installed at every station, and every station runs ten times faster. Between the stations, though, nothing has changed. Humans still pass the work along, with a ticket, a spec, a review, a sign-off. Those handoffs take exactly as long as they did before.
Sprints and agile were designed to coordinate humans who write code. When the code takes minutes, the two-week sprint becomes the main thing people wait for.
The conclusion is uncomfortable. On the line as it's organised today, the human placed at the handoff is what slows everything down. That is why AI's productivity gains aren't showing up yet. The answer is not to remove humans. It's to take them off the line and put them above it, where they decide, explain and sign, while agents keep the work moving.
The data job market will be redrawn in three phases. First a transfer: large teams will need fewer people, and those skills will go to smaller companies, which get access to a real data platform for the first time and need one to get anything out of AI. Then a delegation to broader IT roles, as agents become more autonomous. Then full automation. There won't be more data jobs. There will be data platforms everywhere.
Code is no longer the bottleneck
For twenty years, data work was organised around one scarce thing: the time of someone who could write a query, a pipeline, a model, a dashboard. Teams, roles, methods and schedules all followed from that.
That scarcity vanished in two years. A transformation, an ingestion, a quality test, a dashboard, an interface: an agent designs and builds them in minutes, often better documented than hand-written work.
Lead times should have collapsed. Sometimes they did. Most of the time they barely moved. Ask any business owner how long they wait for a new metric. The answer is the same as in 2020. In a mid-sized or large company, a dashboard still takes three months to go live, when the query behind it sometimes fits in twenty lines.
If the code runs ten times faster and the result arrives barely sooner, the code was never the problem.
The line and its handoffs
A data platform works like a production line. A request goes in one end; out the other comes a delivery: a dashboard, a reliable metric, a flow to another system, or simply a new feature. In between, four stations: ingest the source, model it, analyse it, report it.
Each station belongs to a different trade, data engineer, analytics engineer, analyst, BI developer. And between any two stations there is a handoff. Work sits in a backlog, a spec goes round, someone reviews, someone else signs off. A handoff needs a human who is available, who understands what they've been given, and who decides to start now rather than next week.
Industry has known this problem for forty years. Goldratt wrote it down in The Goal: speeding up a station that isn't the bottleneck doesn't speed up the line. Figure 1 shows it on a typical request.
The journey of one metric request
That's the paradox. Teams really have become more productive, station by station. But the lead time the business sees is made of handoffs, and handoffs are made of calendars, meetings, queues and “I'll look at it next week”.
Time is no longer lost in the work. It's lost between hands.
The sprint is a human tempo
The Agile Manifesto dates from 2001. It was written for human developers who needed to coordinate, cut work into digestible pieces and check in often. The two-week sprint answered a very real constraint: a human takes days to build a feature.
That constraint is gone. A task that takes an agent twenty minutes still waits, on average, a week for a sprint to start, then for the end-of-sprint review before it ships. A sprint batches work, and a batch makes everything that arrives between two departures wait.
Agile isn't to blame. It was designed for a world where work was slow and coordination expensive. In a world where work is almost instant, the cadence itself becomes the latency.
Same stream of requests, two organisations
An electric motor in a steam factory
This paradox has a precedent. By 1987, computers had taken over offices, and yet company productivity was growing no faster than before. The economist Robert Solow summed it up in a line that stuck: the computer age shows up everywhere except in the productivity figures. The gains took another decade to appear.
A few years later, the economic historian Paul David found the explanation in electricity. The first electrified factories swapped their central steam engine for one large electric motor and kept the same line shafts and the same floor plan. The gains were tiny. The break came when factories were redrawn around a small motor on every machine. That took a generation.
- 1880The central steam engineOne engine, line shafts, belts. The factory is designed around it.
- 1900The electric motor replaces itSame floor plan, same shafts, same belts. The gains are minimal.
- 1920One small motor per machineThe factory is redesigned around the machines. Productivity finally takes off.
- 2026AI inside the existing organisationSame roles, same tickets, same sprints. The factory plan is still to be redrawn.
Data teams are living through that moment. AI has been plugged into the existing organisation, with the same roles, the same tickets, the same sprints, the same committees. An electric motor in a steam factory.
AI's gains won't come from more powerful tools. They'll come from a different floor plan.
Taking humans off the line
On the line as it exists, humans slow everything down. Not because they work badly. Because they're in the wrong place. We ask them to carry work from one station to the next, which an agent does without sleeping, without meetings and without a queue.
Removing humans is not an option. A platform without humans produces numbers nobody understands and nobody stands behind. What has to change is their position. Humans leave the line and sit above it.
In practice, agents keep the work moving around the clock. They ingest, model, test, document, monitor, fix. Humans step in where nobody can replace them. They decide what matters, explain what the numbers don't say, sign before anything touches production, and give the context that makes the agent better.
From a chain of handoffs to a loop
This shift solves the handoff problem. No more passing work between stations, so no more waiting between stations. What remains is one human touchpoint per decision, at the moment it counts. And that touchpoint gets more interesting. Instead of reviewing a pipeline, you explain why margin fell in Roubaix this quarter.
Agents report the facts. Humans provide the context. It's a loop, not a chain.
Data jobs, redrawn
If the handoffs disappear, the jobs that lived off them disappear too, in their current form. Better to say it calmly, because it's already happening. Execution roles are getting scarcer. Judgement roles are taking their place, for a while.
A mid-sized company's data team today has ten to twenty people, organised by station. Tomorrow, a few platform pilots will be enough to run the whole thing, with business teams far more involved in defining and explaining the numbers.
For data people, this isn't necessarily bad news. It is bad news for the roles that consisted of carrying work from one place to another. People who can judge, arbitrate and explain will become rarer, and therefore more valuable. And the least interesting work in the trade, the part everyone avoided, is the first to go.
The data platform moves down to smaller companies
You could read everything above as bad news: fewer people on data teams. That's half the story.
Until now, a data platform worth the name was a luxury for large groups. It took ten to twenty people to build and run one. An SME, a mid-cap or a freshly funded scale-up couldn't afford it. It lived on Excel exports, a BI tool plugged straight into the ERP, and one person who did a bit of everything (#swissarmyknife).
Two things are changing at once. When agents run the line, the cost of a platform collapses, because you no longer need a team, you need a pilot. And these companies suddenly need one. They want AI to work for them the way it does for large groups, and an AI plugged into scattered, unverified data with no shared definitions produces wrong answers with great confidence. No reliable data, no useful AI.
Large teams will shrink. The skills they release will equip tens of thousands of companies that never had a data platform.
The transfer: fewer data people in the large ones, far more companies equipped
This isn't about creating jobs. In the first phase, it's a transfer. There will be less need for specialists and hyper-specialisation, fewer data engineers per large group, fewer roles dedicated to a single step of the line. But large companies number in the tens of thousands and small ones in the millions, so the total holds. One data person for every seven small companies is already two hundred thousand people across Europe, more than double what will remain in the large ones.
The transfer will take several forms. Data engineers and analysts will become platform pilots, in-house at a fifty-person company or freelance for several companies at once. Consultancies will sell the operation of agent-run platforms rather than billable days.
That's only the first phase. In every sector that has already been through automation, what follows is well known.
- Phase 1 · The transferData jobs leave large teams and go to equip SMEs. The total stays roughly flat. The jobs change: fewer specialists, more pilots.
- Phase 2 · The delegationDomain agents gain autonomy. Running the platform stops being a trade of its own and becomes one responsibility among others in a broader IT management role, the way networks, security or workstations are managed today.
- Phase 3 · Full automationThe platform runs itself. All that's left to humans is governance: deciding what matters, signing what goes to production, explaining what the numbers don't say. With no dedicated data role.
Dedicated data jobs, in three phases
What this movement leaves behind isn't jobs. It's platforms. Tens of thousands of companies equipped with a state-of-the-art data platform, where only a few hundred large groups had one. That is where AI will really spread through the economy, into the mid-sized companies that until now had neither the data nor the team to use it.
There won't be more data jobs. There will be data platforms everywhere.
Domain agents, BI rebuilt
Everything above describes today's agents. They're young, but the direction is no longer in doubt: whatever happens, data work will run through agents. What remains is to say what that changes for the people who use them.
Domain agents, fed with context. Today's agents are generalists you hand instructions to. Tomorrow's carry the know-how of one specific trade and know the situations that come up on a platform. Their raw material is context: definitions, business rules, past decisions, reference data. Context engineering replaces prompt engineering, and the memory of what has been decided becomes a company asset. Their autonomy grows as that context thickens. They handle known situations on their own and escalate only real decisions to humans.
A platform nobody opens. Today, a data platform is read through screens. Tomorrow, it will mostly be queried by agents: business agents, assistants and in-house agents, will ask the platform's agents directly for a definition or a certified number, agent to agent. Nobody will open the platform any more; it will become a back-office layer, managed in the background the way the network is today.
BI, rebuilt. The static dashboard, built by hand and maintained by hand, will not survive this floor plan. Hex showed what comes next first: a workspace where you ask a question, an agent writes the query and the code, leans on the company's shared definitions so it doesn't reinvent them, produces the chart and the explanation, and a human reviews, comments and publishes. In that model, dashboards become agent outputs, regenerated on demand, versioned and certified. The analyst no longer clicks around a tool; they talk to a production line and sign what comes out of it.
Interfaces for three gestures. Today we talk to the machine through tickets, threads and dashboards. Tomorrow the interface is built for the three gestures left to humans: sign, question, add context. A morning report that shows only what's waiting for a decision. A signature in one gesture. An explanation dictated where the number appears, at the moment you remember it. The handoffs that remain, the ones where a human has to decide, go from days to minutes.
Agents become the operators. Humans decide, explain and sign.
What opens up
This is the moment to step back. Not to defend positions. To decide where humans keep their impact on a line that runs without them. That impact sits upstream, in choosing what matters and what a metric means, and downstream, in the explanation nobody else can give. It also sits in trust. A line run by agents is more traceable, more tested and better documented than any human team has ever been.
There is also a less visible effect. A platform run by humans accumulates entropy: tables nobody reads any more, duplicate pipelines, dashboards nobody opens, and a cloud bill that grows with them. Agents that run continuously spot these leftovers and remove them as they go, instead of letting them pile up. The platform stays clean, and cloud costs fall with its entropy.
And look at what opens up. For twenty years, data teams lived saturated, deciding what they wouldn't do. When the line runs on its own, the question flips.
Agent autonomy takes nothing from humans. It gives them back the time to do what the line had stopped doing.
What people will say against this
“We can't trust an AI with our numbers.”
Trust isn't given to an AI, it is given to the guardrails around it: limited rights, trials outside production, tests, traces, a budget. Those guardrails are what make every action verifiable, and they are what you tune. Flow by flow, you choose to let the agent act alone or to keep fine-grained control, signature included, and you loosen it as trust builds. What works is the right guardrails, with adjustable autonomy.
“Handoffs are also how we keep control.”
Partly. But a control that arrives three weeks after the work doesn't control much. Automated tests at every step and a signature at the right moment control better than a review at the end of a sprint.
“Our organisation isn't ready.”
None is. That's what this manifesto is about. It isn't about switching everything at once. You start with a few non-critical flows, make them autonomous, measure the lead time, then extend.
“Every forecast says more data jobs, not fewer.”
True. The US Bureau of Labor Statistics projects 35% more data scientists by 2035, the World Economic Forum 41% more analysts by 2030, and the European Commission more than ten million data professionals in 2030 against 8.2 million today. These forecasts extend the past decade, when every new use case called for new hands. They say nothing about what happens when the hands are agents. Phase 1 as described here is in fact compatible with them: the total holds. It is afterwards that they go wrong.
“What about security and sensitive data?”
Well-designed agents are more traceable than humans. Every action is logged, every right is explicit, every AI model can be chosen according to company policy. But it has to be built that way from the start.
Predictions to revisit
Dated predictions, on a short horizon. To be checked by 2029.
- late 2026The first mid-sized companies run their data platform with fewer than five people, with no loss of quality.
- 2027The lead time between a business request and a reliable delivery, dashboard, metric or feature, becomes the main metric of data teams, ahead of the number of dashboards delivered.
- 2027Most questions asked of company data go through a business agent that queries the platform's agents, agent to agent, rather than through a dashboard.
- 2027A fifty-person company has a data platform as solid as a large group's in 2020, run by a single pilot.
- 2027Domain agents handle most incidents and schema changes on their own; humans see only the decisions.
- 2027Execution-level data job postings fall sharply; pilot, architect and governance roles grow.
- 2028The sprint has disappeared from data teams, replaced by continuous flow punctuated by human signatures.
- 2028Legacy BI tools have either built in agents or lost their place; the dashboard is a regenerated, versioned, certified output.
- 2028The data platform is no longer a product people open but a back-office layer: business teams talk to their agents, which talk to the platform's agents.
- 2028Running the data platform is no longer a trade of its own: it's an IT management responsibility, like networks or security.
- 2029The first fully autonomous data platforms run with no dedicated data role; humans keep only governance.
- Eliyahu Goldratt, The Goal (1984). The theory of constraints: speeding up a non-bottleneck station doesn't speed up the line.
- Robert Solow, “We'd better watch out”, New York Times Book Review (1987). The productivity paradox.
- Paul David, “The Dynamo and the Computer”, American Economic Review (1990). Why electricity took a generation to deliver its gains.
- European Commission, Annual Report on European SMEs 2022/2023 (enterprises and employment by size class, EU-27, 2022); Eurostat, structural business statistics 2023 and ICT survey 2025 (in-house data analytics: 28% of small enterprises, 79% of large ones).
- Bureau of Labor Statistics, Occupational Outlook Handbook, data scientists (275,600 US jobs in 2025, +35% by 2035); World Economic Forum, Future of Jobs Report 2025.
- European Commission / IDC, European Data Market Study 2024–2026, first report (8.2 million data professionals in the EU in 2024, broad definition).
- Agile Manifesto (2001), agilemanifesto.org.