AI · MBA

A pale project timeline broken mid-span on a dark technical poster field

This is not an IT project

A project structure carries an end date and a closed scope. Value from AI appears only in the continuous adaptation of a process, so the project finishes roughly in the week the work properly begins, and it leaves a working system with no owner. The decision is made before the start, not after the pilot. Either you name the person who on day 400 answers for that result out of their own profit and loss account (P&L), or you accept that on day 400 no such person will exist. There is no third option.

Project mode and product mode

Start with a distinction that most companies never put on paper, though it decides the outcome. Sriram Narayan set it out in Products over Projects (martinfowler.com, 2018) as the difference between project mode and product mode, and broke it into six dimensions. Each of those dimensions behaves differently after the day of handover.

DimensionProject modeProduct mode
Fundingscope of the solution fixed up frontstable team, rolling budget
Life-cycle spanbuild onlyidea, build and run
Organisationseparate build and run departmentsone team from idea to operation
Team longevityweeks or monthsyears
Prioritisationproject portfolio officeempowered product owner
Measure of successon time and on budgetimprovement in a business outcome

The last row is where the whole thing breaks. A project counts as successful when it delivered the declared scope on the declared date for the declared money. Narayan notes that project mode usually has no mechanism at all for verifying actual benefits after delivery. Bad faith plays no part in it. There is nobody to do the checking and no budget to do it from, because the organisational unit created for the work stopped existing on the day the acceptance certificate was signed.

Mik Kersten described the other half of this arrangement in Project to Product (IT Revolution, 2018). Project management funds milestones against a scope defined up front, so every new need requires a new project to be chartered. It looks like administrative friction. It is in fact a throttle setting on adaptation: if changing a process requires a fresh budget request, the system learns at the speed of an investment committee that meets once a quarter. Kersten builds this on client experience rather than a controlled study. I therefore treat his frame as a model, not as evidence. A model is enough, because everything that follows can be derived from first principles.

Conway's law and the disbandment date

Melvin E. Conway put his observation into print in How Do Committees Invent? (Datamation, April 1968). It reads: any organisation that designs a system will produce a design whose structure is a copy of that organisation's communication structure. It is usually quoted in microservice arguments. Yet Conway's law answers directly why an AI initiative dies inside a project structure, and it does so without a single failure statistic.

Take it apart. A project team is a communication organisation with one unusual property: its founding document contains the date on which it will be dissolved. If an organisation imprints its shape on the system, then a system built by an organisation with a dissolution date inherits that date. It does not inherit that in code. It inherits it in who understands why the code looks the way it does, and in who has the authority and the money to change it. Once the team is disbanded, no organisation exists whose shape matches the shape of the system. That is what the ownerlessness consists of.

Conway himself quoted a corollary he attributed to Brooks: since the first design is almost never the best one, the system concept will have to change anyway. Organisational flexibility is therefore a condition of effective design. It is the organisation that has to be flexible, not the architecture. A project organisation is flexible up to the day of handover and not one day longer.

Forty years later somebody measured it. Nagappan and co-authors presented The Influence of Organizational Structure on Software Quality at ICSE 2008. The material covered 3,404 Windows Vista binaries and over 50 million lines of code worked on by several thousand engineers. The authors built eight metrics describing the organisation alone, without looking into the code, and tested how well they predicted a component's proneness to failure.

Predictive modelPrecisionRecall
Organisational structure86.2%84.0%
Test coverage83.8%54.4%
Code churn78.6%79.9%
Code complexity79.3%66.0%
Dependencies74.4%69.9%
Pre-release bugs73.8%62.9%

Organisational metrics beat every classical code metric. For our subject the interesting one is the second of them, which the authors call NOEE: the number of engineers who worked on a component and left the company before the release date. Their rationale is explicitly about knowledge transfer — whoever inherits the component knows neither the reasoning behind the design decisions nor the reason for particular fixes. Somebody priced the dispersal of a team and showed that it is visible in defects. And in a project, the dispersal of the team is scheduled from the first day.

Two caveats, because without them the number claims too much. This is one product in one company, a predictive model on historical data rather than a causal experiment, and the authors themselves flag limited external validity. Conway's law has not been proved causally by this. My claim is smaller and large enough: where somebody took the trouble to measure, the shape of the organisation turned out to be a better predictor of system quality than the system itself.

Four things that die at the project boundary

Now the specifics. The day of handover cuts four things at once, and every one of them is needed only after that day.

The owner goes missing. What disappears is not a job title but accountability for the outcome. The model passes into the care of IT operations, which answers for service availability, not for whether the customer service team actually shortened its response time. Nobody has it in their annual objectives. Nobody loses a bonus when quality falls from 91% to 74%, because nobody reports that number to any governing body. The project sponsor, who could have asked, already has the next project in the calendar.

The budget goes missing, and it is the one that matters. In 2015 Sculley's team at Google presented Hidden Technical Debt in Machine Learning Systems at NIPS. The most reproduced figure in that paper shows machine learning (ML) code as a small box inside a real-world ML system. Everything else is taken up by data collection, feature extraction, serving infrastructure, configuration management, monitoring, analysis tools and resource management. The authors also name the CACE principle, changing anything changes everything: in a learning system every change to an input changes the behaviour of the whole, because the model has no modular boundaries to respect. Put two things side by side. Technical debt is by definition not repaid during the build; it is repaid during operation. The project funds the build and ends before the first instalment falls due, and CACE says that in this system there are no cheap small fixes made in passing. Post-warranty care performed "within routine maintenance" is, in ML systems, an accounting fiction. The maintenance contract signed alongside the project usually covers service availability and incident repair, not retraining the model on fresh data or revising a decision threshold after a regulation changes. Neither line item is in it.

The loop that learns from mistakes goes missing. A model put into production does not stand still, because the world underneath it moves. Vela and co-authors measured this in Temporal quality degradation in AI models (Scientific Reports, 2022). They took 32 datasets from four domains: weather, hospital operations, flight delays and stock prices. Four standard algorithms gave 128 model–dataset combinations on that material. Temporal degradation appeared in 91% of them. The shape of the phenomenon matters more than the percentage: a model could hold its quality for a long time and then collapse abruptly. A flat chart across a year is no proof that things are fine. The authors treat this ageing as a phenomenon distinct from data drift and recommend continuous error measurement with an automatic alarm once a threshold is crossed. They add that unreflective continuous retraining carries costs of its own. They studied classical predictive models, not large language models, so what I carry over to generative models is the mechanism, not the figure of 91%. That is enough for the organisational conclusion: somebody has to be looking at that chart on day 400, and the team was disbanded on day 120. The learning loop also needs somebody who reviews, every month, the cases where the system got it wrong, and who, on that basis, may change both the model and the procedure around it. A project buys no such person.

The mandate to change the process goes missing. This is the deepest of the four losses and the one least often named. A project manager receives a mandate over a tool: deliver the system, train the users, close the budget. He does not receive a mandate to abolish an approval step, push a decision one level down, or tell the sales director that his quoting process no longer makes sense in its current shape. The RAND report The Root Causes of Failure for Artificial Intelligence Projects (2024) puts first on its list the divergence between what the project optimises for and what the organisation actually needs. That is precisely the mandate gap, described from the technical team's side. The fourth cause on their list is more literal still: under-investment in deployment infrastructure, which leaves the pilot with nowhere to live.

That report deserves a methodological note, because a number circulates around it that does not mean what everyone thinks it means. RAND opens with the sentence that by some estimates more than 80% of AI projects fail, twice the rate of projects without AI. That is somebody else's estimate, introduced with the phrase "by some estimates", and not a figure RAND computed. The authors' own contribution is a qualitative analysis of 65 interviews with practitioners who had at least five years of experience building models. A purposive sample, no statistical representativeness, zero projects counted. Anyone who writes "RAND found that 80% of AI projects fail" is quoting gossip from an opening paragraph.

What the data says about who succeeds

Two surveys enter here, on condition that we do not mistake them for an audit.

Gartner surveyed 782 infrastructure and operations (I&O) leaders in November and December 2025 and published the results on 7 April 2026. As reported, 28% of AI use cases in that area fully succeed and 20% end in failure. The unit of analysis is a single use case in IT infrastructure and operations, not a project and not a company, and the ROI assessment is the respondent's own self-report. What interests me here is not the percentage but the attribution. Among those who succeeded, success is attributed to integrating AI with existing workflows and systems, and to full backing from the business side. Not to model quality. Gartner turns this into a recommendation: join AI to the business at the execution stage, not at planning alone. That sentence describes the difference between a project mandate and a process owner's mandate.

McKinsey's 2025 edition of The State of AI asked 1,993 respondents in roughly 105 countries. The result that matters for this thesis: 21% of organisations using generative AI say they have fundamentally redesigned at least some workflows. Four-fifths, then, are laying AI over the process they found. Of some twenty-five organisational practices examined, it is fundamental process redesign that shows the strongest relationship with AI's impact on operating profit (EBIT). I treat that as a strongly suggestive hypothesis, not as proof. It is a correlation drawn from self-reports, and companies able to redesign their own process may simply be better managed at everything at once. With that caveat the picture is consistent with the derivation: the largest identified lever sits in territory where a project manager has no authority.

I am not, however, building this essay on any failure rate, and not out of rhetorical caution. Magne Jørgensen and Kjetil Moløkken-Østvold took apart the Standish Group's CHAOS reports, the most-cited statistic on IT project failure. They found an undisclosed research design and unspecified sampling criteria. The definitions of success turned out to be unstable, so those numbers are unfit for comparison. The same doubt applies to the fresher numbers circulating around AI. An argument about structure has to stand on a mechanism rather than an indicator, because an indicator can be demolished in a single paragraph.

The decisive test: day 400

Now the procedure. Before the start you want to know whether what you have in front of you is a project or a change to the operating model. Do not ask about technology or the size of the budget; both answer a different question. Ask three questions about day 400, counting from the start — a moment long after the project has closed. That date falls outside the warranty period and outside the sponsor's bonus cycle, which is to say in the place where the organisation has already forgotten why it began.

Question one: who owns this on day 400? Not the sponsor and not the IT-side custodian of the system. It has to be a person with a first and last name who answers for whether the thing still produces a result. If the answer is "operations", "the data team" or "we will settle that after go-live", you have the answer to the whole test.

Question two: whose P&L does this result sit in? An owner without their own P&L is a reviewer, not an owner. The benefit from AI almost always shows up in a business unit's accounts and the cost in IT's, and that asymmetry manufactures ownerlessness all by itself. If you cannot point to the line where benefit and cost meet in one person, you have not built ownership.

Question three: what happens on the day the model degrades? The answer must contain a numeric threshold, the name of the person the alarm goes to, and the budget that pays for retraining or withdrawal. Missing any one of the three means the customer will be the one who notices the degradation.

Notice what the test does not contain. There is no question about model quality or about choice of vendor. I write about that separately in The ROI is not in the model; here the practical consequence is enough: the ownership test is settled entirely outside the technical layer.

QuestionAnswer that says projectAnswer that says operating-model change
Who owns it on day 400?"IT operations", "we will settle it after go-live"A name, written into annual objectives before the start
Whose result is it?Cost in IT, benefit nowhereCost and benefit in one P&L
What happens on degradation?"We will raise a support ticket"Threshold, alarm, budget for retraining or withdrawal
Where does year-two money come from?A new investment requestA maintenance line, with no end date
Who can change the process?Nobody on the delivery teamThe process owner, the same person as the system owner

If you have fewer than three answers in the right-hand column, you are running a project and you will get a project's result, whatever the business case says.

Moving the line from build to run

One operation follows from the test, and a finance director understands it without translation. Industry summaries citing Gartner data put maintenance at roughly two-thirds of the IT budget, with the remainder going to growth and transformation. No single citable study sits behind that [TO BE COMPLETED: source for the two-thirds share of the IT budget]. I take the proportion as an order of magnitude. What matters is the split itself, because a transformation line has an end date by definition and a maintenance line does not. Moving an AI initiative from build to run is not an organisational metaphor; it is a specific change of row in the budget. That change has two effects. The cost becomes recurring, and it lands in a unit that has an owner. The project, meanwhile, loses the right to close the subject with an acceptance certificate.

In practice this means a handful of very mundane moves. Name the owner before the first line of code and write the thing into their objectives before they have seen a demo. Plan the second- and third-year budget together with the first, in the same request, because a request split in two never gets its second half. Define a business-outcome measure rather than a technical one, and take the baseline before the start so that on day 400 there is something to compare against. Establish who has the right to change the process, and check whether that person sits in the same reporting line as the system owner. When those two roles sit in different lines, a working model changes nothing in the result.

Where the legislator has already settled it

For some use cases the debate about operating mode is closed, and not by a consultant's opinion. The AI Act, that is Regulation (EU) 2024/1689, imposes on the provider of a high-risk AI system two obligations that amount in substance to a prohibition on treating deployment as a project.

Article 72 requires a post-market monitoring system to be established and documented. That system must actively and systematically collect data on the performance of the AI system throughout its lifetime. The provider documents and analyses the data collected so as to be able to evaluate continuous compliance with the requirements. The monitoring plan is part of the technical documentation. Article 17 goes further and requires a documented quality management system covering the entire life cycle, from design through post-market monitoring, with change-control and continuous-improvement processes.

Read as an organisational document rather than a legal one, this sets out two requirements: a permanent owner and permanent measurement. Both cost money in every year, so both pull a permanent line into the budget. A structure with a closing date satisfies neither condition. In those use cases, treating deployment as a project stops being ineffective and becomes unlawful.

The boundary of this argument is sharp, however, and it has to be held. Both provisions bind the provider of a high-risk system within the meaning of the regulation. The sentence "the law requires continuous monitoring of every AI deployment" is simply untrue, and it must not be used as an argument inside a company rolling out an assistant that searches internal documentation. For the remaining cases the value of these provisions is different. The legislator has described the minimum shape of an organisation that keeps a working model alive, in language that gets through an investment committee. You may copy that template voluntarily.

When it really is an IT project

The strongest objection to this whole thesis is not "projects work". It is more uncomfortable than that: in practice the sentence "this is not an IT project" is often a request to be excused from discipline rather than a diagnosis.

Bent Flyvbjerg and Dan Gardner reported in How Big Things Get Done (2023) results from a database of more than 16,000 large ventures built and curated by Flyvbjerg. Of those, 8.5% come in on budget and on schedule at once. Add delivery of the promised benefits to the two conditions and 0.5% remain. The authors call this the iron law of megaprojects and point to optimism bias and to strategic misrepresentation, that is deliberately understating costs and overstating benefits so that the venture gets started at all. The benefits data there comes from a markedly narrower subset than the cost and schedule data, so the 0.5% stands on a thinner base than the 8.5%. I treat it that way.

What matters is the prescription they draw from it. It does not read "abolish projects". It reads "think slow, act fast", and standing beside it is the case for building modular. Both demand still more discipline before the start and a still shorter execution window. If the biggest killer of large ventures is a falsified business case and a scope that spreads, then a programme with no end date and no reckoning is not the cure for that disease. It is the disease in pure form, with better vocabulary. A transformation programme that reports progress in slides every quarter and never reports an outcome will survive inside a company longer than a failed project, because it has no gate at which it could fail.

To that add a classification argument, which honestly closes the scope of the thesis. Some AI initiatives really are IT projects. Replacing an optical character recognition engine with a language model has a finite scope and a clear acceptance criterion. So does switching to a cheaper API vendor, or moving recommendations onto a newer version of a library. Work like that happens inside a system that already has an owner and already sits in the maintenance line. Elevating such a swap to the rank of an operating-model change costs two quarters of alignment and buys nothing. Narayan himself draws this distinction. In construction, project mode makes sense because the skills of building and of operating genuinely differ; in software they differ less, so separating build from run costs more than it returns. Product mode is therefore not universally better; it is the right mode for work that has no natural end, and a classification error is expensive in both directions.

So where does the boundary run? It runs where the mandate runs. If an initiative requires changing somebody's process, moving a decision between roles, or changing what people are paid a bonus for, no project manager can deliver it, and a project structure is a classification error. If an initiative touches nobody's process and has an owner from the day before the start, run it as a project, set hard gates and close it on time. The test for that is one sentence long: check whether the initiative's success requires somebody outside the delivery team to start working differently.

Both disciplines at once

The worst thing you can do with this distinction is turn it into a slogan and stop setting gates. So hold both tools at once. Hold to account as a project whatever has an end, and hand whatever has no end to somebody who on day 400 loses a bonus when the chart turns down. If you cannot give that name today, you are building something nobody will take over.

Sources

  1. primaryMelvin E. Conway, How Do Committees Invent?, Datamation, April 1968 — the canonical formulation of Conway's law and the corollary on organisational flexibility attributed to Brooks.
  2. primaryNagappan, Murphy, Basili, The Influence of Organizational Structure on Software Quality: An Empirical Case Study, ICSE 2008 — 3,404 Windows Vista binaries; organisational metrics, NOEE among them; a correlational case study, limited external validity flagged by the authors.
  3. primarySriram Narayan, Products over Projects, martinfowler.com, 20 February 2018 — six dimensions of the difference between project and product mode; no mechanism for verifying benefits after delivery; the author's own distinction regarding work of finite scope.
  4. primarySculley et al., Hidden Technical Debt in Machine Learning Systems, NIPS 2015 — the share of ML code in a real-world system, the CACE principle, debt repaid in operation.
  5. primaryVela, Sharp, Zhang, Nguyen, Hoang, Pianykh, Temporal quality degradation in AI models, Scientific Reports, 2022 — model ageing as a phenomenon distinct from data drift; 128 model–dataset combinations, degradation in 91%; classical predictive models, not language models.
  6. primaryRyseff, De Bruhl, Newberry, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, RAND RR-A2680-1, 2024 — a qualitative study, 65 interviews; the "more than 80%" figure is somebody else's estimate quoted in the report, not a RAND finding.
  7. primaryRegulation (EU) 2024/1689 of the European Parliament and of the Council (the Artificial Intelligence Act), Art. 17 and Art. 72 — a quality management system covering the whole life cycle and post-market monitoring; obligations of the provider of a high-risk system.
  8. primaryJørgensen, Moløkken-Østvold, critique of the Standish Group CHAOS reports [TO BE COMPLETED: title of the paper, journal and year] — undisclosed research design, unspecified sampling criteria and unstable definitions of success; IT project failure statistics are unfit for comparison.
  9. secondaryGartner, press release of 7 April 2026 on a survey of 782 infrastructure and operations (I&O) leaders conducted in November and December 2025 — 28% of use cases fully successful, 20% failed; respondent self-report, the unit of analysis is a use case in I&O. Figures taken from press coverage, because Gartner's newsroom does not make the release available for automated retrieval (as of 12 September 2026).
  10. secondaryMcKinsey / QuantumBlack, The State of AI, 2025 edition, n = 1,993 — 21% of organisations report fundamental redesign of some workflows; the strongest relationship of that practice with EBIT impact; self-report survey, correlation.
  11. secondaryMik Kersten, Project to Product: How to Survive and Thrive in the Age of Digital Disruption with the Flow Framework, IT Revolution, 2018 — milestone funding against scope defined up front; a practical frame based on the author's experience.
  12. secondaryBent Flyvbjerg, Dan Gardner, How Big Things Get Done, 2023 — the iron law of megaprojects, 8.5% and 0.5%, optimism bias and strategic misrepresentation; the author's own database, benefits data on a narrower subset.