Every major programme produces two readings of itself: the one in the PMO pack and the one the project director carries in their head. The two are rarely the same, and the second is more reliable. The pack represents what the reporting system has been able to digest in the weeks since the underlying events occurred. The director is describing what they can see today. The PMO is the institutional response to that gap. The operating model we have settled on for it, at the scale most major UK programmes now run, does not close the gap. It widens it.
This is a critique of an operating model and the scale at which it is now configured, not of PMOs as a class of institution and not of the people who work within them. Good PMO professionals are typically the first to notice that the apparatus is failing them, and they work hardest to compensate. If anything, the case for system-level integration is the case for giving those practitioners back the senior, judgement-bearing role the PMO was originally meant to play.
What kind of load is this
Cognitive overload on major programmes is currently diagnosed three ways. It is called burnout, and met with resilience training. It is called a tooling gap, and met with AI copilots. It is called a delivery problem, and met with more headcount in the PMO. None of these is the right diagnosis. The load on the project director is a structural feature of running real-time governance through a human apparatus never designed to carry it.
Four kinds of load sit on the project director's desk:
- Judgement load: weighing partial evidence, balancing competing principals and committing to a course of action under uncertainty. This is the work the role exists to do.
- Integration load: holding two adjacent systems together by hand because no interface exists between them.
- Reporting load: preparing the same information in different formats for different audiences.
- Ambiguity load: operating where authority has not been clearly assigned, so the next decision falls to whoever is unwilling to let it drop.
Only the first is healthy. The other three are governance debt being paid in human attention. The PMO was supposed to absorb the back three so that the director could focus on the first. Under the current operating model it absorbs them in a way that adds latency and strips context, then escalates the residue back upward. Zika-Viktorsson and colleagues named this "project overload" in 2006 [1], and the NAO's 2025 lessons-learned review of UK government mega-projects calls for governance that is clear about who has authority and accountability [2]. The governance debt is real, named, and being paid where it has always been paid: at the integrator seat.
The role is the symptom; the missing system is the cause.
Information velocity has outrun the human apparatus
The amount of information passing through a major programme has grown by an order of magnitude in twenty-five years [unverified]. The cognitive equipment of the people receiving it has not changed at all. Asana's Anatomy of Work Index 2023 puts roughly 58% of knowledge-worker time on "work about work" [3]. Microsoft's 2025 report Breaking Down the Infinite Workday counts 275 interruptions a day on average [4]. The two measure different things, but both point to the integrator seat as the place where the volume lands.
Construction has the same problem in its own dialect. The Navigant Construction Forum's 2013 study of 1,362 projects found a mean of 796 formal RFIs per project, eight hours of designer time per RFI, and a first-reply lag of 6.4 days [5]. It is a US-weighted sample from 2001-2012, the cleanest available proxy rather than a current UK benchmark. PlanGrid and FMI's 2018 Construction Disconnected survey put project-team time on rework, conflict resolution and information-hunting at 35% [6].
The information ought to be flowing as events (a change request raised, a clash report returned, an RFI logged, an instruction issued), each captured at source, each carrying its own metadata, each available to the board the moment it occurs. Instead the events are caught by people, re-typed, summarised and routed by hand.
Latency and context-loss
The first failure is latency. The PMO operates at committee cadence. Reports are written weekly, reviewed fortnightly, escalated monthly, packaged quarterly. The pattern, directional rather than dataset-strength, is that problems surface to boards roughly a quarter after they actually occurred. By then they are not problems any more; they are facts. The board cannot act on a fact the way it could have acted on a signal.
The second is context-loss. The raw event becomes a paragraph, the paragraph becomes a status, the status becomes a RAG rating, the rating becomes a colour on a slide. The board cannot interrogate what it cannot see. So it asks for more reports. The PMO writes them. The chain gets longer. The signal gets thinner.
The lean tradition has a word for this. Map Taiichi Ohno's wastes onto a project director's desk and they are descriptive, not metaphorical. Over-production is the report no decision-maker reads. Waiting is the sign-off chain. Defects are the rework caused by out-of-date information. Koskela's "making-do", working with incomplete prerequisites because the information cannot arrive in time [7], is the decision taken without the information needed. Latency and context-loss are not side effects of the PMO operating model; they are the production process. The apparatus we have built is, in effect, manufacturing the very muda it was designed to manage.
AI is redistributing the load upward
The fashionable answer is that AI will resolve all of this. There are real productivity gains in status reporting, scheduling and summarisation. But the framing is wrong, and the wrongness matters. Lee and colleagues at Microsoft Research found in 2025 that knowledge workers reported reductions in cognitive effort when AI handled a task, alongside a drop in confidence in their own judgement on the same task [8]. Bent Flyvbjerg argues that generative AI prioritises persuasion over truth [9]. In major-programme settings, the AI gives the answer that looks most like the answers it has seen, not the answer the situation requires.
The training-ground tasks that produced the next generation of senior judgement (the bid review, the schedule resynthesis, the first-draft risk register) are precisely the tasks AI handles cheapest. What remains is pure judgement, governance and ambiguity. We are making the job narrower and harder, not easier.
AI is well placed to be the integration layer that finally retires the human routers. It is poorly placed to be a load-reliever for an unchanged PMO operating model. Confuse those two roles and you keep hiring people to babysit the AI that was supposed to retire the people we hired to babysit the system.
People as proxies for systems
We are using people as a proxy for systems. The PMO operating model is the worked example. So are the layers of expediters, document controllers, RFI managers, interface managers and progress reporters that have grown up around the integrator role. Each role exists because information that ought to flow through a designed integration layer is being routed by hand. The role is the symptom; the missing system is the cause.
An airline operations centre runs on a single event stream: every flight, gate change and crew swap is a record updated at source by the party with the authority to update it. A clearing bank's risk function keeps one canonical record per exposure, with the narrative generated from the underlying events. Construction has people doing it instead.
The industry does not have a people problem. It has a systems-design problem that it keeps trying to solve with more people. Brooks's Law has described the mechanism since 1975: adding people to a late project makes it later, because communication channels grow faster than throughput. A large-scale GitHub study has since measured those diseconomies of scale at team level [10].
Four shifts the board can demand
The remedy is not more governance and not less governance. It is a different delivery model for the governance boards rightly demand: real-time over monthly reporting; context preserved, with a click-through from any status to the underlying entries and source documents; integration platforms over PMO bloat, so that where two systems need to talk the answer is an interface, not another expediter; and small senior decision teams over large reporting bureaucracies.
Governance is the requirement. The operating model is the problem. The people inside today's PMOs have been carrying the gap between governance and system for two decades. The proposition here is not to displace them; it is to build the system that gives them their senior, judgement-bearing role back. That is the work, and it is the thinking behind Synaform's PMO Manager.
Sources
- [1] Zika-Viktorsson, A., Sundström, P. and Engwall, M. (2006). "Project overload: An exploratory study of work and management in multi-project settings." International Journal of Project Management 24(5): 385-394.
- [2] National Audit Office (2025). Lessons Learned: Governance and Decision-Making on Mega-Projects. Session 2024-25, 14 March 2025. https://www.nao.org.uk/
- [3] Asana (2023). Anatomy of Work Index 2023: Rise of the Connected Enterprise. https://asana.com/resources/anatomy-of-work-hub
- [4] Microsoft (2025). Breaking Down the Infinite Workday. WorkLab special report, aggregated M365 telemetry. https://www.microsoft.com/en-us/worklab
- [5] Navigant Construction Forum (2013). Impact and Control of RFIs on Construction Projects. 1,362 projects, US-weighted, 2001-2012. https://www.cmaanet.org/sites/default/files/resource/Impact%20&%20Control%20of%20RFIs%20on%20Construction%20Projects.pdf
- [6] PlanGrid and FMI (2018). Construction Disconnected. Survey of c. 600 construction leaders.
- [7] Bolviken, T. and Koskela, L. (2016). "Why hasn't waste reduction conquered construction?" IGLC 2016. https://eprints.hud.ac.uk/id/eprint/29121/
- [8] Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R. and Wilson, N. (2025). "The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers." Proceedings of CHI 2025. DOI 10.1145/3706598.3713778.
- [9] Flyvbjerg, B. (2025). AI as Artificial Ignorance. Working paper, SSRN 5119382.
- [10] Gote, C., Mavrodiev, P., Schweitzer, F. and Scholtes, I. (2022). "Big Data = Big Insights? Operationalising Brooks' Law in a Massive GitHub Data Set." ICSE 2022. https://arxiv.org/pdf/2201.04588