Five Projects, Five Names, No Committees

Ask a large organisation who owns its first retrieval-augmented assistant and the answer tends to arrive as a room. A working group. A steering forum. A partnership between the data function and the business, which is how people say co-owned when they want it to sound deliberate.

Push once and the room comes apart. The person who can stop the work, the person paying for it and the person who has to answer for what it tells a colleague turn out to be three different people, and each of them describes what happens next as something somebody else decides.

That is not a compromise between two claims on the same initiative. It is a device for making sure nobody is ever the person who was wrong.

I keep returning to one page from an internal digital strategy proposal at a global pharmaceutical company, written for a governance forum, setting out an operating plan across three annual stages. The page is the plainest thing in the deck: a table with three columns, an item reference, a topic and an owner. It lists five initiatives. It names five people, one against each. No initiative carries two names, and no name carries two initiatives.

This is a proposal to a governance forum, so what it records is an intent to run the walkthrough, not a minute of it having happened. The design is still the design, and the design is what I want.

The page is an agenda, which is the whole point

A list of owners in a deck is cheap. Most transformation plans have one in an appendix, assembled afterwards so the governance pack looks complete. Those lists are inert. Nobody reads them out, and nothing happens if a name in one is wrong.

This list sits as lettered sub-items under one numbered agenda item, which puts the five initiatives inside a single slot in a meeting that recurs. The owner is not a field in a register. The owner is who speaks when their letter comes up. That converts a name from documentation into a scheduling fact, and scheduling facts enforce themselves in a way registers never do. Somebody has to stand up. Had two people been written into a row, the meeting would have had to decide which of them stood up, in front of everyone.

Committees survive because that moment never arrives. Co-ownership is comfortable in a document and impossible in a diary.

Five different kinds of thing, one rule applied to all of them

The interesting part is the heterogeneity. Had the five all been technology projects, one owner each would be unremarkable, because a project has a manager whether or not anybody thinks about ownership. These five are not commensurable. One is the alignment of a programme spanning markets. One is a therapy area's digital plan, cutting across brand, medical and commercial. One is a whole market's integrated customer experience. One is a single physician-facing platform. One is a discipline that touches everything and belongs to nothing, which is why content strategy is the item most organisations would hand to a group.

Nothing in that list suggests a common unit. The plan applies the rule anyway. The unit of accountability is not the unit of work, and once you separate the two, the awkwardness of naming one person for something as diffuse as content strategy stops being an argument against doing it. The named person does not do content strategy. The named person is who you ask.

The topics are described as alignments and overviews, which is honest about how early they were. Early is exactly where co-ownership does its worst damage, because there is no delivery date yet to expose the drift. A named owner is the only instrument available at that stage.

The same work, listed twice, named once

Elsewhere in the deck sits the prioritised technology portfolio: seven numbered entries covering analytics and shared customer data, integration across touch points, campaign management, two pieces of portal work, a therapy area programme and a platform upgrade. It is a prioritisation artefact, ranked one to seven, and no owner is named against any entry.

Four of the five walkthrough topics reappear there. The campaign management programme, the therapy area plan, the market portal and the physician-facing portal are in both lists. The fifth, content strategy, sits among the plan's stated priorities instead.

I read that duplication as deliberate rather than sloppy, because the two pages answer different questions. The technology list answers what gets built and what comes first, a question about capacity and dependency, to which a person's name adds nothing. The walkthrough answers who is accountable for the outcome the build is meant to produce, to which a place on a priority list adds nothing. Where the deck lists systems it names none, and where it lists commitments it names exactly one each.

Most portfolio reviews I sit in have collapsed those two views into one spreadsheet, and the collapse is where accountability leaks out. A row carrying both a delivery date and an owner invites the owner to report on the date, which is a question about the vendor and the sprint, rather than on the outcome, which is a question about them.

Six mandates, none of them shared

The governance page does the same at the level of functions. Six are set out, each with one sentence of what it owns: setting direction and issuing the guidance that regions and brands work from; designing new services and improving the ones already running; holding the customer relationships and generating demand; sourcing, reusing and maintaining the content assets; running shared services and the reporting above them; and building and evolving the platform the content sits on.

Read the six together and what stands out is the absence of a shared clause. Nothing is jointly held with the function beside it. The seams stay visible, because content sourcing and content platform development have to meet somewhere, but the plan puts the seam between two named mandates rather than dissolving it into one shared one.

Even the analytics remit inside the strategy group is split rather than pooled. One job is checking whether what was published is being used, and feeding that back as course correction. The other is measurement and target tracking for the group, its regions and its markets. Different jobs, different customers, and merging them is how an analytics team ends up accountable for everything and answerable for nothing.

What this has to do with the programmes being funded right now

Almost everything I am asked to review this quarter is an assistant of some kind. Retrieval over a document set nobody has curated. A copilot for a drafting task. A few carefully scoped autonomous experiments in low-consequence queues. The MLOps around them is better than it was two years ago, and the data readiness underneath is worse than anybody admits.

In every one of those programmes, three things reliably have no owner. The evaluation set has none, so nobody can say whether the assistant improved and the demo becomes the evidence. The retrieval corpus has none, so stale documents stay in it and the assistant confidently cites a superseded procedure. The lineage of the underlying data has none, so when somebody asks where a figure came from, the answer is a diagram rather than a name.

None of those three is a technical gap. Each is a co-ownership decision made by default, usually because assigning it felt premature while the pilot was still a pilot.

Political agreement on the European AI Act has been reached, and whatever the final obligations require, the question underneath them will not be which forum oversaw a system. It will be who is accountable for it. An organisation already running a standing walkthrough with one name per initiative can answer on the day. An organisation with a working group will be reconstructing an answer under time pressure.

5

Initiatives on the walkthrough page, each with one named owner

5

Distinct owners, no name repeated

7

Entries on the prioritised technology list, no owner column

6

Functions on the governance page, one mandate each

Different instruments for different work

The plan does not use names everywhere, and that restraint is worth copying too. The guidance backlog, roughly a dozen documents covering search, media, social, mobile, design, email and agency selection, is governed by completion state, by priority band and by date. Not one of the dozen carries a person. A usable draft was estimated within thirty days and final publication to the markets was set at sixty, and both of those are figures in a plan the deck never comes back to grade.

So the deck reaches for a different instrument depending on the work. Rank order for the systems, banding and dates for the publishing backlog, a named person for the initiatives whose outcomes somebody has to defend. Only the last of the three puts a human name anywhere. The failure I see most often is one instrument applied to everything, usually dates, because dates are easiest to put in a dashboard and least revealing about whether anyone is accountable.

Two owners is not twice the accountability. It is a standing excuse, available to both of them, for as long as the initiative runs.

What I would take from it

Write the list. One line per initiative, one person against each, and check the column both ways: no initiative with two names, no person with more than one. If a name appears twice, one of those initiatives is not real yet, and finding that out is worth the exercise on its own.

Put it on a recurring agenda rather than in a register, so ownership is expressed as who speaks rather than as who was written down. Keep the portfolio priority list on a separate page, without names, so nobody reports on a vendor timeline when asked about an outcome. Give the shared artefacts of your AI work the same treatment, because those are the pieces otherwise owned by the programme, which means by nobody.

None of this needs a platform, a tool or a reorganisation. It is a table with three columns. Consult exists at RealAI mostly to run the argument that produces that table, because the argument is short and organisations avoid having it for years.

Details are as set out in an internal digital strategy and operating plan proposed to a governance forum at a global pharmaceutical company: its walkthrough page, its prioritised technology list, its guidance publication backlog and its governance operating model. That deck is a plan put forward for approval, not a record of results. Reading its ownership convention as the transferable part is ours.