The cost of holding it.

Fourth and final piece in a series on accountability in regulated AI. The Named Owner argued that every AI deployment needs a person, not a function, against the decision. You Cannot Acquire Accountability argued that the responsibility cannot be outsourced, bought, or merged away. The Judgement Muscle argued that the owner has to stay sharp enough to actually exercise it. This piece is about what doing all three actually costs, and about who quietly wins by paying that cost while everyone else hopes the bill will not arrive.

Three pieces in, the natural reader question is the one I have been ducking. If accountability cannot be delegated, cannot be acquired, and cannot be safely automated, what does it actually cost to hold it properly? And if the cost is real, why are so few operating models openly budgeting for it?

The honest answer is that holding accountability is expensive, the costs do not show up where most finance teams look for them, and the organisations paying the bill are not the ones celebrating the productivity gains. The accountability bill is borne by the people who refuse to let the structure off the hook. The productivity reporting accrues to everyone else. That asymmetry is not an accident. It is the operating model working exactly as designed, and it is the reason the cost of holding accountability is the most underdiscussed line item in the AI conversation.

The plumbing changes hands

The cleanest worked example of this asymmetry is currently running in adtech. Publicis is acquiring LiveRamp for $2.2 billion. Most of the marketing-industry coverage has framed the story as a neutrality question: LiveRamp has historically sat as connective infrastructure between advertisers, publishers, data providers, platforms, and measurement vendors, and the acquisition by a major agency holding company changes the meaning of that connective layer.

Jason Bauer's piece this week made the argument as precisely as I have seen it made. Even if LiveRamp remains contractually open, rival agencies, publishers, platforms and marketers will now ask the harder questions. Who controls the identity spine. Who sees the data flows. Who benefits from marketplace economics. Who influences activation paths. Who owns the measurement layer. The market will hedge accordingly: dual-source identity, scrutinise clean-room governance, ask for stronger auditability, build redundancy into the connective layer that was previously treated as a utility.

That hedging behaviour is the cost of holding accountability, made visible in real time, in observable budget decisions. It is not a future projection. It is what the procurement teams are already doing this quarter. Every advertiser who now runs a second identity resolver is paying the bill. Every publisher who builds a direct integration alongside the LiveRamp pipe is paying it. Every clean-room agreement that now requires stronger auditability clauses is paying it, in legal hours, in slower onboarding, in capability redundancy.

The total bill is going to be substantial, and almost none of it will appear in a line item labelled the cost of holding accountability. It will appear as legal spend, integration cost, vendor diversification, technical debt, and slower decision cycles. It will appear, in other words, as everything except what it actually is.

The same shape, eighteen months out

Pharma will see this pattern next, and it will arrive in AI procurement rather than in the connective infrastructure that started the adtech story, but the shape will be identical.

Look at the capabilities most pharma functions are buying in 2026: model-assisted medical writing, AI-driven KOL prioritisation, agentic systems in early R&D, promotional content review tooling, regulatory submission acceleration. They sit in exactly the position LiveRamp sat in five years ago. They are being treated as utility layers. They are not utility layers. They are decision infrastructure, and the named-owner question, the accountability question, and the judgement-muscle question all collapse onto them the moment a regulator, a payer, or a patient asks who signed off on the output.

The pharma equivalent of the LiveRamp hedge is already visible in early form. The organisations that take the accountability argument seriously are dual-sourcing critical AI capabilities, building model-agnostic workflow layers, refusing single-vendor lock-in on anything that touches a regulated decision, and writing decision-rights frameworks that survive a vendor changing hands. They are building the redundancy and audit infrastructure before they need it, not after. They are paying the bill now.

The organisations that are not doing this are running faster on quarterly metrics. That is the part that needs naming clearly. The cost of holding accountability is paid out of the same budget that pays for the productivity story, and the productivity story is the one that gets to the board. The teams paying the bill look slower than the teams who are not.

Why the bill never lands on the right desk

Three structural reasons the cost of holding accountability is so consistently understated, in pharma as in adtech, and as it will be in every other regulated function that adopts AI at pace.

First, the costs are diffuse and the benefits are concentrated. The named owner who insists on a second model for cross-validation, the procurement lead who refuses to sign a contract without portability clauses, the compliance director who slows a deployment to write a decision-rights framework. These people generate cost that lands across multiple budgets and timelines. The productivity gain that comes from skipping those steps lands cleanly in one quarter and one P&L. The accounting structure rewards the shortcut and punishes the prudence. This is not a critique of individual leaders. It is the operating model behaving exactly as designed.

Second, the benefits of holding accountability properly only become visible when something goes wrong. A function that has built redundancy, audit infrastructure, and a real named owner looks identical to a function that has built none of those things, until the day the regulator calls, the model misclassifies a safety signal, or the vendor restructures and the data position evaporates. On that day, the two functions look completely different. On every other day, the function that paid the bill looks like the slower one. The board sees the bill but not the disaster that did not happen.

Third, and this is the one Martha put more sharply than I had, the cost of holding accountability is not really paid by the function that holds it. It is paid by the function that refuses to let the structure off the hook. In her phrase, it is those who can step into the void and deliver outside the holding-company structure who win. The accountability cost, in other words, is a competitive advantage in disguise. The organisations that pay it cleanly are not absorbing a tax. They are buying a position the market will value later, when the bill the others avoided finally arrives.

Who wins by paying

Three categories of organisation are quietly winning the accountability argument right now, and worth naming.

The independents. In adtech, the operators who can step outside the holding-company structures and offer genuinely connective layers, direct publisher integrations, alternative identity stacks, independent measurement, are seeing demand they have not seen in years. In pharma, the same dynamic is starting to favour boutique advisory firms, independent data-science partners, and validated point solutions whose entire value proposition is that they have no incentive to drift toward the buyer's interests. Their independence is not branding. It is operating model, and it is now a procurement criterion.

The regulated functions that hold the line internally. Quality, regulatory affairs, medical, compliance. The functions that have historically been described as friction by the commercial side are, in the AI era, the functions that protect the institution from the consequences of accountability drift. The organisations that resource these functions properly, give them the authority to refuse, and accept that they will sometimes slow deployments down are paying the bill on purpose. They are also the organisations least likely to make the headlines for the wrong reasons.

The leaders who can articulate the cost without apology. This is the smallest group, and it is the one that decides whether the other two get the resourcing they need. A leader who can explain to a board that a slower deployment with proper governance is a better deployment, and who can quantify the avoided cost rather than only the realised one, changes the conversation. Almost no one is doing this well in 2026. The few who are will set the standard the rest will eventually have to meet.

Where the series lands

Four pieces. One argument, told in four parts.

The Named Owner established that AI accountability has to sit with a person, not a function. You Cannot Acquire Accountability established that the responsibility does not transfer with an acquisition, a restructuring, or an outsourcing arrangement. The Judgement Muscle established that the owner has to remain capable of exercising that ownership, and that strong model performance is the anaesthetic that quietly erodes the capacity to do so. This piece establishes the last part. Holding accountability properly costs real money, the cost shows up everywhere except in the line item that names it, and the organisations paying it cleanly are not absorbing overhead. They are buying a position the rest of the market will eventually have to buy at a higher price.

The four arguments are sequential, but they collapse into a single instruction. Name the owner. Hold the accountability. Keep the owner awake. Pay the bill before someone else sends it to you.

That is the entire job. It is harder than it sounds, more expensive than most operating models admit, and rarer than the market currently rewards. It is also, as every regulated industry eventually learns the expensive way, the only version of AI adoption that survives contact with the first real incident.