The commercial design space.

Sixth and final piece in the accountability series. The Named Owner argued that every AI deployment needs a person against the decision. You Cannot Acquire Accountability argued the responsibility cannot be outsourced or merged away. The Judgement Muscle argued the owner has to stay sharp enough to exercise it. The Cost of Holding It argued the discipline is expensive in places the accounting does not name. The Design Space argued the discipline already exists in pharma quality, transposed from ICH Q8. This piece is for the rooms ICH Q8 does not reach. Medical affairs. Market access. Promotional review. KOL prioritisation. Commercial analytics. The functions where the design space has to be drawn from scratch by people who did not author the framework they are required to defend.

The previous piece in this series, The Design Space, argued that pharma already knows how to govern probabilistic systems. The discipline lives in ICH Q8, has lived there since 2009, and the right transposition into LLM governance is to define the bounded territory within which the model operates reliably, characterise the edges, and accept that operating inside the space is a probabilistic claim with documented controls at the boundary. The piece anchored on Nuno Valério's Trust Architecture newsletter, which had made the transposition first and properly.

Nuno replied to the piece in the comments with a question I had hoped would not get asked, mostly because I had not yet worked out the honest answer. The character mapping to your prior series is sharp, he wrote, but I'd love to see what the extension actually adds that the source piece doesn't. He was right. The transposition was clean. The extension was framing. The substance was his.

This piece is the answer. The genuine extension does not sit inside the quality function. It sits in the rooms ICH Q8 does not reach.

The functions without a quality manual

ICH Q8 reaches GMP manufacture. It reaches process development. It reaches CMC sections of regulatory submissions. It does not reach medical affairs. It does not reach market access. It does not reach promotional content review, KOL prioritisation, segmentation models, commercial analytics, or any of the AI deployments now multiplying across the commercial side of every pharma in the world. None of these functions inherits a design-space discipline. None of them has fifteen years of small-molecule rigour pre-loaded into the way it thinks about probabilistic systems.

This matters because the commercial functions are where AI deployment is moving fastest, where named owners are being asked to sign against decisions the rest of the industry has not been trained to govern, and where the consequences of getting it wrong land hardest. A promotional review tool that approves a borderline claim is not an out-of-spec batch. It is an interaction with the regulator, a competitor, a patient advocacy group, and a national press, all at once. The boundary the named owner is being asked to defend matters more than the quality manual ever pretended. The institutional scaffolding that helps them defend it is mostly absent.

The map is being drawn at the wrong altitude

Where the discipline does exist in commercial pharma, it usually does not exist locally. It exists in a global function. The AI Centre of Excellence in the New Jersey campus, the Digital and Data office in Basel, the enterprise rule-book drafted in the regional headquarters, with sign-off from the Chief Digital Officer and review from group legal. These functions are competent and well-resourced. They are also, in almost every case, drawing the map at the wrong altitude.

The map being drawn is the global design space for the commercial AI deployment. The cartographers are headquarters digital and AI leaders. The territory is every country the company operates in. The problem is that the territory contains regulators the map does not see clearly. HPRA in Ireland. BfArM in Germany. ANSM in France. PMDA in Japan. AGES in Austria. The IPHA code, the ABPI code, the Swiss promotional code, the LIF code in Sweden. Each of these is a real boundary the global cartographer is drawing across without being able to see properly from the altitude they are working at.

None of which is the cartographer's fault. It is structural. A digital leader in Basel cannot hold the IPHA promotional code in working memory alongside the ABPI code alongside the FSA code alongside the practical operating norms of fifteen other national markets. The map they can draw is the one they can see from where they are. The map the local commercial team has to operate inside is the one the local regulator can see from where they are. These are not the same map.

A composite Tuesday

Composite vignette, drawn from several real client conversations, identifying details changed. The local commercial team in a European affiliate of a top-twenty pharma is preparing to deploy an AI-assisted promotional content review tool. The tool was selected centrally, configured against a global content policy, trained against a corpus of approved materials from across the enterprise, and validated against the FDA-aligned promotional standards the global compliance function uses as the conservative baseline.

The local medical affairs lead, the named owner of the deployment in her market, reads the configuration documentation on Monday. By Tuesday morning she has a list of fourteen things the tool will not catch and seven things the tool will refuse that the local code permits. The seven things include a comparative efficacy claim that is permitted under the local self-regulatory code with appropriate sourcing, a competitor reference structure that is standard practice in her market, and three patient-language constructions that the local patient advocacy bodies have explicitly endorsed but the global model treats as overclaim. The fourteen things it will not catch include a transparency reporting trigger specific to her market, an off-label proximity rule that does not exist in the FDA regime the model was trained against, and the local rules on HCP interaction disclosure that her affiliate will be audited against next quarter.

She raises this with the global digital lead on a Wednesday call. The call is friendly and unproductive. The global lead's framework is the framework. Her market is one of fifteen the framework has to serve. The framework cannot be configured per market without breaking the auditability claim that justified the central procurement in the first place. The compromise offered is a local exception register that the local team can use to log overrides. Every override increases her personal exposure as the named owner. Every refusal of the tool slows her team down against competitors whose tooling was drawn at a different altitude. She is being asked to operate inside a design space that is simultaneously too loose and too tight for the territory she is responsible for, and to absorb the consequence of both.

This is not a hypothetical. It is happening in every major commercial pharma right now, in some form, in most markets. The names change and the tool changes. The shape does not.

Two failures, the same map

The vignette names a thing that needs naming directly. The global commercial AI design space fails in two opposite ways at the same time. It is too loose, because it permits things the local regulator will sanction. It is too tight, because it forbids things the local regulator permits and that competitors are doing freely. Both failures sit inside the same framework simultaneously, because the framework was drawn at an altitude that cannot see the edges of either.

This is the failure mode the ICH Q8 transposition does not protect against. ICH Q8 design spaces are drawn against a single regulator and a single product. The cartographer has line of sight to the boundary they are drawing. The commercial AI design space is drawn against fifteen regulators and a portfolio of products, by a function that does not have line of sight to most of them. The discipline is the same. The conditions for the discipline to work are different. The transposition cannot be left to do all the work.

The named owner sitting under that map carries the consequences of both failures. The accountability series argued repeatedly that the named owner has to be a person, not a function, and has to have the authority that matches the consequence they are signing against. In the commercial pharma case the consequence is local. The authority is global. The named owner has been set up to lose.

Where the map has to be drawn from

Three operational moves, for global digital and AI leaders who recognise the failure pattern in their own organisations, and for local commercial leaders who are tired of carrying the consequence of a map they did not draw.

First, the design space has to be drawn at the altitude where the regulator is visible. This does not mean fifteen separate frameworks. It means a global framework that explicitly carries a thin layer of local boundary documentation, authored locally, signed off locally, and treated as part of the validation dossier rather than as an exception register. The local boundary is part of the map, not a footnote to it. If the global framework cannot accommodate that, the framework is at the wrong altitude.

Second, the named owner has to sit at the level the boundary is drawn against. If the framework permits or forbids things that are decided locally, the named owner is local. If the framework operates uniformly across markets, the named owner is global. Mixing the two, which is the current default, is the configuration that sets the local owner up to lose. The accountability series argued that the seat and the authority must match. In the commercial pharma case, the seat and the altitude must match.

Third, the override register has to be honest about what it is. Every override the local team makes against the global framework is information about where the framework is drawn wrong. It is not a list of local nuisances to be tolerated. It is the empirical edge of the design space, being mapped by the people closest to it. A digital function that reads its override registers as feedback rather than as friction is the function that draws better maps in the next iteration. A digital function that reads them as exceptions to be minimised is the function that keeps drawing the map at the wrong altitude and wondering why the local teams keep complaining.

Where the series lands

Six pieces. One argument, told in six 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 a restructuring or an acquisition. The Judgement Muscle established that the owner must remain capable of exercising the ownership. The Cost of Holding It established that holding accountability properly costs real money in places the accounting structure does not name. The Design Space established that the discipline of doing all four properly is the design-space discipline pharma quality already runs on small molecules. This piece established the last part. In the commercial functions where ICH Q8 does not reach, the discipline has to be drawn from scratch, and it has to be drawn at the altitude where the regulator is visible. Otherwise the named owner is operating inside a map they did not author and cannot defend, against consequences that land locally even when authority sits globally.

The six arguments collapse into a single instruction, slightly longer than the five that preceded it but earning the additional clauses. Name the owner. Hold the accountability. Keep the owner awake. Pay the bill. Run the deployment inside a design space that is bounded, documented, and defended. Draw the map at the altitude where the regulator is visible, and put the named owner at the same altitude as the map.

That is the entire job. Six pieces to say it. The map for AI governance was not missing. It has been sitting in the quality manual the entire time. The remaining work in the commercial functions is to draw the map at the altitude where it can be defended, and to stop pretending the local consequence can be governed by a global cartographer who has never sat in the room where the consequence lands.

This concludes the six-part accountability series. The Named Owner, You Cannot Acquire Accountability, The Judgement Muscle, The Cost of Holding It, The Design Space, and The Commercial Design Space are on the Wild Geese site.

With thanks to Nuno Valério, whose comment on The Design Space asked the question this piece exists to answer, and whose Trust Architecture newsletter supplied the framework the series closes on. The cartography argument and the global-versus-local extension are mine. The discipline they extend is his and Georgi Tancev's.