At a Reuters Events pharma conference in Barcelona earlier this year, roughly fifty senior industry professionals sat down with pens, paper and a question they had not been formally asked before: if your organisation’s human capacity were no longer a limiting factor, what would you do differently? The question came from a 75-minute design session, and it was less theoretical than it sounds. Its purpose was to move experienced practitioners out of the abstract discussion about artificial intelligence, which had filled the preceding day’s panels and keynotes – and into the practical mechanics of what an AI agent would actually do in their specific corner of the business.
The context that preceded the session was not short on urgency. The average pharmaceutical company operates across approximately 78 different systems and more than 50 manual process steps. Data sits in separate pockets across commercial, medical and clinical functions – producing what Ihab Fakhouri, GTM Lead at Salesforce, described as “disconnected point solutions”: a collection of capable parts that rarely communicate well enough to drive consistent output at scale. A survey of senior health leaders found that 64% identified inadequate data infrastructure as a significant barrier to AI adoption. The technology, in short, is rarely the source of the difficulty. Data fragmentation nearly always is.
A McKinsey study referenced in the session’s opening put some numbers to the opportunity. Agents – AI systems capable of autonomously executing sequences of actions – could free up between 25 and 40% of workforce capacity across the industry. Rather than eliminating roles, the same research projected the creation of more than ten new categories of work: agent orchestrators, governance specialists, quality controllers, and others still taking shape. The argument for involving practitioners in the design process, rather than handing it over to technology teams, ran throughout the session. As Tom Smith, VP Life Sciences at Salesforce, put it, “the people that create the best use cases, that drive the best ROI are the people closest to the workflows.”
Before the design work began, participants saw what an agent looks like in practice. The demonstration followed a fictional field sales representative as he began his day with a single-agent conversation that drew together prescribing patterns, approved clinical literature, unstructured field notes, and commercial context into a single briefing. Venkat Swaminathan, Account Director at Salesforce, who presented the demonstration, described the aim as moving beyond “next best action” to something he called “next best interaction” – a distinction that turns on whether the agent is simply routing a task or genuinely calibrating for context, channel, tone and timing. Each table then received a design framework built around five questions: what is the agent’s role, what data does it draw on, what actions is it authorised to take, what guardrails govern its behaviour, and through which channels does it operate?
The five groups produced five distinct agents in roughly 35 minutes, each anchored in a different function, objective and desired outcome. The marketing group designed Marketing Navigator - called Mark-eteer - an end-to-end campaign execution tool that draws on CRM data, content libraries, sales information, and prescribing data, and is structured to comply with promotional policy and GDPR from the outset. The conference engagement group produced Harmonia, which records live professional conversations, cross-references the profiles of the people involved, and generates personalised follow-up communications – with consent capture built in as a structural requirement. The commercial business strategy team built Sally, designed as a general manager’s strategic briefing tool that draws together financial planning data, competitive intelligence, and external market signals to explain the factors behind commercial numbers. A targeting and segmentation team designed Susanna, focused on identifying the right healthcare professionals for engagement using internal engagement data, external profiling and social listening. And the key account management group developed Kammie, a growth opportunity identification agent for KAM representatives, accessible by voice, via CRM, or via Microsoft Teams, with careful controls over access to sensitive pricing data.
Several things were consistent across all five designs. Every group made compliance and guardrails a structural feature from the start, treating data access questions – who sees what, under what conditions – as design principles rather than constraints to solve once a capability was proven. And every group anchored their work in commercial or customer-facing functions, even though the evidence presented in the session’s closing section suggested that the stronger returns from AI adoption tend to sit elsewhere in the organisation, as per MIT research.
That closing section drew on a study from MIT examining why large enterprises have failed to generate measurable returns from AI investment. Smith described the phenomenon as “pilot purgatory” – a state many organisations will recognise. The study identified five contributing factors. The primary obstacle to adoption is data integration within existing systems, not model quality – a finding that landed with some weight in a room where every team had imagined drawing on multiple data sources as though the connections were already in place. Investment tends to cluster in sales and marketing, where the ROI case feels intuitive, while back-office and business process functions – which the data suggests offer greater returns – face more compliance concerns and attract fewer pilots. Organisations building their own language models are achieving results more slowly than those deploying available commercial capabilities, a significant point given the resources some companies have committed to proprietary development. Adoption does not follow automatically from deployment: the use case has to generate enough perceived value by including impacted users earlier on, which drives the behavioural change required for adoption. And measurement is made harder by what might be called shadow AI – employees using consumer tools outside company systems to handle parts of their jobs - which widens the gap between real productivity gains and anything the organisation can formally track.
The design exercise was calibrated to that last problem as much as any other. Participants were not asked to assess AI strategy in the abstract, or to evaluate vendor claims against a procurement checklist. They were asked to think through a specific problem in their specific context and sketch what a working solution would look like. The gap between that sketch and a deployed system – with its data plumbing, governance requirements, adoption questions, and measurement challenges – is precisely where the industry’s current difficulty lies. As the session made clear, working out what an agent should do is, by some distance, the easier part of the job, which requires the architects to think through those same elements of the underlying complexity to be designed for from day 1.
For more insights on commercial pharma, including latest articles, interviews and more – click here.
