Generative AI reached businesses with unusual speed, but adoption alone does not create lasting value. Paul O’Sullivan, Salesforce’s senior vice president of solution engineering for the UK and Ireland, framed the moment as a choice between major productivity gains and equally serious questions about reliability, privacy and workforce readiness.
His argument places trust at the centre of deployment. A company can give employees a powerful assistant, but the system will remain peripheral if staff cannot rely on its answers or customers fear what happens to their data. The transformation of work therefore depends on management choices as much as model capability.
Rapid Adoption Raises the Productivity Ambition
O’Sullivan used ChatGPT’s rise to 100 million users in two months as evidence of how quickly generative AI entered public life. He contrasted that pace with Netflix taking roughly a decade to reach the same audience. The comparison helps explain why executives moved AI towards the top of their agendas: employees and customers encountered the technology before many organisations had established a formal strategy for it.
The economic promise is correspondingly large. O’Sullivan cited an estimate that generative AI could contribute £3.5 trillion to the global economy and said it might automate as much as 40 percent of an average working day. Automation in this context does not necessarily mean eliminating an entire role. It can mean drafting, summarising, locating information or preparing a first analysis, leaving a person to judge and refine the result.
That distinction matters because productivity is not measured by the number of generated words. A faster draft only helps when it is accurate, relevant and easy to verify. Organisations need to identify tasks where an assistant removes genuine friction rather than adding a new review burden. The most useful early deployments are likely to have clear inputs, observable outcomes and a person who remains accountable for the decision.
Data Quality and Privacy Shape the Trust Gap
O’Sullivan said that more than half of customers remained sceptical about the safety and security of AI applications. He connected that lack of confidence to familiar enterprise problems, especially poor data quality and hygiene. A model grounded in duplicated, incomplete or outdated information can produce polished answers that reproduce the organisation’s underlying disorder. Fluency may make those errors harder to notice.
Improving the model without improving the data is therefore an incomplete strategy. Businesses need to know which sources an assistant can access, who owns those sources and how corrections propagate. Representative data also matters because gaps and historical bias can lead to uneven service. Trust emerges when an organisation can explain where an answer came from and correct the system when it fails, not merely when a demo looks convincing.
Privacy creates an additional boundary. O’Sullivan warned against allowing large language models to store and learn from sensitive customer information. That requires more than a general promise of confidentiality. Teams need explicit rules for retention, access, model training and deletion, supported by technical controls. Employees should understand which information is permitted in a prompt, while customers should receive a clear account of how their data is used.
Human Skills Remain Part of the System
O’Sullivan advocated workforce upskilling, including the ability to write effective prompts and understand basic AI terminology. Prompting can help employees state a task, provide context and request a usable format. More importantly, it teaches them to treat the model as an interactive tool whose first response may need questioning, correction and evidence rather than as an unquestionable source.
The wider skill requirement is judgement. O’Sullivan described AI as a co-pilot rather than a replacement for human expertise, noting that current systems lack qualities such as empathy, emotional intelligence and ethics. Those capabilities become especially important in customer service, management and other contexts where a technically plausible answer can still be inappropriate. A person must interpret the situation and own the consequence.
Training should therefore be tied to real work. Generic awareness sessions may explain what a language model is, but employees also need examples of acceptable use in their own roles, criteria for checking outputs and a route for reporting failures. Managers should measure whether AI improves quality and cycle time without silently increasing risk or after-hours review. Upskilling is strongest when it combines experimentation with clear professional standards.
Responsible Adoption Is an Operating Model
The ethical issues O’Sullivan highlighted include hallucinations, embedded bias and toxic output. These are not one-time obstacles that disappear after procurement. Models change, internal data changes and employees discover new use cases. Responsible adoption consequently requires ongoing evaluation, documented ownership and the authority to restrict a system when its performance falls below the level a task demands.
A useful operating model separates low-risk assistance from consequential decisions. Drafting an internal meeting summary may permit a different error tolerance than recommending a financial action or responding to a vulnerable customer. Each use case needs controls proportional to its possible harm. Human review, access limits, logging and testing should be designed around that context instead of applied as a single universal checklist.
Transparency also supports adoption. Leaders should say what the system can do, what it cannot do and how employee performance will be assessed when AI is involved. If staff suspect that productivity tools are primarily surveillance or an undeclared route to job cuts, they have little reason to share useful feedback. A collaborative implementation gives workers a role in finding benefits and spotting risks.
O’Sullivan’s view of the generative AI workplace is optimistic, but it is not automatic. Fast adoption and large economic estimates establish the scale of the opportunity; they do not prove that every deployment deserves trust. Businesses must connect technology with clean data, privacy protection, practical training and human accountability.
The most durable productivity gain may come from designing that entire system well. When employees know how to use AI, customers know how their information is protected and leaders respond openly to limitations, a co-pilot can become a dependable part of work. Without those conditions, even a capable model remains an impressive tool searching for institutional permission to matter.
