This blog explores the themes from Episode 2 of Technically Speaking, Audacia's podcast for technology leaders navigating real decisions in complex organisations, in which Richard Brown is joined by Chris Barber, Chief Executive Officer of Lhasa Limited, and Philip White, Managing Director of Audacia, to discuss the impact of AI on the workforce.
A widening gap in adoption
The next step in AI adoption is for organisations to move away from manual AI processes to agentic ones. However, most organisations are at very different stages in the adoption process. According to OpenAI, Enterprise AI use is becoming more agentic, with frontier firms (those in the top 10% of OpenAI usage each month) generating '8.3× as many output tokens per active user as typical firms, up from 2.6× in January.' Additionally, some organisations are seeing real value from AI, whilst others are failing to take AI into live production, or to see a return on investment.
The distinction between the varying levels of success rarely comes down to the technology itself. It comes down to whether the organisation started with a defined problem or started with the ambition to "do AI" and worked backwards from there. Organisations adopting AI for fear of falling behind fall into the trap of picking up the nearest available tool and looking for somewhere to apply it. However, as in software projects, organisations achieving genuine value tend to be the ones that can answer key questions before they touch the technology. What is the AI tool being used for? What problem does it solve? What does success look like once it's built? Those struggling tend to be the ones still trying to answer them during or after the event.
Redesigning work
A common constraint of AI adoption is that the processes AI has been layered on top of never get redesigned.
For example:
- AI generates code at speed, but every change still needs review before release
- AI surfaces fresh insights daily, but decisions still only get made at the monthly steering committee
- AI helps generate vendor proposals or business cases quickly, but budget sign-off only happens at quarterly planning
- AI flags anomalies across thousands of transactions instantly, but each flag needs manual investigation by the same small fraud team
The underlying point is that improving a process built around human checkpoints will always be limited by human capability and time, however much AI is layered on top. The emerging conversations being had now are about redesigning how work is structured so that people aren't the default checkpoint at every stage.
That in turn means finding new ways to measure whether something is good enough. Speaking purely about technology, code coverage, line-by-line review and the other proxy's teams have relied on for years don't map cleanly onto a world where the code itself may not need to be read by a person at all.
Redesigning work to remove human bottlenecks raises the question: how will the workforce change? Previous waves of technological change, from the industrial revolution to the arrival of the internet, tended to be absorbed by a corresponding rise in demand rather than a simple reduction in the workforce required. Current figures from the World Economic Forum suggest that AI is currently creating roughly twice as many jobs as it displaces. Whether that continues is an open question. The scale of this shift could be different from what came before, and the answer may only become clear once organisations move from early adoption into full efficiency gains, rather than during the experimentation most are still in.
Trust, accountability and determinism
A more fundamental issue is that software has always been deterministic, whereas AI often isn't. That unpredictability means organisations should expect to treat AI increasingly as a black box. Instead of understanding exactly how a task gets done, teams will define clear inputs, clear outputs, and a way of measuring whether the result is correct and consistent. A useful small-scale comparison is an older shift in how developers work: memory management used to be a manual, closely guarded responsibility, and the arrival of garbage collection made it invisible to most developers without making software worse.
That shift raises questions about accountability. If a model is trained on data selected by one team, built by another team and deployed to act with increasing autonomy, where does responsibility sit when something goes wrong? There's no settled answer, but the response emerging is a move away from prescriptive, step-by-step rules and toward strong validation and guardrails: being explicit about what an output has to satisfy, rather than dictating exactly how a model gets there.
Change management for AI
Traditional change management assumes a period of stability once a decision is made, but that assumption doesn't hold in regard to AI. A CRM replacement a few years ago might have justified months of evaluation across several vendors, because the decision would stand for years. Applying that same pace to AI risks the tools and the surrounding thinking becoming outdated before the evaluation is finished.
The response to this change of pace is a well-defined, but short and adaptable, strategy document. Rather than a fixed plan revisited annually, a workable AI strategy sets out an organisation's appetite for risk, change and innovation clearly enough that decisions can be made quickly as circumstances shift. Alongside this strategy there needs to be consistent communication to the wider organisation through clear, ongoing conversation: people who aren't told where the organisation is heading tend to assume the worst about what AI means for their own role.
What organisations are hiring for now
Uncertainty and the pace of change are also reshaping recruitment. Where hiring decisions used to start with hard skills, specific languages or specific qualifications, the emphasis has shifted toward curiosity, adaptability and the ability to work well through change. Skills that felt central only a couple of years ago, prompt engineering among them, have already been overtaken by the time formal courses caught up with them. Culture may be increasingly placed at the front of the interview process rather than treated as a secondary consideration.
The bigger structural question is what happens to career progression as entry-level tasks become increasingly automated. If junior roles have traditionally been where people built the judgement to become effective managers, and those tasks are handed to AI, organisations are still working out how that experience gets built instead. There's a sense that some roles rooted in analysing context and processes, rather than data alone - such as Business Analysts - will remain valuable for longer than most.
What leaders can act on now
- Start with a clear, if brief, AI strategy document outlining appetite for risk
- Be explicit about the outcomes a piece of work is meant to achieve before choosing the tool
- Start defining what "good enough" looks like for AI-assisted work
- Settle accountability before deploying anything with autonomy
- Communicate direction consistently, even amid uncertainty
This was an episode of Audacia's podcast series, Technically Speaking.
Future episodes will explore the challenges shaping technology leadership today, whether that's navigating the governance questions that come with AI, rethinking what technology leadership looks like, or building the infrastructure and culture that makes it all possible.
Listen now on your preferred platform.
If there are topics you'd like to see covered, we'd welcome your input at [email protected]
