It’s easy to talk about enterprise AI in the abstract of glossy slide decks, promises of 10x productivity, start-ups who work 2 hrs a day and earn 6 figures, and the race to get the newest flagship models into everyone’s hands.
But when you get senior leaders, IT directors, and enterprise practitioners around a dinner table for an honest conversation, the narrative shifts fast.
Last week, Cisilion brought together a mix of our customer business and technology leaders alongside our partner team from Paul Cawley / Microsoft and Matt Brown / Cisco to explore the realities of enterprise AI adoption.
Over a very enjoyable roundtable dining experience in London, what unfolded was a candid deep dive into jobs and skills, governance, compliance, human behaviour, agentic architectures, business transformation and the shifting definition and need around ROI, FinOps and ethical AI.
Here are the six defining takeaways topics that resonated through the evening:
Buying "AI" Licenses Isn’t an AI Strategy
Too many organisations (especially in the earlier days) treat AI adoption as a procurement checkbox: get budget, buy and assign licenses, deploy tools like Copilot, GPT Enterprise and wait for magic to happen.
People go to work and typically do a job – many don’t have time (or know what they are allowed to do) to learn, experiment, fail, and try again. Technology without structured change management, training and people looking at improving at how “things” get done is still not taken seriously enough if you want to drive real AI value and make that AI investment deliver value rather than just a cost on the P&L.
The best initiatives come from department leads, not IT. Paul Cawley shared valuable information from the Microsoft WorkTrend Index Report around the maturity curve that Adoption (of AI) typically follows:
for example, saving an hr a day drafting emails, reviewing documents, or creating a first draft
specialised agents tracking meeting actions, onboarding, or policy workflows
end-to-end multi-agent architectures running across systems, with humans governing exceptions
With technologies like Copilot Studio, Codex, GitHub Copilot and the emerging tools to really govern, control and secure our “agent ecosystem” [Agent 365 for example], these organisations are already rapidly moving rapidly from personal and team productivity tools into bespoke Agentic Design and Build.
The general view was that overall, most companies haven’t (yet) redesigned job roles around these new capabilities yet outside of development/coding roles.
Real transformation happens when you redesign the workflow, and re-define the job description – not when you hand out a new digital assistant or chat bot.
The Governance Pendulum: Security, Residency, and Shadow AI
Shadow AI
Restricting tools without offering an alternative can push employees towards unmanaged Shadow AI. Policies alone aren’t enough. Education and controlled environments, such as innovation sandboxes, can give employees freedom to experiment while keeping corporate data and IP protected.
Residency
For regulated organisations, data sovereignty is non-negotiable. Businesses need clarity on where data is stored and processed, particularly across the UK, EU and US, without resorting to complete isolation or restriction.
Security
The focus is on creating secure enterprise guardrails without restricting innovation. Tools such as Microsoft Purview, Defender, Cisco SASE and Zscaler can help, but organisations also need clear internal rules around what AI can access and what data can be shared.
Output Inflation, “AI Cheating,” and the Trust Gap
A major topic of discussion was the quality and authenticity of work. We’ve all seen it: a colleague generates a 10-page document or meeting summary with AI, hits send, and never actually reads what was created. Worse still, the receiver uses AI to summarise that doc and draft a response – how much human was really in that loop?
When content is inaccurate or bloated, organizational trust collapses – and the failure is often misattributed to the tech rather than a lack of human accountability. In parallel, AI is driving an escalation in expected output speed and volume. The same goes for building agentic flows or agents (especially for example HR, Customer Service etc). If the Agent/App Builder does not thoroughly test and train and coach the output and its wrong – people lose trust.
One member of the roundtable said that it’s actually worse than that. The tolerance for a human making a mistake is actually quite low. If an agent gets something wrong, it’s much lower but getting an “AI” to admin “I do not know” is much harder than it looks.
Responsible adoption requires teaching teams to distill, verify, and take true ownership of AI-assisted outputs.
What Happens to the Junior Talent Pipeline?
Often referred to as the “missing rung” of the career ladder this was also a hot topic. In sectors like insurance and law, multi-agent frameworks are stepping in to ingest and classify complex, 18-section documents, replacing manual assistant tasks. There are many junior/intern level roles that AI can just do out of box far better this is short-term blessing a longer-term problem, but the roots go way beyond what we hire for, it actually impacts human learning evolution, the school system and beyond (it’s ok – we did not get this deep into it, but the point was valid and resonated).
There were a few real examples explored over our starters and wine! While this approach (legal agents, Customer Service agents, etc) can create immediate capacity, extend hour coverage and enable existing assistants to step up into more decision-making roles, it introduces a vital long-term question:
“If junior administrative and review work is automated away, how will the next generation gain the foundational experience required to become senior leaders?
As we plan our future of work with agents and automated things, we must deliberately construct AI-accelerated apprenticeship and learning paths to protect the professional pipeline and make sure we know who is ultimately accountable.
AI Infrastructure, Tokenomics, and FinOps
One of the biggest questions right now is how we make the economics of AI actually work.
AI is an infrastructure and cost challenge just as much as an application layer and adoption challenge. Matt Brown at Cisco shared some interesting stats around the impact of multi-agent workflows on the network infrastructure. Unlike human workers, AI doesn’t generate the same intermittent traffic as human do – they produce continuous, machine-speed API calls that can surge network loads by over 400%.
Unmonitored token consumption can cause costs to outpace projections by four times. Deploying employee agents at scale demonstrates the power of a hybrid model: running the majority of inference on premises for predictable cost control, security, and GPU consolidation, while selectively leveraging cloud models. Enterprise AI requires disciplined AI FinOps to rightsized models and monitor consumption.
The Need for Continuous Strategic Road mapping
Given the dizzying pace of model updates, security enhancements, and licensing tiers, an AI roadmap cannot be a static annual document. Organisations need ongoing strategic advisory to cut through the noise, align with emerging capabilities, and adapt their architectures as the ecosystem evolves.
There are roadmaps, updates, new features, new rules, new prices, new governance, but it is hard to keep up. There were discussions around the room about (for example), early negative Copilot experience compared to say Claude and ChatGPT which triggered a fair bit of debate. Many in the room challenged those who had previous negative experiences of it (especially those paying for multiple tools) citing the huge advances in this space in last 9-12 months along with the simplicity of ensuring governance and security and ability to easily understand people and organisational data thanks to Work IQ.
This exposed the conclusion that this space is moving so quickly – it’s no longer just about models, it is about the connected eco system, understanding costs, architecture (what talks to what, what does that talk cost, what security and governance is needed, what third party tools are “still” needed) and above all to review regularly… This is a role for the PEOPLE not the AGENTS.
How AI Ready is your organisation?
What stage is your organisation currently navigating on the AI journey – personal productivity, multi-agent workflows, or infrastructure readiness?
