Scottish tech conference ScotSoft enjoyed a packed agenda as its organiser, ScotlandIS turns 25. 

This year’s event, which took place during Scotland’s inaugural National Innovation Week, combined technical excellence with leadership insight through keynote sessions, breakout discussions, and networking opportunities.

One of the standout talks, centred around AI-driven workforces, saw Pete Gordon, Principal Consultant at Waracle, explore how LLM-based agents can become true teammates, not just tools.

His talk, titled Designing Your Ideal Agentic Employee shared a step-by-step design thinking approach for rapidly prototyping these AI agents (while stressing the need for strong governance and human oversight).

His message blended classic management wisdom with cutting-edge AI: extrapolating from a quote by business theorist W. Edwards Deming in which he said “the greatest waste in America is failure to use the abilities of people,” Gordon suggested we must update this to include “peoples and machines” in today’s workplace.

So, what does that look like?

Defining the AI “Agentic Employee”

Gordon proposed a simple, tech-centred definition: an AI agent is “an LLM plus a memory plus tools.”

In other words, these agents combine a large language model with long-term memory and actionable capabilities. The memory lets the agent “remember what they’ve done” and recall past interactions, while the tools enable it to act without prompt, based on its own ‘judgement’ – insofar as we can use the word in this context. 

In practice this means giving the AI access to services (APIs, databases, etc.) so it can carry out tasks, not just chat.

Crucially, Gordon emphasised that agents require clear roles and goals – each agentic ’employee’ needs a defined purpose, plan, and constraints that align with the organisation’s culture (and also keeps them on the right side of compliance/regulation).

For example, he explained, an agentic AI working for a compliance team would be programmed to flag anything that might break “the letter of the law” – and even consider “the spirit of the law” – which is something akin to approaching the outcomes that the law is meant to engender. Likewise, a marketing agent would focus on brand consistency and tone across all content.

Obviously, there’s nuance in this discussion around things like hallucinations, quality of data, prompt engineering and 100 other things that Gordon acknowledges.

Balancing Innovation and Risk

Broadening that out, a major theme of the talk was enterprise risk versus innovation.

Gordon noted the exponential explosion of shadow AI, with employees using the technology “under the radar” to shocking degrees.

Citing MIT’s landmark Project NANDA, State of AI in Business 2025, he revealed the rise of a “thriving shadow AI economy,” where staff use personal ChatGPT or other tools to automate parts of their jobs without IT approval.

That report revealed that roughly 40% of companies now subscribe to LLM services, but 90% of employees said they were using them in their work. As Gordon put it, the official tools provided by many firms are underserving users, prompting workers to hack together their own solutions (the “shadow IT” effect).

He warned that this gap has serious consequences – even consumer-facing AI features like the new ChatGPT Developer Mode – which can plug arbitrary Model Context Protocol (MCP) servers into ChatGPT – may expose companies to uncontrolled or malicious tools.

As he candidly noted, when anyone can “grab a URL off the internet” and connect it, “that’s really scary for the enterprise,” especially in regulated sectors.

Rapid Prototyping with Design Thinking

To bridge the gap between informal AI use and official strategy, Gordon advocated a design-driven discovery process – he outlined a compressed design thinking approach (akin to a week-long design sprint) to identify real pain points and build prototypes.

At Waracle, the team used workshops and “divergent/convergent diamonds” exercises to map out work processes and data flows. As Gordon recounted from a case study at a fictional firm “PG Investments,” the goal was to see which routine tasks could be automated and which needed human judgment.

They asked simple questions: “My tasks that I do daily – do they involve judgment? Or are they just processes?” Tasks that followed fixed steps were prime candidates for an AI agent, leaving humans free for more creative or sensitive work.

Mapping the customer journey revealed the real bottleneck.

In the example, the marketing team’s workflow was painstakingly drawn: create a message, send it through two layers of approval (compliance and marketing directors), publish it, then wait weeks for customer feedback.

With such a long feedback loop, any change took ages to validate. Gordon explained that this often forces teams into “big batches” of updates (to minimise trips through the loop) rather than trying small improvements.

To break this logjam, the team explored automating each role in the loop. Gordon described asking the room: could they “automate the head of marketing? The compliance? The customer?”

In effect, they proposed turning each into an agent: a compliance agent to vet legal and regulatory issues, a marketing agent to check brand tone, and even a customer agent to gather feedback.

Ensuring Trust: Governance and Testing

Gordon stressed that technical prowess alone is not enough – companies must also address accountability and trust.

He outlined the need for clear governance frameworks: assigning ownership for each agent’s behaviour, setting up oversight mechanisms, and planning maintenance.

Testing was a big part of this. The team’s QA engineers evaluated the agents on functional quality, robustness, performance and self-correction. They even tried to automate as many tests as possible, so the agents could be scored objectively. All this was to ensure the agents could be deployed responsibly, so that AI becomes an asset and not a liability.

Recommended reading

• OpenAI to Launch AI-Powered Jobs Platform in 2026
• Is the UK Falling Behind in the Race for AI Talent?
• PwC: Firms Now Pay a 56% Premium for AI Talent

In closing, Gordon restated his core message: people and AI should work together, not in silos.

“I think the greatest waste today is not utilising people and agents together,” he asserted, echoing Deming.

He warned that unless firms build AI safely, staff will continue to find risky shortcuts.

His challenge to the audience was clear: “Build your own work agent… try to see if you can extend yourself across the organisation.” By proactively creating internal agents with proper design and oversight, companies can harness those ideas and free up human talent.

As he concluded, a thorough discovery process – ensuring the data and context are in place – remains “the key to unlocking that path.”