EP15: Agentic AI & Societal Shifts: Balancing Speed, Ethics & Impact with Sagittal

An AI Team Member, Not Just Another Tool

Sagittal AI's core product treats AI as a genuine team member within a software development team — spanning designers, engineers, testers, and QA — integrating directly with the tools teams already use, including Jira, GitHub, Notion, Figma, and Google Docs. Rather than working in isolation, the AI (which the team calls Neo) gathers context from existing systems just as a human colleague would, completes assigned tasks — such as turning a Jira ticket into a full pull request — and puts the results back into those same tools, following an organisation's documented coding standards and best practices.

Michael is clear that this isn't about full automation: human judgement remains essential, with AI lifting the more mechanistic, repeatable elements of the workflow.

Why Hybrid Teams, Not Just Individual AI Tools, Unlock the Real Value

One of the most interesting insights Michael shares is around adoption dynamics. Early AI adoption tends to happen at the individual level — a single enthusiastic user picking up a tool and experimenting. But the real transformative value, he argues, comes from embedding AI between team members, rather than simply speeding up each person's individual work in isolation. An AI positioned as a genuine collaborative team member can significantly reduce overall project time — not just individual task time — while also helping teams better understand and articulate their own unique human value: what only a person can bring, versus what AI can now handle.

The Technical Challenge: Making AI Fluent Across Existing Tools

Sagittal AI deliberately focused on software development because, as Michael explains, the tools across the entire development lifecycle — from product ideation through to QA and deployment — have spent over a decade becoming highly API-accessible, giving AI systems a realistic path to fluency. He contrasts this with the informal, ad-hoc use of the Model Context Protocol (MCP) many developers currently rely on — often running an unsecured local process alongside a codebase downloaded to a personal laptop, with no clear tracking of what was AI-generated, or where security exposure has been introduced.

Sagittal AI's alternative approach is to bring AI formally into the organisation's existing tools and access controls, rather than working around them — a model Michael believes is reproducible in other team-based domains like legal review, customer success, or marketing, though he notes those fields currently have less mature, less accessible APIs to build on.

"Adapting AI to Humans" — Not the Other Way Around

A central philosophy at Sagittal AI, reflected in one of its taglines, is adapting AI to humans, rather than retraining organisations to fit around AI. Michael argues that decades of refinement have gone into designing existing team workflows to bring out the best in human collaboration — and that it's more effective, and less disruptive, to teach AI to meet people within those existing systems than to force organisational change to accommodate the technology.

Designing for AI's Quirks: Sycophancy and Context Management

Michael candidly discusses some of the practical challenges of deploying AI as a team member — including a tendency toward excessive sycophancy, where the AI misinterprets a simple "thank you" as a request for further, unnecessary work. Sagittal AI had to build explicit classifiers to detect genuine follow-up requests versus polite acknowledgement.

On memory and context, Michael describes Neo's role as primarily an orchestrator — carefully sequencing what information is fed into a large language model at each stage of a task, rather than simply dumping an entire codebase and documentation set into a massive context window. He notes a key limitation of LLMs: there's a meaningful difference between finding "a needle in a haystack" and finding "a needle in a needle-stack" — when an LLM is given too much relevant, unfiltered information at once, output quality can deteriorate quickly.

Security and Accountability: Treating AI Like a Contractor

On the growing security risks of ad-hoc AI integration, Michael highlights a common blind spot: organisations that would never skip an architecture review for a new system often fail to properly map data flow when simply adopting an AI subscription informally. Sagittal AI's approach is to treat its AI as a distinct user and "contractor" within an organisation's existing access control systems — using the same authorisation and token-based permissions already applied to human staff, rather than granting AI systems access "as" a specific person, which he warns creates serious accountability problems.

Because Sagittal AI is embedded within existing software development lifecycle processes — including mandatory human code review — AI-generated work retains clear attribution, with humans required to explicitly validate AI contributions, just as they would a colleague's.

The Risk of Blind Trust in Improving AI

Michael raises a subtler, longer-term security and governance concern: as AI systems become more capable and reliable, humans may become increasingly likely to accept their output without proper scrutiny — citing real-world examples of lawyers mistakenly submitting AI-hallucinated legal citations to courts. He argues organisations must resist the temptation to abdicate control entirely to AI, and consciously identify which processes genuinely warrant a stringent human review step, even as overall AI accuracy improves.

Is AI More Like the Industrial Revolution Than the Internet?

Reflecting on AI's broader societal trajectory, Michael shares a striking reframing offered by one of his customers: rather than comparing AI's impact to the internet or mobile technology — which largely digitised and streamlined existing knowledge work — AI may be closer to the Industrial Revolution in scale, fundamentally redefining large swathes of work rather than simply making it more efficient. Unlike the Industrial Revolution's roughly 100-year transformation across two distinct phases, however, Michael expects AI's disruption to unfold far more rapidly, and across a broader range of jobs — making thoughtful, early consideration of governance and responsibility especially urgent, rather than something addressed only after significant societal disruption has already occurred.