Pierre Legrand · AI enablement
Pierre Legrand
Pierre LegrandAI enablement and adoption · Melbourne

Adoption is no longer the constraint.

Most AI use is one person helping themselves with one task, and it never leaves their desk. A few people cross into building something others rely on. The usual response is to applaud them. Deploying them is how a small function reaches an organisation it could never physically get to.

Everyone using AI, each on their own. Most organisations measure this picture and call it progress.
43% or 12%?

One government survey says 43% of Australian SMEs are adopting AI. The ABS counts 12% of businesses using it. Plenty are trying AI. Far fewer are getting anything back.

See the data on AI adoption

What I do

I build the AI experiences people actually use, and the standards that keep them working after launch. In practice that means sitting with a team, finding the friction in the job they really do, building the thing that fixes it, and staying until it has been handed over.

Two rules govern everything I ship. An agent has to live inside the tool people already have open, because the moment somebody stops, opens a chatbot somewhere else and pastes the answer back, adoption quietly dies. And anything only I can run is a dependency rather than a solution, so nothing is finished until it belongs to somebody else.

A third rule covers the arithmetic. Models are unreliable with numbers, so the model never produces a number that has to be correct. A deterministic engine does that, with an audit trail behind it.

Selected work

In production, owned by others

Work breakdown quality agent

A manual quality review replaced by an agentic workflow: a Copilot Studio agent orchestrated through Power Automate, scoring delivery items against a defined set of quality criteria, writing to SharePoint and surfacing through Power BI. It now runs as an ongoing diagnostic, showing where work was being broken down poorly so coaching went to the teams that needed it.

My first version was a standalone application. It worked, but nobody else could maintain it, so I rebuilt it on tools the organisation already licensed. Other people own and maintain it now, and they decide what to improve next.

Live · 18december.com.au

18December

A commercial platform I designed and built end to end with Claude's agentic coding tools, guiding Australian families from a terminal diagnosis through to settling an estate. Next.js on Vercel with Supabase, Clerk and Stripe. I own the commercial model, the pricing, the vendor relationships and the responsible-AI guardrails: hard checks for privacy, cultural safety and clinical accuracy that fire before anything publishes, on a platform holding wills, medication records and final wishes.

Independently assessed and certified to CyberCert Gold. Built on a context architecture I designed, where an orchestrator reads a task and routes it to specialist workers with defined trigger conditions. Have a look.

Two day prototype

Advisor quoting assistant

Built in two days on Copilot Studio, Power Automate and SharePoint, so an advisor could give an accurate answer while still on the call. The agent ran the conversation and a deterministic engine produced the number, because a quote has to be right every time.

It didn't go ahead. The lesson I took: a working prototype is a persuasion tool and carries no mandate of its own. My first question now is who has to say yes, and what would make it easy for them.

Leadership summit

Leading AI, panel

Invited to speak on a Leading AI panel at a leadership summit, to a room of 450 leaders, on moving an organisation from individual experimentation to capability it can keep, and on where human judgement has to stay.

The longer view

The conversation almost nobody has started

AI in most organisations is a twelve month conversation about which tools, which licences, which training. The one that matters sits three to five years out, and it is about shape.

Most structures were designed for a world where every step needed a person. Handoffs between teams. A layer that checks the layer below. Roles that exist to carry information between functions that could not talk to each other. That design was correct when it was drawn. What has changed is that fewer of its steps now need a person in them.

Put agents into that shape and you get a faster version of the same shape. Automation creates capacity, and capacity is only worth something if you point it somewhere.

The useful questions are structural and slow. What does a role look like when the first pass is already done, and done consistently? Where does review belong once you are checking exceptions? Which team boundaries exist only because a handoff used to need a person? Who owns an agent when it is running, and what happens when it is wrong?

You can buy a tool in a quarter. You cannot redraw an organisation in one.

Background

Two decades across insurance, superannuation and financial services, beginning as an analyst programmer, with six of those years consulting inside client environments. People leadership along the way, including a team of 13 analysts, an acting Head of Digital Delivery post, and five years running my own company.

I read organisations vertically, at enterprise, portfolio and team level, because the people signing the cheques are rarely the people doing the work. Production work runs on the enterprise stack: Microsoft 365, Copilot Studio, Power Platform, Atlassian Rovo and Jira. I build with Claude and Claude Code every day, in Node.js, React and Python.

Get in touch

Melbourne based. If your organisation is working on any of this, I would be glad to hear from you.