When you sit down with KPMG to talk about AI agents, you're not talking to an organization chasing buzzwords. You're talking to a firm that's re-engineering its own delivery models with agents while advising some of Australia's largest organizations on how to do the same.
Two people sit at the center of that story: Levi Waters, Partner in Charge of AI Solutions, and Lachlan Hardesty, Director of the AI Lab in KPMG Futures. Together, they're a barometer for where serious enterprises actually are with AI agents—and where they're heading next.
Beyond Pilots: The ROI Imperative
When I ask Levi what's dominating conversations in Australian boardrooms, he doesn't hesitate: return on investment has become the dominant theme.
Most large organizations have already run pilots. The question isn't "Can we get an agent to do something?" It's: "How do we get from proof-of-concept to scale—and show clear ROI?"
He's blunt about one thing: the "all pilots fail" narrative doesn't match what KPMG is seeing. They're seeing real deployments, real scale, and increasing benefit—when foundational work is done properly.
The organizations that are winning:
- Get their data into the right format, or use AI to help prepare it
- Expose systems via micro-APIs so agents can plug into real operations
- Standardize processes so AI can be trained on how the organization actually works
- Obsess over adoption, not just experimentation
The ones that stall spin up technically impressive pilots nobody uses, underinvest in data and documentation, and refuse to tackle the operating model changes required to make agents part of how work gets done.
Why KPMG Went Deep on Agents Early
Long before "agentic AI" became a marketing tagline, Lachlan's team was experimenting with open-source frameworks like CrewAI and AutoGen. Those tools had power, but they had a fundamental problem: they took too long to get to a client-ready, commercially assessable outcome.
Enterprise clients don't want to wait weeks or months to find out if something is viable. They want to know quickly: Is this worth backing?
That's where Relevance AI came in. Lachlan's team discovered they could use our platform to prototype agents rapidly, build multi-agent systems with low-code, and assess commercial value fast.
One of their earliest test cases: a KPMG-incubated aged-care scheduling startup. Lachlan's team used Relevance to build the agent solution end-to-end and delivered six pilots into home care clients using that stack. It wasn't a demo. It was operational.
That experience proved something important internally: agents weren't just a research toy. Used properly, they could underpin real products and services for paying customers.
From Digital Strategy to Workforce Strategy
Internal adoption forced KPMG to rethink how it designs AI strategies for clients.
Lachlan puts it simply: AI strategy, digital strategy, and workforce strategy are now the same conversation.
Instead of just asking "Which system are we implementing, and how do we train people on it?" you have to ask: "Which parts of our workforce should become agents—and what does that organization look like?"
That leads to different questions:
- Do you want to be an agent manufacturing house for your own organization?
- Do you primarily want to integrate third-party agents?
- Are you going to build an internal platform or library where agents live?
- Or are you going to play whack-a-mole with individual use cases?
Lachlan's experience is clear: when companies don't decide what kind of "agent organization" they want to be, their investments fragment. They don't know which use cases to back, how to prioritize, or what all the pilots are actually building toward.
The Accuracy Problem Nobody Was Ready For
One uncomfortable truth KPMG is surfacing: most enterprises never had to define what "accuracy" meant for white-collar work. Now they have to—and they're not ready.
It's easy to get output: draft a report, summarize a document, generate a call note. What's hard is agreeing on accuracy:
- What does "good enough" actually look like for this process?
- What's the acceptable error rate?
- Which parts must be perfect, and which can be approximate?
- Why does this process exist in the first place?
Agents have forced companies back to first principles. To automate a process, you have to articulate the process itself, articulate why it exists, and define what "good" looks like in a way an agent can be evaluated against.
That work was always latent. Agents just exposed how often it hadn't been done.
Governance and Monitoring: Where We Fit
As Levi puts it, KPMG will always take governance, controls, and risk seriously. That's the firm's DNA. But the part Levi is most focused on isn't just the policy layer—it's the runtime layer: monitoring what agents actually do, and whether their output matches the original intent.
That's where Relevance has become a key part of their toolkit:
- Vendor agnostic – They don't have to bet the future on a single LLM that will be leapfrogged next month
- Avoids tool sprawl – Instead of dozens of niche AI tools with separate licenses, they use a horizontal platform across functions
- Controls, monitoring, and visibility – Exactly what large enterprises and their auditors care about
- Configurable governance – They can align agents with their own standards and those of their clients
Levi's team has gone further and used Relevance as an educational tool. Leaders build agents themselves, see the building blocks behind "agentic AI," and get the aha moment of understanding what they're buying and deploying.
Standardizing how they use Relevance has had a practical effect: onboarding for new people has gone from weeks to hours. When you train AI on how you work and codify those patterns, new hires can ramp by collaborating with agents instead of reading through tribal knowledge.
The Adoption Gap: Waiting Is Dangerous
Levi recently spoke with a client who told him they were "two years away from Copilot." His view is unambiguous:
Organizations that wait for AI to "settle" are making a category error—it's not going to. Those that are already adopting and iterating will improve at an accelerating rate. Those that don't adopt will remain static.
The gap between those curves—accelerating versus flat—becomes a structural competitive disadvantage.
The goal is creating a network effect of adoption:
- Leaders use AI personally and are overt about it
- Teams feel safe to adopt and share patterns
- Organizations build feedback loops and usage metrics around agents
- Every new use case benefits from the ones that came before
That compounding effect is what KPMG is creating for its clients—and what late adopters will struggle to catch up to.
Looking Ahead: Agent-to-Agent Commerce
When Lachlan looks 12–24 months out, he's talking about B2B communication becoming agent-to-agent by default. Invoices, compliance checks, order updates, contract drafts—all the routine, structured communication between organizations will be generated by agents, read by agents, and logged with full traceability.
He sees something similar coming for sales: mass use of agents in outreach and procurement, a flood of bot-written messages where traditional outbound loses signal, and a world where agents not only sell but also evaluate and procure—on both sides of the transaction.
What KPMG's Journey Signals
KPMG's work with Relevance AI isn't about a single "hero use case." It's about a repeatable playbook:
1. Get clear on what kind of agent organization you want to be – Don't fund disconnected experiments
2. Start narrow and close to value – Pick a pain point with real economic weight
3. Use AI to get AI-ready – Don't wait years to "clean your data"
4. Define accuracy before you automate – Know what "good" looks like and why
5. Treat agents as part of the workforce – They need monitoring, maintenance, and support
6. Make adoption the KPI – Measure usage and embed agents in the flow of work
7. Lead visibly – Use AI yourself and normalize it
KPMG isn't treating agents as a novelty. They're treating them as an inevitable part of how professional services—and their clients' businesses—will operate. That's what it looks like when a global advisory firm moves beyond hype, does the hard foundational work, and starts using AI agents as a serious lever for how work gets done.
By Jacky Koh, Co-Founder & Co-CEO, Relevance AI
Watch the full interview premiering at Agents@Work, December 10 2025.
