Agentic AI — Everyone's Talking About It. But What Does It Actually Mean for Your Finance Team?
I got an email from KPMG last week about agentic AI. Which isn't unusual — the Big 4 have been flooding inboxes with AI thought leadership for the past two years. But this one was different. They weren't talking about AI as a future possibility. They were announcing actual products, actual partnerships, and an actual view that 2026 is the year agentic AI moves from pilot to production at enterprise scale.
It made me sit down and think seriously about what this actually means — not for a Fortune 500 company with a dedicated AI transformation team, but for the finance professionals I work with every day: CFOs and Heads of Finance at Australian SMEs, NDIS providers, and NFPs who are genuinely interested in AI and genuinely wondering where the line is between marketing and reality.
This post is my honest take. Not hype, not dismissal — just a CPA who's spent the last two years actively building and using AI tools trying to answer: is agentic AI something we should be paying attention to right now, or is it still a developer's game?
What Agentic AI Actually Means (In Plain English)
💡 The One-Sentence Version
A regular AI tool answers your questions. An AI agent does jobs — autonomously, across multiple steps, connecting to your systems, making decisions along the way, and handing results back to you when it's done.
The analogy I find most useful: think about the difference between a really smart colleague you can ask questions, versus a capable team member you can delegate a whole task to and trust them to figure out the steps.
When I ask Claude to summarise a compliance document, that's AI as a tool. When an AI agent monitors my accounts payable inbox, identifies invoices due in the next 7 days, cross-references them against the approved supplier list, flags any that need additional approval, and sends me a daily summary — that's agentic AI. It's not responding to prompts. It's working through a process.
The technical building blocks are: a reasoning model (the "brain"), memory (so it retains context across steps), tools (so it can connect to external systems like your accounting software or email), and a runtime loop (so it keeps working until the job is done or a human needs to step in).
What KPMG and the Big 4 Are Actually Doing
The KPMG email I received connects to a broader pattern that's worth understanding — because it tells you something about where the technology actually is versus where the marketing wants you to think it is.
KPMG acquired the team behind PrivateBlok in February 2026 — described as a "premier AI development platform" focused on multi-agent systems for finance, due diligence, and change management. They've also formed a strategic partnership with Uniphore to build AI agents for regulated industries, with a focus on governance, compliance, and integration with enterprise data environments. And they've published research estimating that agentic AI could unlock $3 trillion in global corporate productivity.
Deloitte, PwC, and EY are running similar plays — all racing to productise agentic AI capabilities for their enterprise clients, all emphasising governance and compliance as the differentiator for regulated industries.
Here's what I take from this:
- The Big 4 are building products, not just writing papers. KPMG now has an agentic AI product team. That's a material signal that enterprise demand is real.
- The target market is large enterprises. KPMG's Pulse Survey sample is C-suite leaders at organisations with $1 billion+ in annual revenue. When they say "99% of companies plan to deploy agents," they're not talking about a 50-person NDIS provider in Western Sydney.
- Governance and compliance are unsolved problems. The emphasis on "governed, monitored, secured" agents tells you that even at enterprise scale, trust and control remain the hard problems. These aren't solved at the SME level either — but the SME version of the problem is actually more manageable.
What's Actually Real for Australian SME and NFP Finance Teams
Let me separate three distinct things that all get called "agentic AI" but require very different levels of technical capability to use:
Level 1 — AI Assistants With Extended Capabilities (Available Now, No Developer Needed)
Claude, ChatGPT, and Gemini now all have features that let them browse the web, execute code, read documents, and take actions in connected tools — within a single conversation. This is the entry point to agentic capability, and it's available on paid subscription tiers today.
For a finance manager, this means: upload your payroll report and your roster, ask Claude to reconcile them and flag discrepancies. Or give it your bank statement and your accounts receivable ledger and ask it to identify unpresented items. These are agentic in the sense that the AI is doing multi-step reasoning and taking actions — but you're still in the loop for each session.
Level 2 — Low-Code / No-Code Agent Builders (Available Now, Moderate Learning Curve)
There is now a category of visual workflow builder platforms that support AI agent logic — connecting triggers, AI reasoning steps, and actions without requiring you to write code. A finance manager with reasonable technical confidence can build a workflow that monitors a specific email folder, extracts invoice data, checks it against a supplier list, and produces a daily summary for review.
This is genuinely accessible to a motivated non-developer. I've built functional workflows at this level. The learning curve is real but not prohibitive — and the gap between "interesting demo" and "reliable production workflow" is where most of the work actually sits. Data quality, exception handling, and human review design matter as much as the automation logic itself.
Level 3 — Production-Grade Multi-Agent Systems (Still Developer Territory)
What KPMG and the Big 4 are building — orchestrated systems where multiple specialised agents collaborate, with governance layers, audit trails, integration with enterprise ERPs, and production-grade reliability — this requires engineers with significant AI infrastructure experience. The code-first frameworks used at this level assume programming fluency as a baseline, and the governance requirements alone represent a substantial engineering project.
KPMG's email to me this week promoted their KymX platform — described as a "governed, KPI-driven AI platform" — alongside a framework called TACO for navigating different agent types. It's a polished enterprise product designed for organisations with dedicated IT teams and transformation budgets. It's not where most of us are starting from.
For most SME and NFP finance teams, Level 3 is not the current opportunity. Level 1 and 2 are.
⚠️ The Honest Reality Check
Despite the marketing volume around agentic AI, only 11% of organisations have actually put agents into production — even though 99% say they plan to (KPMG, 2026). The gap between intention and execution is enormous. The barriers are data quality, governance, and the gap between a promising demo and a reliable production system.
This doesn't mean you should wait. It means you should start at Level 1 and 2, build confidence with real use cases, and not let the enterprise marketing create unrealistic expectations about what you can deploy in a weekend.
What Finance Teams Can Realistically Build Right Now
Based on what I've built and seen work, here are the agentic-adjacent use cases that are genuinely within reach for an NDIS provider, NFP, or SME finance team today — without an enterprise IT budget or a developer on staff:
| Use Case | Still Requires Developers ❌ | Accessible Now ✅ |
|---|---|---|
| Payroll reconciliation | Fully automated end-to-end, integrates live with payroll and rostering systems, flags exceptions before anyone wakes up | Upload exports, use AI to reconcile and flag discrepancies, human reviews exception list before action |
| Invoice processing | Agent monitors inbox, extracts data, validates against approved suppliers and POs, routes for approval — autonomously | Structured workflow monitors a designated folder, extracts key fields, populates a tracking register for human review |
| Compliance monitoring | Agent monitors regulatory feeds in real time, identifies relevant changes, updates internal policy documents automatically | Scheduled AI review of ATO, Fair Work, or NDIS updates — summarised and flagged to the finance manager for action |
| Management reporting | Agent pulls live ERP data, generates variance commentary, distributes to stakeholders on schedule — without human touch | Structured prompts generate reporting commentary from data you provide — with your review, judgement, and sign-off |
| Cash flow monitoring | Agent monitors live bank feeds, flags anomalies, updates rolling forecast automatically | Regular data export into a structured template, AI flags items needing attention, human confirms and acts |
The pattern is consistent: the accessible version keeps you in the loop at key decision points. The developer version runs without you. Both add real value — the question is where your data readiness, process maturity, and technical capacity currently sit.
The Security Question Nobody Talks About Enough
The conversation around agentic AI tends to focus on capability — what it can do, how fast it can do it, what the productivity gains look like. The security conversation gets much less airtime, and that's a problem.
A few things worth having on your radar:
Data Privacy and Model Training
Any time you're putting work-related data into an AI tool, you need to know where that data goes. Free and lower-tier AI subscriptions typically allow the provider to use your inputs for model training. For casual personal use, that's a reasonable trade-off. For finance data — payroll figures, supplier details, management accounts — it isn't. Enterprise or business plans from major providers typically include data protection provisions, but you need to verify this for each tool you use, not assume it. And if you're building automated workflows that pull data from your systems on a schedule, the data governance question applies to every step in that workflow.
The Vibe Coding Risk
One pattern that's emerged with the rise of accessible AI-assisted coding is what's being called "vibe coding" — where non-developers use AI to generate functional code quickly, often without fully understanding what the code does. The productivity gains are real. The risk is equally real: AI-generated code can contain security vulnerabilities, hard-coded credentials, inadequate input validation, or dependencies with known issues — and a non-developer reviewing it may not spot any of these. There have been documented cases of vibe-coded tools inadvertently exposing sensitive data or creating exploitable access points.
This doesn't mean you shouldn't build tools with AI assistance. It means you should be deliberate about what data those tools touch, who can access them, and whether anyone with security awareness has reviewed the code before it goes anywhere near production data.
Agent Permissions and Access Scope
An AI agent needs access to your systems to do anything useful. The principle of least privilege applies here exactly as it does in any IT environment: the agent should have access to what it needs for its specific task, and nothing more. An agent that processes invoices doesn't need access to payroll data. An agent that monitors compliance updates doesn't need write access to your accounting system. Scoping access carefully before deployment isn't paranoia — it's basic governance, and it's the same discipline that makes finance controls work.
⚠️ The Bottom Line on Security
The security maturity required to safely deploy agentic AI is higher than most SME and NFP finance functions currently have in place. That's not a reason to avoid the technology — it's a reason to move deliberately, start with low-risk use cases, and make sure your data governance thinking keeps pace with your automation ambitions.
If an AI tool or workflow would be embarrassing or damaging if it exposed the data it touches, that's your signal to think carefully about access controls, data minimisation, and whether you have the right subscription tier before you build.
Where PFL Fits in This
I've been building finance automation tools for the past two years — starting with vibe coding simple Python scripts, progressively working with more sophisticated workflows. PFL sits at the intersection of finance expertise and practical AI capability: we understand the finance process deeply enough to design automation that's actually useful, and we have enough technical fluency to build tools at Level 1 and 2 without needing a developer.
What we can help with:
- Assessing where agentic automation genuinely makes sense for your finance function — and where the data and process readiness isn't there yet
- Building Level 1 and 2 automation tools configured to your specific workflows — payroll reconciliation, exception reporting, compliance monitoring — not generic products
- Designing the human-in-the-loop governance layer that makes automation trustworthy, not just fast
What we're not going to promise: that we can deliver the kind of fully autonomous, enterprise-scale multi-agent systems that KPMG is building for large enterprises. That's Level 3, and it's not where the opportunity is for most of us right now.
💡 The Question Worth Asking
Before asking "how do I get agentic AI?" — ask "what is the single most tedious, well-defined, repetitive task in my finance function that currently requires a human?" That's your starting point. A well-designed automation for one specific, real problem will deliver more value than a sophisticated agent architecture built around a vague aspiration to "automate finance."
The organisations getting real value from agentic AI right now are the ones that started specific and got it working — then expanded. Not the ones that tried to build everything at once and got stuck in a governance review.
Interested in what Level 1 or Level 2 agentic automation could look like for your specific finance function? PFL is happy to have that conversation — no enterprise budget required.
Talk to PFL →- KPMG — Q4 AI Pulse Survey: Enterprises Shift from Experimentation to Production (January 2026)
- KPMG — PrivateBlok Agentic AI Team Acquisition (February 2026)
- KPMG — Agentic AI Untangled: Build, Buy or Borrow Framework
- Anthropic — Building Effective AI Agents
- World Economic Forum — How AI Agents Can Become Strategic Partners (January 2026)
- McKinsey — How Finance Teams Are Putting AI to Work Today
Comments
Post a Comment