Empathetic AI | June 2026 | 6 min read
Anthropic recently made its AI Capabilities and Limitations course freely available through Anthropic Academy. It is a short course, around two to three hours, and it is one of the more honest pieces of educational content I have seen from a major AI lab. Rather than selling capability, it teaches a mental model for when AI will fail and why. Finance professionals should pay attention, because the failure modes it describes show up constantly in our domain.
The four-property framework
The course organises everything around four properties that shape what a large language model can and cannot do. Each one sits on a spectrum from capability to limitation. The practical insight is that in real-world tasks these properties interact, and the interactions are where the problems live.
Next Token Prediction. AI generates responses one token at a time, selecting what is statistically most plausible next. It is extraordinarily good at producing fluent, structured text. It is not actually reasoning from first principles.
Knowledge. The model knows what was in its training data, up to a cutoff. It can be confidently wrong about facts that have since changed, and it has no real-time access to regulatory updates, rate changes, or rulings.
Working Memory. The model works within a context window. For long documents, complex reconciliations, or multi-step workflows, content at the edges of the window gets less attention. This is not laziness. It is a structural property.
Steerability. AI responds to instructions, but imprecise instructions produce imprecise outputs. The quality of your prompt directly caps the quality of the result. This is why prompt engineering is a real skill, not a workaround.
Where this lands in finance
These are not abstract properties. In finance and accounting work, each one shows up with specific consequences. The course frames them generically. Here is how they play out daily across our work with tax accountants, CFOs, and finance teams across APAC.
The fluent hallucination problem. AI will produce a tax calculation, a GST treatment, or a regulatory citation that reads with complete confidence and is structurally correct. The numbers may be wrong. The section reference may not exist. The model is optimising for what sounds right, not what is right. In finance, a confident wrong answer is worse than no answer. This is not a bug to be fixed. It is the underlying mechanism. The correct response is to verify outputs against primary sources before any client or compliance use.
Tax law changes faster than training data. ATO guidance, IRAS circulars, GST rulings, and transfer pricing safe harbours change regularly. A general-purpose AI model has a knowledge cutoff and will apply prior rules with complete confidence, unaware that they have been superseded. This is why institutional relationships matter more than model capability. Our partnerships with IRAS, ATO, CPA Australia, and CA Singapore allow us to ground AI outputs in verified, current regulatory materials rather than relying on what was in the training corpus. The model knows a lot. It does not know what changed last quarter.
Long reconciliations lose the thread. A month-end close package, an audit workbook, or a multi-entity consolidation is exactly the type of long document that stresses a model’s context window. Errors introduced earlier in a document tend not to be caught at the end. Dependencies between entries get missed. The workflow design matters as much as the model capability. Breaking long tasks into bounded steps, with explicit checkpoints, is not a workaround. It is how you use the tool correctly.
Vague instructions produce expensive ambiguity. Ask an AI to “review this contract” and you will receive a general answer. Ask it to “identify clauses that create exposure under Australian consumer law for a B2B SaaS contract, and flag any non-standard indemnity terms” and you will receive something useful. Finance professionals who invest in prompt design get materially better outputs. This is a transferable skill, not a technical one.
When properties collide
The Anthropic course dedicates its final section to what happens when all four properties interact simultaneously. In practice, every real finance task triggers multiple properties at once. This is where general AI tools struggle and domain-specific design becomes essential.
A tax position memo for an IRAS filing puts stale knowledge in tension with a confidently fluent tone. The result can be a persuasive document citing superseded guidance, presented as current. A multi-entity reconciliation review puts a long document against a vaguely worded task, and errors at the end of the document get missed entirely. A cross-jurisdictional transfer pricing analysis asks the model to hold both ATO and IRAS rules simultaneously across a complex input, and jurisdiction-specific nuances quietly drop out of attention. FY close variance commentary asks the model to explain drivers it cannot actually see, so it produces a technically coherent narrative that does not match what happened in your business.
The honest framing is this: AI is not a junior analyst who can be given work and trusted to flag uncertainty when it arises. It is a sophisticated drafting and pattern-matching tool that requires a human with domain expertise to set the scope, verify the outputs, and own the judgement. The firms getting real value from AI in finance are the ones who have internalised that distinction and designed their workflows accordingly.
What this means for how we build
Empathetic AI was built around a single insight: the finance domain has requirements that general-purpose AI cannot meet out of the box. The Anthropic framework validates this. Knowing the four properties is the starting point. Designing for them is the work.
That means grounding knowledge in live regulatory sources rather than training data, structuring workflows to respect context window limits, building prompt frameworks that make steerability repeatable rather than dependent on individual skill, and maintaining human review as a structural part of the process rather than an optional extra.
It also means being direct with clients about what AI does and does not do. The hype cycle has created unrealistic expectations in both directions. Finance professionals who understand these four properties will be better placed to deploy AI in ways that actually reduce risk rather than obscure it.
Anthropic’s AI Capabilities and Limitations course is free, takes around three hours, and requires no technical background. For finance professionals who want a grounded starting point rather than vendor-produced enthusiasm, it is worth the time. We recommend it to every team we work with as foundational reading before deploying AI in any compliance or client-facing context.
Source: Anthropic Academy, AI Capabilities and Limitations course (anthropic.skilljar.com), accessed June 2026. Available free with no Anthropic account required.
Leave a comment