Does Your Client Actually Need AI?

Ask a room full of executives whether their organization needs AI, and almost everyone says yes. Ask what “need” means, and the case usually falls apart. A genuine need implies an essential, operational gap that only that tool can close. For most accounting and finance functions, AI rarely clears that bar, at least not first.

Large language models earn their strongest results where language, code, and technical reasoning are the work itself: software development, product design, and R&D, where a model can participate directly in producing the output. Accounting operations look different. The constraint is rarely a shortage of intelligence. It’s a shortage of discipline clean master data, standardized processes, consistent controls, and systems that reconcile. Research on generative AI’s economic potential finds the bulk of realized value clusters in customer operations, sales, software engineering, and R&D, not in back-office transformation pursued for its own sake.

That distinction should reshape how practitioners advise clients. The common mistake is starting with “where can we use AI?” A sharper question is: is this even a problem AI can solve? In finance and accounting, models too often sit downstream of data entered inconsistently, approvals with no standard, and account mappings that shifted without anyone tracking the change. NIST’s guidance on trustworthy AI stresses that reliability, transparency, and accountability depend on the environment a model is deployed into weak inputs and weak controls simply pass their weaknesses through to the output.

Four Questions Before Recommending AI

Is it a process problem? If people can’t describe the current workflow the same way, if there’s duplicate effort, unclear ownership, or reliance on individual workarounds, the fix is to simplify and standardize the process not automate a mess or hand it to a model.

Is it an automation problem? Where rules are predictable and the work is repetitive moving data between systems, routing tasks, generating routine reports conventional workflow automation will usually be cheaper and more reliable than AI.

Is it a data problem? When information can’t be trusted, connected, or consistently defined, the fix is data governance: master data standards, system integration, clear ownership. This matters especially in accounting, where auditability and control outweigh novelty. CPA Ontario and RPA Canada’s guidance on AI governance centers on trust, professional judgment, and accountability precisely what erodes when weak data hides behind a polished AI interface. AI depends on good data; it is not a substitute for it.

Is it actually an AI problem? This is the case when work requires interpretation, prediction, pattern recognition, or generation and the business can tolerate a probabilistic tool with human oversight. Good candidates include document classification, forecasting, anomaly detection, summarization, and decision support, tied to a measurable pain point and a workflow where people can check the output.

Where AI Actually Belongs in Accounting

None of this means AI has no place in accounting. Once process, automation, data, and genuine intelligence problems are properly separated, AI can meaningfully help summarize reconciliations, flag unusual transactions, explain variance drivers, draft disclosure language, or surface policy guidance. But those applications only work sitting on reliable data pipelines, documented controls, and clear ownership. Without that foundation, AI becomes a confidence machine attached to an unreliable source of truth.

For practitioners, the professional obligation isn’t to sell AI into every conversation it’s to protect clients from solving the wrong problem elegantly. Clients will keep asking for AI because it’s visible and exciting, or because a competitor has it. The better service is often telling them, clearly, where AI doesn’t belong yet.

The Real First Step

Most clients don’t need AI first. They need a sharper understanding of their own workflow, cleaner data, and automation applied where the rules are already clear. LLMs create real value in software and technical fields; in accounting and finance, the first question is rarely model capability. It’s whether the data, controls, and process are sound enough for any intelligent system to be trusted at all. Ask what problem you’re solving before asking how AI might solve it, and AI stops being a buzzword and becomes what it should be: a tool used only where it belongs.