After five years of automating business processes across industries, I have a mental tier list of what works and what does not. The tier list is not what most people expect. The functions that sound the most exciting to automate are often the worst candidates. The boring ones are gold.
Tier one — automates beautifully: accounts payable, accounts receivable, bank reconciliation, invoice processing, payroll calculation, inventory reordering, and compliance reporting. These share a common trait: they are rules-based, data-heavy, repetitive, and the cost of errors is high but the errors themselves are predictable. We deployed AI accounting across five countries for a corporate services group. The system handles SFRS, HKFRS, and AASB standards simultaneously. It works because accounting has clear rules, even when those rules differ across jurisdictions.
Tier two — partially automates with human oversight: customer service triage, contract review, financial forecasting, and quality inspection. These are tasks where AI can do 80% of the work but a human needs to review the output. For a healthcare group running 40+ clinics, we built SARIMA-based financial forecasting. The model is good. But the clinic managers still review the forecasts because they know things the model does not — a new competitor opening nearby, a doctor leaving the practice, a seasonal pattern specific to their location.
Tier three — automate with extreme caution: strategic planning, client relationship management, creative work, and complex negotiations. These involve too much ambiguity, context, and human judgment. I have seen companies try to automate sales with AI and end up with a system that sends tone-deaf emails to their most important clients. Some things need a human.
The biggest trap is automating tier-three functions because they sound impressive. A board presentation about AI-powered strategic planning gets more applause than one about AI-powered invoice processing. But invoice processing actually works, and it saves real money. We automated the entire accounts payable workflow for a furniture retailer with hundreds of stores. It is not glamorous. It saves them thousands of hours per year.
Within tier one, the order of implementation matters. Start with the function that has the most structured data and the clearest rules. For most companies, that is bank reconciliation or invoice processing. These functions also tend to have the best historical data for training and validation. And the people doing them are usually eager for relief — they know the work is tedious.
The transition from tier one to tier two is where most companies stall. Tier one gives you clean wins. Tier two requires you to design human-in-the-loop workflows, and that is a fundamentally different kind of system. You are no longer replacing a process; you are augmenting a person. The UX matters more. The confidence scoring matters more. The exception handling matters more.
One pattern I have noticed across every engagement: the functions that automate worst are the ones where the rules are not written down. If the process lives in someone's head — if the answer to how do you do this is it depends and then it is complicated — that function is not ready for automation. It might never be. Before you automate, you need to codify. And sometimes the act of codifying reveals that the process is more arbitrary than anyone realized.
The financial case is straightforward. Tier-one automation typically pays for itself within six to twelve months. Tier-two projects take eighteen to twenty-four months. Tier-three projects rarely pay for themselves at all, though they can have strategic value if scoped correctly. If your CFO is asking about AI ROI, point them at tier one and let the numbers speak.
My recommendation to any company starting an automation journey: make a list of every business function, honestly assess which tier each one falls into, and work through tier one before you even think about the rest. It is not exciting. It is effective. And that is what matters.