Every conversation about automation eventually arrives at the same question: what should we automate first? It sounds like the practical question. It is actually the second question, asked too early. The first is: by what criterion would we rank the candidates? Skip it, and "first" gets decided by whatever a vendor demonstrated most recently — which is how companies end up with a chatbot on a website that receives eleven visits a day while three people retype supplier invoices in the back office.
The ranking criterion is not mysterious. A process is a strong automation candidate when five things are simultaneously true, and the discipline lies in checking all five before falling in love with any one of them.
The five tests
It repeats.
Frequency is the multiplier on everything else. A task performed forty times a day repays automation forty times faster than the identical task performed weekly. This is why the glamorous, occasional work — the quarterly strategy deck, the annual report — ranks low even when it consumes painful hours, and why the invisible daily work ranks high.
It follows rules more than judgement.
The current generation of tools has widened what counts as "rule-following" — extracting fields from an unstructured document is now routine where five years ago it was research. But the test still holds: if two competent employees would process the same input the same way, it is automatable. If they would argue about it, it is a judgement call, and judgement calls belong with people — supported by better information, not replaced.
The data it needs is reachable.
An automation is only as good as its access to the systems of record. If the process lives in the ERP, the accounting platform, and a shared inbox, integration is engineering work with a known shape. If it lives in one veteran employee's memory, the first project is documentation, not automation — and discovering that early is cheaper than discovering it in week six of a build.
Its cost of error is bounded.
Start where a mistake is detectable and reversible: an internal report, a draft the human approves, a data entry that reconciliation would catch. Processes where an error reaches a customer or a regulator unreviewed come later, after the system has earned trust on lower-stakes work. Sequencing by risk is not timidity; it is how confidence is built at a price the business can afford.
Its result can be counted.
Hours per week, error rate, cycle time in days. If nobody can state what the process costs today, nobody will be able to prove the automation changed anything — and the research is unambiguous about where that leads. MIT's 2025 study of enterprise AI deployments found roughly 95% of pilots produced no measurable P&L impact, with unscoped, unmeasured initiatives dominating the failures. Measure it first; that work is never wasted.
Where the tests usually point
Run honestly, the five tests tend to converge on the same unfashionable territory: document-heavy, back-office, cross-system work. Invoice and order intake. Data transfer between systems that do not talk to each other. Report assembly. Confirmation and status communication. The MIT research found exactly this mismatch at scale — over half of AI budgets flowing to sales and marketing pilots while the measurable returns sat in back-office operations that received a fraction of the attention.
For companies operating in Romania, one category currently outranks everything else on all five tests at once: fiscal reporting workflows. The obligations are no longer approaching — they are here. e-Factura has been mandatory for domestic B2B invoicing since July 2024, with enforcement extended to the smallest taxpayers from July 2026 and penalties for non-compliant B2B invoices of up to 15% of the invoice value. SAF-T (D406) reporting now covers small taxpayers as well, transmitting standardised accounting data to ANAF on a fixed rhythm. Every one of these flows is repetitive, rule-based, connected to systems of record, catastrophic to get wrong manually at volume, and precisely measurable. Some companies absorbed this early by wiring their ERPs directly to ANAF's platforms; any firm still handling these flows semi-manually — whatever its size — is running the highest-frequency, highest-penalty manual process in the building. It should not lose a ranking exercise to a chatbot.
The exercise itself
The method fits on one page, deliberately.
List every process the team touches in a normal week — not the org chart's version, the actual one, including the workarounds. For each, estimate frequency, hours consumed, error consequences, and which systems it touches. Score the five tests. Rank by expected return: hours recoverable per month, multiplied by honesty about integration difficulty.
Then — and this is the step that separates an automation programme from a purchasing spree — take the top candidate and write down the success criterion before selecting any tool: "Supplier invoice processing currently averages 18 minutes per invoice with a 4% correction rate; the automated process must average under 3 minutes with corrections under 1%, measured over 60 days." A sentence of that shape does three things. It forces the current state to be measured, which most companies discover they cannot yet do. It makes vendor comparison rational, because every demo is now auditioned against the same number. And it makes the result checkable by a sceptic, which is the only audience whose approval compounds.
The tool comes last. This ordering feels slow to a team eager to start, and it is slower — for about three weeks. Then it is dramatically faster, because the build has a specification instead of a vibe, and the go-live has a finish line instead of a fade-out.
What should we automate first? Whatever survives the five tests with the highest count of hours attached — which, in most companies right now, is something nobody would put in a demo. That is not a disappointment. It is where the money has been sitting the whole time.
Sources: MIT NANDA, "The GenAI Divide: State of AI in Business 2025" (2025); ANAF / Romanian Ministry of Finance e-Factura and SAF-T implementation timeline as consolidated in VAT-compliance industry reporting (2024–2026). Figures on Romanian fiscal obligations reflect rules in force at the time of writing.
