The Automation ROI Audit: Finding the Six Processes Worth Automating This Year

Ask any operations team what to automate and you get either everything ("we're drowning") or the pet grievance (the report someone hates compiling). Both answers waste the first, most credibility-rich automation budget. The audit that works is unglamorous: enumerate, score, rank, and let the arithmetic argue. We run it in two weeks; here is the method.

Enumerate honestly

Workshop per team, one question: what do you do repeatedly? Capture frequency, time per instance, who does it, systems touched, and what happens when it goes wrong. The list always surprises: the Monday copy-paste rituals, the approval chains where the approver rubber-stamps 97%, the return-reason coding nobody reads, the supplier emails re-keyed into the ERP. Fifty to ninety processes is typical for a mid-size commerce operation. Most will not survive scoring — that is the point.

Score on four axes

• Volume × time — the gross prize, in hours. High-frequency-short beats low-frequency-long at equal totals: less variance, faster learning loops.
• Specifiability — can a competent colleague execute from written rules alone? Fully specifiable → deterministic automation; mostly-specifiable-with-judgement → agent with approval gates; tacit-expertise → leave it, for now.
• Error cost, both directions — what a mistake costs (caps autonomy, adds review) and what current human error already costs (the case for change; audits regularly find the manual process erring more than anyone admitted).
• System accessibility — API-reachable systems score up; portal-only and PDF-shaped inputs score down but are increasingly model-tractable; the licence that forbids automation scores zero and goes on the renegotiation list.

Multiply, rank, and take the top six — not sixteen; program credibility is built by finishing.

The portfolio shape that pays

The ranked list almost always yields: two deterministic integrations wearing automation costumes (build the boring pipeline, no AI required — fastest payback in the set), two agentic workflows (document intake, reconciliation, triage — the judgement-shaped middle), one assist-not-replace (drafting for human send, the highest-acceptance pattern in customer-facing work), and one decline — the pet grievance, scored honestly at last, returned with numbers and condolences.

Instrument before victory laps

Baseline before build: hours, error rate, cycle time, unit cost. Report against baseline monthly; feed exceptions back as eval cases; expand autonomy only when the log earns it. Automation programs die of unverifiable claims and unowned failures — the audit's real deliverable is a measurement habit, which outlives any single bot.

The two-week version — enumeration workshops, scored register, top-six briefs with honest integration estimates — is a fixed-fee engagement precisely because the method is the same every time. The ranked list is different every time; that is why the audit exists.