Which Tasks Should You Hand to AI? A Practical Scoring Framework

The way many companies prioritise AI automation is a single meeting: processes go up on a whiteboard, and people vote by gut feel. The outcomes tend toward two extremes — either the hardest target in the room ("automate all of customer service") or the least painful one. The meeting ends with a list, but still no standard for judging it.

Here is a four-dimension scoring framework. It needs no technical background — a line manager can classify their own processes with it, and the score pattern points straight to a disposition.

Four dimensions, each an answerable question

Is the input already digitised? AI can only process what it can read. The test is direct: does this task's input already exist as files, tables or system records? Customer emails, tickets and e-invoices count; paper slips, verbal instructions and chat logs scattered across a salesperson's personal WeChat do not. If a person must transcribe the input first, transcription eats back the time automation saves.

Can the rules be written down? A useful test: ask your most experienced person to write the handling rules out, to the level where a new hire following them gets eight out of ten right. If it can be written, AI can probably execute it. If the writing keeps hitting "well, it depends" and the list of cases never closes, the task runs on accumulated feel — not something to hand to AI on its own yet.

Is the output easy to verify? Someone must be able to check AI output quickly. Whether invoice fields were copied correctly is verified at a glance against the original; whether a market analysis is sound may take longer to verify than writing one yourself. Where verification is expensive, review becomes the new bottleneck after automation.

Are errors reversible? A wrong number in an internal weekly report is fixed by a corrected resend. A quote sent to the wrong client, or a liability clause missed in contract review, cannot be recalled. Reversibility sets how much autonomy you can afford to grant.

Four score patterns, four dispositions

Rate each dimension high or low, and the combinations settle into four groups:

• High on all four: automate outright, with human spot checks. Clean input, clear rules, checkable and correctable results — hand the whole segment to AI and sample-check a share of it.
• Rules are clear, but verification is hard or errors are costly: AI drafts, a person confirms. AI produces the draft or suggestion; a person keeps the final say. This is the most common and most stable arrangement in companies — for how to tier the review effort, see Reviewing AI Output: A Tiered Checklist.
• Input not digitised, or rules cannot be written: fix the process before discussing automation. Moving input into systems and turning veterans' judgment into written rules is the groundwork — worth mapping the whole flow first, as covered in What to Map Before You Automate.
• Rules unclear and errors irreversible: leave it alone for now. Keep people on it, and reassess once rules accumulate or tools improve.

Run seven everyday tasks through the framework

In the automate-outright group, the textbook case is invoice data extraction: the input is electronic, the rule is fixed-field extraction, right or wrong shows at a glance, and a wrong record is simply corrected. Weekly report consolidation behaves the same way — the data already sits in systems, the aggregation rules write down cleanly, and a mistake gets fixed the following week. After-sales ticket classification usually lands here too: the ticket text is ready-made, the categories are enumerable, and a misclassified ticket costs one extra handover.

In the draft-and-confirm group, the classic case is customer-service FAQ replies: answers to high-frequency questions are documented and the rules are clear, but a wrong reply to a customer is a public-facing incident — so the sound pattern is AI suggests, the agent confirms and sends, or full automation only within an explicitly fenced set of questions. Résumé screening belongs here too, with one caveat: hard requirements can be written down, but "a good candidate silently filtered out" is an error you will never discover. Let AI rank and annotate; do not let it reject.

In the fix-the-process-first group, the recurring example is quote approval: in many companies, half the pricing rules live in policy and the other half in the boss's head ("key accounts can get two more points — depends who introduced them"). Until the rules are written down, there is nothing for AI to execute. Making the rules explicit is valuable to the business in its own right, AI or not.

In the leave-it-alone group: the final call on contract clause review. Clause risk depends on context and commercial background; verification is hard and errors are irreversible. What AI can do is flag suspicious clauses and check the draft against a checklist for omissions, with the decision left entirely to a person — which in practice folds back into draft-and-confirm, with the human share turned up.

How this relates to choosing your first scenario

If your company has not started with AI at all, you do not need this framework yet — go straight to which scenario your first AI workflow should target, which answers "where to begin": one high-frequency, low-risk, quick-payoff spot. This framework answers the question that comes after: once candidates multiply and resources are finite, how to rank the whole board. Sweep every candidate through the four dimensions: order the automate-outright and draft-and-confirm groups by payoff, move fix-the-process items into the process improvement plan, and give every leave-it-alone item a review date.

Scores change; the framework does not

One closing reminder: scores are dynamic. Roll out a new system, and a task once stuck at "input not digitised" moves up a tier; when a veteran turns experience into a written manual, "rules unclear" becomes "rules written". The framework is worth re-running every six months. Its real value is not computing a precise score — it is moving the prioritisation discussion from "I think this one could work" to "this task is blocked on exactly this dimension". The second conversation is the one you can act on.