Automate Work Tasks: Draft First, Approve Second

“Automate work tasks” is usually typed by someone staring at the same five-minute chore for the tenth time this week, not someone wanting a platform comparison. The useful question is not which tool to install. It’s which part of the task is actually safe to hand to a model, and which part still needs a person to look at the result before anything happens because of it. That split — drafting versus acting — is the method below, and it works whether the tool is a chat window, a spreadsheet formula or a connector platform.

The split that actually matters

A model can read something unstructured — a pile of receipts, a messy inbox, a spreadsheet with gaps — and turn it into a structured draft: categorised rows, a summarised list, a filled-in template. That is drafting. Acting is anything that leaves the draft stage: submitting an expense report, sending an email, updating a record another system depends on, approving a payment. The first half is low-risk and gets more reliable every model generation. The second half is exactly as risky as it always was, because the model does not know what it does not know, and a wrong action is harder to undo than a wrong sentence.

Large-scale research on how people actually use AI tools at work keeps finding the same shape: most real usage is drafting, summarising and classifying, not a model autonomously finishing a transaction end to end. That is not a limitation to engineer around. It is roughly the right boundary to build a workflow on top of.

A four-step method

1. 01Name the task in one sentence, specifically enough to hand to someone new without a meeting — not “handle expenses”, but “turn ten photographed receipts into expense-report rows, flagged by category”.
2. 02Decide which half is drafting. Usually everything up to “produces something a person reads” is safe to automate first; anything past “now something happens because of this” is not, yet.
3. 03Give the model a fixed, closed set of outputs — categories, fields, a template — instead of open text, plus an explicit “flag for review” option so it has somewhere to put a case it is not sure about. Anthropic’s documentation on tool use covers why constraining the output shape changes reliability more than rewording the instruction does.
4. 04Add one human check between the draft and the action, placed where the draft is actually wrong often enough to matter, not added as a formality nobody reads.

A worked example: expense receipts

Take a real version of this. Ten receipts come in after a work trip — a restaurant, two taxis, a hotel folio, a handful of shop receipts, one in a foreign currency. The automation reads each photo, pulls the date, amount and vendor, assigns a category from a fixed list (travel, meals, lodging, other), and writes it to a draft row in a spreadsheet. Nobody’s expense report has shipped yet. A person still opens the sheet, checks the ten rows against the ten receipts, and only then submits it.

Run four kinds of receipt through it before trusting the category it picks: a clear one (easy, and it tells you nothing, because every tool gets this right); a blurry or partly torn one; a duplicate of a receipt already logged last month; and the foreign-currency one, where the model has to decide whether to convert the amount or flag it. The first is the one every demo uses. The other three are where the workflow earns its keep or quietly fails.

Where it breaks

• —A model given an open text field instead of a fixed category list will invent one that does not exist in your chart of accounts, and nothing downstream catches it until a bookkeeper does.
• —A duplicate receipt gets logged twice because the automation has no memory of what it already processed, not because anyone told it to double-count.
• —A blurry total gets read with confidence and no visible uncertainty — fluent and correct are not the same property, and a wrong number in a clean row looks exactly like a right one.
• —A failed step in a connector chain can stop silently: no row appears, no error either, and the receipt simply never gets logged.

A rule-based step either runs or throws an error. A model-based step can run, produce an answer, and still be wrong — and the workflow has no way to tell the difference on its own.

Checks before you trust it unattended

1. 01Keep the category list closed, with “uncertain, check by hand” as one of the options, rather than leaving a bad extraction to invent one.
2. 02Log every run, including the ones that produced nothing, somewhere a person actually opens. NIST’s AI Risk Management Framework recommends exactly this kind of ongoing monitoring, not a one-off accuracy test, for any AI-assisted process.
3. 03Sample-check the real output against the real receipts for the first few weeks, not just the four test cases you ran once.
4. 04Keep the human step before anything the workflow cannot cleanly undo — a submitted report, a payment, a record another system now depends on. Anthropic’s research on where agentic systems still need a person in the loop makes the same case at a larger scale: the newer the automation, the more it benefits from a checked handoff, not fewer checks.
5. 05Revisit the split itself after a month. Some of what you marked “acting” might turn out to be low-stakes enough to automate too — that should be a decision made on purpose, not a boundary that quietly erodes.

What to do Monday

Pick the one task in your week that is mechanical except for a judgement call buried in unstructured input — receipts, an inbox, a form — and automate only the drafting half first. Office automation is the broader version of this for meeting notes and routine replies, and customer support automation applies the same split to a support inbox, where the stakes of acting (replying to a customer) are higher than the stakes of drafting. If the task involves a spreadsheet specifically, Excel accounting software covers when a dedicated tool earns its keep over doing this by hand, and Excel errors is worth reading before you trust any formula the draft step writes into a cell. Cloud automation tools runs the same fit test one layer down, for a copilot that can change infrastructure directly rather than fill in a spreadsheet row. Whichever half you automate, write the instruction itself with care first — five AI prompts worth reusing is a reasonable starting set, and if the draft needs to read like something you would actually send rather than a form letter, shaping AI text so it sounds like you is the narrower fix for that one problem.

The judgement behind this split — knowing which part of a task is safe to hand to a model and which part still needs a person — is close to the skill behind the wage premium PwC has been tracking for workers with measurable AI skills: 62%, up from 57% the year before. Coursium teaches that judgement directly, one short lesson at a time, through ChatGPT Essentials and AI Excel. Stay ahead of AI by learning the tools on your phone.