Most AI projects stall at the selection step, not the prompting step. You pick a task because it sounds automatable, spend two weeks wiring it together, then review the output so carefully that you save nothing. Deciding which AI tasks to automate is a measurement problem before it is a technology problem, and you can get most of the way there with a blank spreadsheet, a timer, and two weeks of honest notes.
The method has four steps: inventory the tasks, score seven columns, apply one decision rule, re-measure after 30 days.
Start With Tasks You Can Name in One Sentence
Open a spreadsheet. One row per task. A task belongs on the list only if you can state three things: what triggers it, what a finished version looks like, and how often it happens.
That rule filters out most of what people try to automate. „Marketing“ is not a task. „Customer support“ is not a task. These are:
- Write a first-draft reply to a shipping-delay email.
- Turn a 40-minute discovery call into six bullets and three action items.
- Categorize last month’s bank transactions into your chart of accounts.
- Rewrite one approved blog post into five short social posts.
Aim for 15 to 25 rows and score nothing yet. Add rows for a week or two as work comes in, because the tasks you forget are usually the frequent, invisible ones.
The Seven Columns That Decide Which AI Tasks to Automate
Now score each row.
- Volume. Times per month. Count it; don’t estimate from feeling.
- Human minutes per run. Time yourself three times and take the median. Memory inflates boring tasks and understates familiar ones.
- Cost per run. Loaded hourly cost divided by 60, times the minutes. Price your own hour at what you would charge a client, not at zero.
- Quality bar. One of three: internal-only, customer-visible, or decision-bearing. Write down who on the team can tell good output from bad.
- Review minutes per run. Time to check the output and fix it. This is the column people skip, and it decides the outcome.
- Data sensitivity. What actually goes into the prompt: public, internal, confidential, or regulated. Client financials, health details, Social Security numbers and card data are their own category and change which tools you may use at all.
- Exception rate. Share of runs that don’t fit the standard pattern. Estimate from the last 20 runs.
Add one calculated column: net minutes saved = human minutes minus AI runtime minutes minus review minutes. Multiply by volume for monthly hours saved. A task with 40 minutes of human work and 35 minutes of review is not a candidate, however good the demo looked.
A Worked Example With Three Candidates
Say you run a four-person e-commerce and consulting shop, and three tasks reach the top of the list.
Task A, shipping-delay replies. 120 per month, 6 minutes each, so 12 hours of human time. The draft takes under a minute; review and light editing take 2. Net saving is about 3 minutes per run, roughly 6 hours a month. Data is internal plus a customer name, quality bar is customer-visible, exceptions run near one in ten and are mostly international orders.
Task B, monthly transaction categorization. Once a month, 180 minutes. AI can suggest categories, but because the output feeds your books you would still read every line, so review lands near 150 minutes. Net saving is half an hour a month on decision-bearing, confidential data.
Task C, call recording to summary. 20 per month, 25 minutes each, review 6 minutes. Net saving is about 17 minutes per run, close to 5.5 hours a month. Data is client-confidential, so the tool needs written terms on retention and model training before anything starts.
By raw hours saved, C nearly ties A. By hours saved per hour of setup, and by how much damage a bad output can do, A goes first: cheap review, small blast radius, and the reviewer is already in the inbox. C goes second, once the vendor question is settled. B never gets automated. The real fix there is upstream, in bank rules and a cleaner chart of accounts.
If you would rather not build the sheet from scratch, the AI Workflow Inventory & ROI Tracker is organized around this same inventory, score and decide sequence.
The Decision Rule
Automate a task when review time is under half the human time, the exception rate is under roughly 20 percent, the reviewer can genuinely judge the output, and one bad result slipping through costs less than a month of the savings.
Then four hard stops. If one applies, the spreadsheet result does not matter.
- The output is advice. Legal, tax, medical or financial guidance going to a third party needs a named accountable human as its author. AI can produce a draft for that person; it cannot be the last step.
- Sensitive data without written vendor terms. No documented position on retention and training means the task stays manual until you have one.
- Nobody can review it. If no one can separate a correct answer from a confident wrong one, you cannot review, so you cannot automate.
- The task exists because something upstream is broken. Automating it just makes bad inputs travel faster.
Then pick the top two survivors. Not eight. Two live workflows with real review discipline beat eight half-built ones.
Measure Again After 30 Days
Before switching anything on, record the baseline: minutes per run and current error rate. Then log every run for four weeks with five fields: date, AI minutes, review minutes, edited or rejected, exception yes or no.
At 30 days, compare actual net savings with your estimate. The usual surprise is review drift, as people find failure modes the demo never showed. Set the kill rule in advance: if actual savings land under half the estimate, narrow the task or switch it off. Count quality escapes separately, meaning wrong outputs that reached a customer. One escape on customer-visible work justifies a scope change even when the hours look good. The answer to which AI tasks to automate also shifts as tools change, so rerun the scoring every quarter or two.
Where the Spreadsheet Stops Being Enough
Be honest about the model’s limits. Columns cannot price these:
- Trust damage. One wrong customer-facing output can cost more than a year of saved minutes, and you cannot know in advance which one it will be.
- Skill erosion. Junior staff learn judgment by producing first drafts. Automate every draft and you stop training the reviewer you will need.
- Induced demand. When summaries become cheap, you produce far more of them, and the review load grows with the volume rather than staying flat.
- Chained steps. When one AI step feeds another, small error rates compound. Chains need process mapping and test cases, not extra columns.
Past a handful of live workflows, the spreadsheet becomes an index rather than the system. What you need then is written procedure: a named reviewer per workflow, a defined escalation path, a disclosure rule for customer-facing output, and a record of what was reviewed and when. Other operational walkthroughs are on the Cursiqa blog.
If you want the scoring sheet already built, the AI Workflow Inventory & ROI Tracker covers the inventory and the numbers, and the Responsible AI Use & Human Review SOP Kit adds a 16-page guide, a 9-sheet workbook and 16 policy and SOP templates for the stage where review has to be written down. Both are in the Cursiqa shop when you reach that point.