Task Decomposition

Step by step, with review time

Break the process into steps and estimate, for each, how much of it AI can handle and how much review time remains. Some steps (downloading files) automate almost fully; others (keying fields) mostly, with review; some (vendor phone calls, exceptions) not at all. The total time per item after automation is usually a meaningful reduction, but far from zero. This is the single most important correction to optimistic estimates.

Per-step time before and after, run

I ran this with plain Python 3. All figures belong to one worked example (an invoice-processing team) with invented but internally consistent numbers; prices are placeholders. Four steps taking 6.5 minutes become 2.69 minutes with AI and review: downloading drops to 0.05 minutes, keying to 1.18, matching to 0.66, and exception handling stays at 0.80. That is a 59% reduction, not 100%.

steps = [  # (step, minutes per item today, share automatable, review minutes if automated)
    ("open email, download PDF", 1.0, 0.95, 0.0),
    ("read and key 12 fields", 3.5, 0.80, 0.6),
    ("match to purchase order", 1.2, 0.60, 0.3),
    ("handle exceptions / call vendor", 0.8, 0.00, 0.0),
]
before = sum(m for _, m, _, _ in steps)
after = 0.0
for name, m, auto, review in steps:
    remaining = m * (1 - auto) + review * auto
    after += remaining
    print(f"{name:<32} {m:>4.1f} min -> {remaining:.2f} min")
print(f"total per item: {before:.1f} -> {after:.2f} minutes ({1 - after / before:.0%} reduction, not 100%)")

Output:

open email, download PDF          1.0 min -> 0.05 min
read and key 12 fields            3.5 min -> 1.18 min
match to purchase order           1.2 min -> 0.66 min
handle exceptions / call vendor   0.8 min -> 0.80 min
total per item: 6.5 -> 2.69 minutes (59% reduction, not 100%)

Validate shares with a pilot

Test the automatable share on real items before trusting it; early guesses are usually too high.

त्वरित जाँच: Why decompose a process into steps?

• To make the process longer
• Different steps automate to different degrees, and review time remains
• Because AI can only do one step
• To avoid measuring anything