TaskVector
Status: field-tested at scale across PE portfolio companies. Formal validation pending.
Before automating a task, score its risk across 9 dimensions. Each one is anchored: pick the description that matches your task and the number comes with it. Low is good - 0.0 is the best possible score and 1.0 the worst. If any single dimension lands above 0.8 it is a landmine, and the composite stops mattering.
Drag across a row to score that dimension (far left = 0.0, no risk; far right = 1.0, landmine). Scores snap to the rubric’s five anchors.
Does the AI know what it is going to see?
Can you enumerate the input formats? · red flag: "Whatever the customer sends"
How fast can you tell if the AI got it wrong?
How long until you discover an error? · red flag: "We’ll know in a few months"
How many steps? How deep is the reasoning?
If each step is 90% accurate, what is the accuracy after all of them? · red flag: "It’s complicated"
Is verifying the output easier than producing it?
What is the ratio of (time to do) / (time to check)? · red flag: "You have to really read it"
Can you split the task into independent pieces?
If item #5,000 fails, do you lose items #1-4,999? · red flag: "It all connects"
Does the AI have everything it needs to succeed?
Will the AI have to make something up to finish the task? · red flag: "You need to understand the context"
What happens when the AI is wrong?
At 99% accuracy and 1M decisions a year, what do 10,000 errors cost? · red flag: "If this goes wrong..."
Does the task need more than text?
Does success require understanding non-text information? · red flag: "There are images and videos"
Can current AI models actually do this?
Has this exact task been done reliably before? · red flag: "I’ve seen demos that do this"
The Landmine Rule
Load the medical diagnosis preset. It averages 0.58 - middling, the kind of number that survives a steering committee. Failure Consequences is 0.95. Averaging is what hides that, so the rule runs before the average and overrides it.
The landmine rule exists because automation failures are not normally distributed. They are fat-tailed. The expected cost of a worst-case error on a landmine dimension dominates every other term in the sum, and a mean is exactly the wrong statistic for a distribution whose tail carries the mass.
Why Low Is Good
TaskVector is a risk vector, not a quality score. Each dimension measures how much of a specific failure mode the task carries, so 0.0 means the failure mode is absent and 1.0 means it is maximal. Scoring it the other way round would make the composite read as “how good is this task,” which is a different question with a different answer.
One consequence worth stating: this instrument answers can it be done reliably, not is it worth doing. Volume, savings and payback are value terms and belong in the NPV model, not in here. A rare task can be perfectly tractable, and a high-volume one can be a landmine.
What Comes Next
TaskVector tells you whether to automate. The rest of the toolkit tells you how:
Verification Quadrant - plot the task by calculation vs. verification difficulty. Calculate the Templeton Ratio.
Dollarized Confusion Matrix - price the error costs. Compute the optimal threshold.
The Promotion Protocol - deploy in HITL, gather evidence, promote to autonomous on statistical proof.
When to Use This
Use when
- +Evaluating one specific task for AI deployment, not an entire function
- +You have real examples to score against, not vibes
- +Multiple stakeholders need a shared scoring frame to align
- +You want to surface landmine dimensions before deploying, not after
Skip when
- -The task is obviously too immature (no data, no users, no rubric)
- -You already have deployment experience with this exact task class
- -Budget is the primary binding constraint - go straight to Automation NPV
- -The scoring itself would take longer than running a small pilot
Rosetta Stone
Four circles, four readings of the same object. Each role reads the artifact through its own lens.
Multi-factor scoring for operating instruments. Analogous to factor models in equities: decompose the task into orthogonal-ish dimensions, score each, rank by composite.
A screener for the next hundred tasks on the backlog. Run the vector, sort, work top-down. Keeps the team from optimizing for the loudest project instead of the most automatable.
A checklist with teeth. Nine questions that usually catch the "sounds automatable, isn't" failure mode before it eats a quarter.
A feature vector in R^9 embedding the task. Similarity in this space predicts automation viability. With enough labeled history, a supervised classifier outperforms hand-scoring.
See also
Which bets to make. Capital allocation, M&A due diligence, portfolio construction.
How to execute at scale. Multi-brand portfolio, turnarounds, P&L ownership.
Builds it, ships it, owns it. Solo full-stack, DevOps, production systems.
Proves it, models it, publishes it. Mathematical modeling, Bayesian frameworks.
TaskVector