Task Framework · Task Economics

Which AI Assistant Fits The Job?

Start with the work that needs doing. Define a usable result, compare assistants under the same conditions and account for human effort.

01 · Start

Better At What?

A useful comparison starts with a familiar job. “Good at writing” is too broad: suggesting a short caption and compiling an annual report place different demands on an assistant.

What Is The Job?Choose the task family and describe the intended result.
What Must It Handle?Specify the information, constraints and finish line.
How Well Did It Do?Check quality, consistency, effort, speed and boundaries.

02 · Families

Six Kinds Of Work

These are task families, not levels of intelligence. A difficult explanation may demand more capability than a simple booking. Classify the main outcome; add other families when the job contains several stages.

Understand

Explain or make sense of something: a concept, document or procedure.

Usable means: accurate, clear and appropriate to the reader. A fluent explanation still needs its facts checked.

Find

Locate information or evidence: a file, a specification, a current price or an available appointment.

Usable means: the relevant information is supported by a verifiable source. Time-sensitive facts need a current source.

Transform

Change existing material while preserving what matters: summarise, translate, rewrite or extract fields.

Usable means: the required meaning and facts survive, and the output fits the requested format.

Create

Produce something new against a brief: an email, campaign ideas, a spreadsheet formula or a working calculator.

Usable means: it meets the brief and works in its intended setting. Attractive output can still be wrong.

Evaluate And Plan

Compare possibilities and recommend a course: suppliers, flooring options, an itinerary or a sequence of work.

Usable means: constraints are respected, assumptions are visible and the trade-offs support the decision.

Act

Make a change in a system or service: save a file, update a record, schedule an appointment or place an authorised booking.

Usable means: the intended change actually happened, within the permissions granted, and the result is confirmed.

For example, planning and booking a trip involves finding, evaluating and acting. Keep the whole trip as the user’s job, but inspect each stage when something fails.

03 · Conditions

Same Family. Different Demands.

Information

Does the work use supplied material, general knowledge, live sources or private connected sources? Record the source and the access the assistant actually has.

Completion Test

Does the result require verification of exact fields or a calculation? Does it require judgement against a creative brief? Is exploration itself the intended outcome?

Complexity

Is this one bounded output or a chain of dependent steps? Note the length of the material, number of files and constraints that must remain consistent.

Consequences

Is a mistake easy to inspect and reverse, or costly to get wrong? Required review and permission are part of the task, not optional extras.

Record the language and domain when they matter. Success with an English invoice does not establish reliability with a Greek one.

A missing connection is a limit of the assistant setup. A wrong answer despite access is a performance failure. Record the distinction.

04 · Finish

Define Usable Before Testing

Write the conditions that make the result usable. Decide them before seeing the assistants’ answers, so a polished response does not quietly lower the standard.

Required content: what must be included, and what must be excluded?
Accuracy: which facts, numbers or source details must match?
Format: what structure or file must work?
Action: is advice enough, or must something actually be changed and confirmed?

Exploration also needs a boundary. “Generate ten distinct campaign directions with one reason for each” is a task with a finish line. Extra ideas belong to the original task when the brief requests them.

This builds on the Effective Task Cost Framework: the assistant finishing a reply and the job being usable are separate events.

05 · Example

One Meeting. Three Assigned Actions.

This fictional meeting recurs throughout the framework, guide and checks. The job is to turn the notes into a usable action table.

Mary: Send the revised budget by Thursday.
Tom: Confirm supplier availability by Friday.
Gloria: Update the launch checklist by Monday.
Peter: A customer FAQ could be useful later.
Meeting note: No owner or deadline was agreed for the FAQ.

Task Family: Transform

The task reorganises existing meeting notes into a table while preserving the assigned actions, owners and deadlines.

Working Conditions

All required information is contained in the supplied notes. The assistant produces one table. No external research or changes to other systems are needed.

Finish Line: include all three assigned actions with the correct owners and deadlines. Preserve the day names. Exclude the unassigned FAQ idea. Return an Action / Owner / Deadline table without extra commentary.

The Usable Result For This Example
Action	Owner	Deadline
Send the revised budget	Mary	Thursday
Confirm supplier availability	Tom	Friday
Update the launch checklist	Gloria	Monday

Equivalent wording is fine if it preserves the action. Inventing a calendar date, adding a fourth assigned action or omitting an owner fails the check.

A later request to put these actions into a calendar adds an Act stage. A request to verify supplier availability adds a Find stage using current information.

06 · Compare

Measure More Than The First Reply

Keep Each Measure Tied To The Same Job And Conditions
Measure	What To Check
Capability	Can it meet the completion criteria on representative examples, with the tools available?
Quality	How well does the result meet the brief? In the meeting example, check completeness and fidelity to the notes.
Reliability	How consistently does it pass across fresh attempts and different examples? Record successes and attempts, not just the best answer.
Efficiency	How much active human time goes into prompting, reading, checking, correcting and finishing?
Speed	How much elapsed time passes from starting the job to a usable result, including waiting?
Scope Discipline	Does it finish the requested job and keep optional suggestions separate? Useful extra work still requires a deliberate decision to pursue it.

One successful answer shows a success. It does not establish reliability. Repeat in fresh chats and vary the source material: longer notes, missing deadlines, conflicting statements and other examples drawn from routine work.

Use the same starting brief, source and correction budget for each assistant. Record the product, model or mode when visible, plan, available tools and date. When comparing complete assistants, tool access is part of their usefulness.

Report failures and time spent on them. A quick result that fails the completion test is not a cheap completed job. Small personal trials are useful evidence for the workflow tested, not a universal ranking.

07 · Check

What Makes The Difference?

The table looks right, but Thursday has become an invented calendar date. Is the job usable?

No. The brief requires deadlines as written. A plausible extra detail is still a failure of accuracy and scope.

The three actions are correct. A follow-up request asks for reminder emails. Is that correction or going further?

Going further. The original table was already usable; drafting emails is an additional Create task. Correcting a missing action would still belong to reaching the original finish line.

One assistant takes longer to respond but needs less checking and editing. Which is more efficient?

Compare active human effort separately from elapsed speed. The slower response can require less human time. Both observations matter when choosing.

08 · Time

Put A Value On Human Effort

Track the active human time spent reaching the usable result. Separately track the time deliberately spent beyond that finish line. The time-cost calculator values those two stages at the chosen hourly rate.

Ten minutes spent editing is different from ten minutes waiting while other work continues. Record elapsed time separately; count waiting in active time only when it requires attention.

The calculator does not measure quality, reliability, the value of discoveries or AI fees. Use the performance checks first, then use the calculator to understand the time consequences.

Understand The Task Framework Compare Two Assistants Value The Time Spent

09 · Basis

A Practical Framework, Not A Leaderboard

This is the publication’s way of organising real jobs and personal comparisons. It is not a validated benchmark or a claim that one assistant wins every kind of work.

The distinction between task scenarios and multiple performance measures is informed by Stanford’s HELM research. Repeated attempts as a test of consistency are also explored in the τ-bench paper. The six reader-facing task families are an editorial proposal, not categories taken from those studies.