Estimate AI work before it runs

An estimate is a promise about how a run will be governed, not a prediction that every detail is known.

Break work into observable steps

Start by turning a broad request into a short sequence of outcomes. For example, a content task may need a brief, research notes, an outline, a draft, and a review-ready version. Each step has a different amount of context, tool use, and expected output. Estimating the whole request as one block hides the choices that drive cost.

Ask what each step needs to read, what it needs to produce, and whether it can reuse an earlier artifact. This creates a practical unit for setting limits. It also exposes work that should not be automated yet because the input, standard, or approval rule is still unclear.

Set a range, then a stop rule

Use a range rather than a single number. A lower bound can describe a straightforward run with existing context. A higher bound can allow for revision, tool use, or a second pass when quality checks find a gap. The range is useful because it makes uncertainty visible before a workflow starts.

Pair the range with a stop rule. A run should pause when it reaches its allocated budget, exceeds its planned number of steps, encounters a blocked dependency, or needs an action outside its permission scope. A pause gives a person a chance to inspect progress and choose whether the additional work is worth it.

Separate preparation from execution

Preparation can reduce both cost and surprise. Remove duplicate files, choose the few inputs that matter, and write a compact task brief before an agent begins. A focused context often helps more than giving an agent every available document.

During execution, capture actual usage by step. The record should distinguish planning, generation, checking, and external-tool actions. If one stage consumes more than expected, the team can improve that stage without assuming the entire workflow is inefficient.

Compare estimates with completed runs

The value of estimation compounds when completed work returns evidence. Compare the planned range with the actual work, then ask why the result differed. Maybe a review checkpoint caught an issue early. Maybe a research task needed an extra source. Maybe a model spent too much effort on an ambiguous prompt.

Use those observations to tune the next estimate: shorten a repeated prompt, reuse an approved artifact, add a preflight question, or move a review earlier. Do not optimize only for the smallest possible run. The aim is a predictable tradeoff between quality, time, and spend.

A visible estimate turns AI work from an opaque request into a deliberate operating decision. It gives people a chance to approve the scope before resources are committed, and it gives agents a clear condition for asking what to do next.