Which work tasks should you delegate to AI?

Start with the decision the work needs to support

You ask AI to compare options. It returns a clear table and a confident recommendation. The table may be useful, but accepting the recommendation is a separate decision.

This is where human–AI collaboration becomes practical. Instead of asking whether AI can do your job, name the next piece of work: choose a content topic, prepare a product brief or compare suppliers. Then separate the steps that organise information from the steps that interpret it or commit you to an action.

At Altiv, our AI Readiness research starts with tasks rather than treating a role as one indivisible activity. Our view is that a task split should tell you what to do next: what to hand over, what to work through with AI, and what you need to check or decide.

What the research tells us and what we add

Anthropic's first Economic Index studied how people used Claude, distinguishing direct task execution from collaborative patterns such as iteration, learning and validation. That supports looking at how AI participates in a task, rather than treating all AI use as automation. It describes usage on that platform; it does not certify that an output is reliable or that a whole role can be delegated. Read the original Economic Index. ↗

The ILO's 2025 occupational-exposure study also works at task level. Its assessment points towards changes within jobs rather than assuming every exposed job disappears. Exposure is a different question from whether a particular output is ready to use. Read the ILO study summary. ↗

Altiv adds the practical interpretation: given this task and these inputs, where would AI assistance be useful, and what would make its output usable? Those recommendations are our assessment, informed by research. They are not rules issued by the research authors.

There is also a more specific marketing example in the research. Anthropic's March 2025 report observed substantial back-and-forth iteration in writing and editing tasks. Our local review of its task data likewise found more collaborative than task-completion patterns for preparing reports of research findings. This is historical evidence of how people worked with Claude, not a current estimate of the share of a marketer's job that AI performs. Read the March 2025 report and access its task data. ↗

A real example: choosing what to write next

We used this approach while researching Altiv's own content plan. The decision was which topic to investigate first for professionals learning to work with AI.

Varun supplied Google Keyword Planner results for the United States, all languages, covering October 2025 through September 2026. Three candidate terms looked like this:

These are keyword estimates, not expected visits to an Altiv article. We did not add them together: related searches may overlap.

AI assisted with recovering the pasted values, comparing the topics and inspecting existing answers. Varun confirmed the location, language, reporting period, currency and column meanings. Without those checks, a tidy table could have carried the wrong interpretation.

The comparison then went beyond the numbers. Atlassian's AI ROI guide addresses organisational leaders and offers an assessment spanning several kinds of business value. A narrowly focused article about one person's time savings would not answer that entire need. See Atlassian's AI ROI guide. ↗

Microsoft already provides practical guidance for choosing which tasks to give Copilot, including impact and how easily errors can be detected. Another generic delegation checklist would add little. See Microsoft's task-selection guidance. ↗

The resulting recommendation was to investigate task-level delegation first, supported by a complete example readers could inspect. AI ROI had the highest volume of these three terms, but volume alone did not determine the choice.

That was a content decision, not a traffic result. We have not established that the selected topic ranks better or brings more customers. The useful outcome was a clearer brief and an explicit reason for choosing it.

Split the work before you delegate it

Here is how we would divide that task for a repeat run. This is Altiv's recommended workflow, based on the research exercise. It is not a claim that every step ran without supervision.

The important boundary is between proposing a recommendation and accepting it. AI can help with both the comparison and the reasoning. You still need a basis for deciding whether that reasoning fits the work.

Enlarge image ↗Open original ↗

Close image ×

Use three questions to choose the split

Can I give AI the inputs this step needs?

An instruction to organise supplied records is different from an instruction to discover missing facts. Be explicit about which information is available and what must remain unknown.

In our example, a dash in the keyword table stayed unavailable. It did not become zero searches. If the input is incomplete, the useful AI output may be a list of questions rather than a finished recommendation.

Can I check the result before someone uses it?

Name the check before delegating the step. A calculation can be recomputed. A summary can be traced to source passages. A recommendation needs criteria against which you can assess it.

If you cannot tell whether an answer is sound, asking AI for more confidence does not solve the problem. Narrow the task, find someone who can review it, or use AI to help you understand the material before relying on the result.

What changes if the output is wrong?

Drafting three alternatives leaves room to revise. Publishing a claim, approving a purchase or committing a team to a roadmap has consequences beyond the draft.

Use that difference to decide where review belongs. A step that is suitable for AI assistance may still need approval before it changes anything outside the working document.

The same approach in marketing, product and operations

The following are Altiv recommendations for applying the method, not measured findings about every person in these roles.

These are examples of AI-augmented work: the professional uses AI's output while continuing to interpret evidence and shape the result. The same task could involve more delegation when the inputs and checks are stable, or less when the situation changes.

For marketing, ONET's Search Marketing Strategists profile includes examining search behaviour and analysing research to understand intent. That helps define the work. The proposed division between AI and the professional is Altiv's interpretation, not an ONET automation rating. See the occupational task profile. ↗

Make a plan for one task you already do

Choose a piece of work you expect to repeat. Write the intended result, divide it into a few steps, and name the AI assistance, your action and the check for each step. Use the planner to assemble and copy that plan.

Your plan does not have to delegate everything. A useful first attempt might hand over one well-defined step and make the check clearer than it was before.