Workforce planning when part of the workforce is AI

Most workforce plans still assume that work is done by people in positions. That assumption is breaking. AI agents now draft documents, reconcile data, answer routine enquiries, screen requests and prepare reports, and they do it inside the same processes your employees work in. If your plan counts heads and ignores agents, it is forecasting a workforce that will not exist.

I spend part of my time designing and building AI agents, including one that turns a written brief into editable slides on a client's own template, and the rest advising HR leaders on organisation and workforce design. This article brings the two together. It gives you a practical way to plan work across people and agents, the roles and oversight that planning has to include, and the traps I see organisations fall into.

Plan work, not jobs

Traditional workforce planning treats the job as the unit. You forecast how many analysts, officers or coordinators you need. That works when the content of each job is stable. It stops working when a large share of a job's tasks can be handed to an agent while the rest cannot.

The practical response is to break roles into tasks before you plan. A job is a bundle of tasks held together by history, grading and convenience. Agents unbundle it. Some tasks go to the agent, some stay with the person, and some become new tasks that did not exist before, such as reviewing agent output or handling the exceptions the agent cannot resolve.

This does not mean building a task inventory for every role in the organisation. Start with the job families where agents are most likely to change the work: high-volume, document-heavy, rules-based roles in operations, finance, HR services, procurement and customer service.

Sort every task into one of four modes

For each task in a job family, decide which of four modes it belongs in. This is the core tool in the method.

The deciding question is not "can an agent do this?" Increasingly, the answer is yes for more tasks than people expect. The better question is: what happens if the agent gets it wrong, and would we notice? A small error in a meeting summary costs nothing. A small error in an end-of-service calculation or a disciplinary letter costs a great deal.

Four tests help sort tasks consistently:

1. Volume: is the task frequent enough to justify building and maintaining an agent?
2. Clarity: are the rules, inputs and expected outputs written down, or do they live in someone's head?
3. Consequence: what is the cost of an error to the person affected, the organisation and its reputation?
4. Checkability: can a reviewer verify the output quickly, or would checking take as long as doing it?

A worked example: an HR services team

Say an HR shared services team has 20 people handling employee letters, leave and payroll queries, onboarding paperwork and monthly reporting. The figures below are illustrative.

After breaking the work into tasks, the team finds that roughly half of its time goes on work that fits "agent-led" or "agent drafts, human approves": standard letters, routine policy questions, data entry and report preparation. Another third is "human-led, agent-assisted": complex cases, payroll exceptions, onboarding for senior hires. The remainder is "human only": sensitive employee conversations and anything touching disciplinary or termination matters.

The naive conclusion is that the team can shrink to ten. It cannot, for three reasons. First, the agent-drafted work still needs review time, perhaps a fifth of what it took to do it manually. Second, someone must own the agents: monitor quality, update them when policies change and handle what they escalate. Third, releasing capacity only has value if it goes somewhere.

A realistic plan might look like this over two years. The team moves to around 14 roles. Two of those are redesigned as agent owners who combine process knowledge with responsibility for agent performance. The capacity released is redirected to work the team never had time for, such as proper workforce analytics and manager support. The six roles no longer needed are handled through attrition and redeployment rather than exits, because the plan was made early enough to allow it.

The point of the example is the shape of the answer, not the numbers. Agent capacity creates new human work, and the plan has to show it.

Oversight is workload you have to staff

The most common mistake I see is treating human oversight as free. A project team deploys an agent, declares that "a human stays in the loop", and assumes the reviewing will be absorbed by whoever is nearby. Within months, reviews become a rubber stamp because nobody was given the time to do them properly.

Plan oversight as explicitly as you plan any other work:

• Agent owner. Every agent needs a named person accountable for what it does, the same way a manager is accountable for their team. This is usually a process owner in the business, not IT.
• Reviewers. For "agent drafts, human approves" tasks, estimate review time per item and include it in capacity.
• Exception handling. Agents escalate what they cannot resolve. Those cases are, by definition, the harder ones, and they need experienced people.
• Maintenance. Policies, templates and systems change. Someone has to update the agent's instructions and test the result.

An agent that nobody owns is not a digital worker. It is an unmanaged risk.

There is a useful parallel with span of control. A manager can only supervise so many people well. A reviewer can only meaningfully check so much agent output in a day. When the volume of output outgrows the review capacity, quality control quietly disappears. Build that limit into the plan.

Redesign roles, grades and career paths

When tasks move to agents, the jobs that remain change. That has consequences HR must manage deliberately.

Job descriptions and grading. A role that loses its routine content and gains review, exception handling and ownership responsibilities may deserve a different grade. Re-evaluate redesigned roles properly rather than leaving old grades in place. Equally, some roles lose the complex parts of their work and should not keep a grade that no longer reflects them.

Skills. The remaining work demands more judgement, more subject expertise and the ability to specify, question and correct an agent's output. That is a different skill profile, and your development plans should reflect it.

Career ladders. This is the issue I would put at the top of any GCC leadership agenda. The tasks most suited to agents are often the tasks that entry-level employees learn on. Remove them and you remove the first rungs of the ladder. In a region where graduate roles are a major route for national workforce programmes such as Emiratisation and Saudisation, automating entry-level work without redesigning how people learn can undermine the national talent pipeline you are required to build. Plan structured apprenticeship-style roles where juniors learn by reviewing and correcting agent output under supervision.

Govern the boundary between people and agents

Decide, in writing, which decisions agents may make, which they may prepare, and which they may not touch. In HR, decisions with serious consequences for a person's employment should remain with an accountable human: hiring and termination, disciplinary outcomes, final performance ratings, promotions and pay decisions. Agents can gather information, draft and flag issues. A named person decides.

Audit agent output regularly for accuracy and for bias, especially where it touches recruitment or assessment. Make it easy for employees to raise concerns about an agent's work without it being seen as resistance. The wider questions of policy, ownership and accountability are covered in governing AI at work.

Treat the transition as organisational change

Agents fail more often through adoption than through technology. People who suspect their jobs are being quietly automated will work around the agent, withhold the knowledge it needs or check its output so heavily that no time is saved.

Three things make the difference. Be clear and early about intent: what will change, what will not, and how affected people will be treated. Build practical AI literacy so employees can use and challenge agents in their own work, rather than sending them to a single awareness session. And start with pilots in one job family, measure honestly, and scale what works rather than announcing an enterprise-wide programme before anything has been proven.

This work fits inside your broader workforce planning cycle. The agent decisions become one of the options for closing gaps, alongside building, buying, borrowing and redeploying people, as I set out in strategic workforce planning in the GCC.

Where to start

1. Pick two or three job families where agents are most likely to change the work, and break each into its main tasks.
2. Sort those tasks into the four modes using the volume, clarity, consequence and checkability tests.
3. Estimate the oversight workload, and name an owner for every agent before it goes live.
4. Re-evaluate the redesigned roles and check what happens to entry-level learning and your national talent pipeline.
5. Run one pilot, measure time saved net of review, and use the result to update your workforce plan.

At Humanyx we design people and agents together, so if you want help planning the roles, the oversight and the agents themselves, that is where we can support you.