Key points
- Decide collaboration at the level of tasks, not jobs or whole processes, because most OD roles contain both kinds of work.
- Use three modes of collaboration, where the agent drafts, advises or executes, and match each task to a mode based on risk and judgement.
- Decisions that affect people's careers, pay or employment stay with accountable humans, and the agent's role is to prepare the evidence.
- Invest in the skills of the OD team to brief, review and challenge agent output, since that is now a core capability.
Most HR teams I speak to are already using AI in some form, often without a plan. Someone summarises survey comments with a chatbot, someone else drafts job descriptions, and a vendor adds an AI feature to the learning platform. None of it is wrong, but none of it adds up to a way of working, and nobody has decided which judgements should stay with people.
Organisation development is a particularly important place to get this right, because OD work shapes people's roles, careers and culture. This article sets out a practical framework I use for designing how people and AI agents work together in OD: how to break work down, which collaboration mode to use, where to draw firm lines, and what it means for your team.
Why OD needs a deliberate collaboration design
OD is a mix of two very different kinds of work. There is heavy analytical and drafting work: reading thousands of survey comments, mapping positions to a new structure, writing role profiles, comparing competency frameworks, building workshop materials. And there is work that depends on trust and judgement: helping an executive team agree a structure, handling a difficult conversation with a leader whose role is changing, sensing when a culture programme is losing credibility.
AI agents are now good at a large share of the first kind of work. They are not good at the second, and they should not be asked to do it. The problem is that both kinds of work are mixed together in the same projects and often in the same afternoon. Without a deliberate design, teams either underuse agents out of caution, or let them drift into decisions they should not be making.
I build agents myself, including one that turns a written brief into editable slides on a client's template. The lesson from that work applies directly to OD: the value comes from being precise about what the agent does, what the human does, and where the handover sits.
Break the work into tasks before you decide anything
The most useful unit of analysis is the task, not the job or the process. "Org design" is too broad to assign to anyone. "Produce a first draft of the span-of-control analysis for the operations division" is specific enough to decide who, or what, should do it.
For any OD programme, list the tasks involved and ask three questions about each.
- Does it require judgement about people or values? For example, deciding who fills a role, or how to handle a sensitive change message.
- What happens if it is wrong? A weak first draft of a workshop agenda costs an hour. An incorrect grading recommendation can cost trust and money.
- Is the input data good enough? An agent working on an inaccurate position list will produce confident but wrong results.
These three questions tell you which collaboration mode fits the task.
Three collaboration modes
In practice I use three modes. They are simple enough for a whole HR team to apply consistently.
Agent drafts, human decides. The agent produces a first version and a person reviews, edits and owns the result. This fits most writing and analysis: role profiles, policy drafts, summaries of focus groups, first-cut structure options.
Agent advises, human judges. The agent analyses data and surfaces patterns, risks or options, and a person weighs them alongside context the agent does not have. This fits diagnosis work: engagement analysis, skills gap mapping, identifying spans and layers issues.
Agent executes, human checks. The agent carries out a defined, repeatable task end to end, and a person reviews at set checkpoints or on a sample. This fits high-volume, well-specified work: formatting job descriptions to a standard template, checking an org chart against the HR system, preparing packs for a talent review.
A fourth option is always available: the human does it, and the agent does not touch it. That is the right answer for more tasks than enthusiasts admit.
A decision table for common OD tasks
Here is how I would assign some common OD tasks. Your own answers will depend on data quality, risk appetite and the maturity of your tools.
| OD task | Recommended mode | Why |
|---|---|---|
| Summarising open-text engagement or pulse survey comments | Agent advises, human judges | Agents are fast at clustering themes, but people must interpret them in context and check for misreadings |
| Drafting role profiles and job descriptions | Agent drafts, human decides | Good first drafts save time, but accountabilities must match the actual design |
| Analysing spans of control and layers | Agent advises, human judges | Pattern finding is mechanical, but whether a narrow span is a problem depends on the work |
| Generating structure options for a redesign | Agent drafts, human decides | Useful for breadth of options, but choosing a structure is a leadership decision |
| Mapping employees to new roles in a restructure | Human does it | Directly affects individuals' jobs and must be accountable and explainable |
| Recommending grades in job evaluation | Agent drafts, human decides | Agents can prepare a first view, but a trained panel must evaluate and own the result |
| Building personalised learning paths | Agent executes, human checks | Low risk per decision and high volume, with L&D reviewing the logic and a sample |
| Preparing talent review packs | Agent executes, human checks | Assembling data is mechanical, but the potential ratings in the pack come from people |
| Designing and facilitating a leadership workshop | Agent drafts, human decides | The agent can draft materials, but the facilitation is human work |
| Communicating role changes to affected employees | Human does it | Requires empathy, accountability and the ability to answer hard questions |
The agent prepares the evidence, and a named person makes the call.
That principle runs through the whole table. Any decision that affects someone's role, pay, grade, rating or employment should have a named, accountable person who can explain it. In the GCC, where many organisations are government or semi-government entities with formal approval chains, this also fits existing governance. The agent sits inside the approval process, never around it.
Learning labs, skills hubs and immersive tools
OD and L&D teams are also being offered a steady stream of new learning technology. Two categories deserve a clear position.
AI-powered skills and learning platforms. These map employees' skills, suggest learning content and sometimes recommend internal moves. The idea is sound, particularly for organisations trying to build national talent or redeploy people as roles change. The weakness is the skills data. If your skills taxonomy is vague and employee profiles are self-reported and out of date, recommendations will be poor. Fix the taxonomy and the data first, then turn on the recommendations, and keep L&D reviewing what the platform suggests.
Virtual and augmented reality. Immersive tools are genuinely useful for a narrow set of purposes: safety training in hazardous environments, practising complex operational procedures, and rehearsing difficult conversations with realistic role play. In sectors like energy, utilities, aviation and construction, which are significant employers across the region, that can justify the investment. For general leadership or knowledge content, the cost and logistics usually outweigh the benefit. Pilot on one clear skill gap, measure whether behaviour changes, and scale only if it does.
In both cases, the same collaboration logic applies. The technology delivers and personalises, and people decide what capability the organisation needs and whether the learning is working.
Governance and trust
Employees will judge your use of AI by how open you are about it. A few practical rules go a long way.
- Tell people where agents are used in processes that affect them, such as survey analysis or learning recommendations, and what humans still decide.
- Keep a register of OD and HR tasks where agents are used, the mode, and the named owner of each.
- Protect personal data. Be clear about which employee data agents can access, and apply the stricter standard where local data protection rules and your own policies differ.
- Review for bias periodically, by checking whether agent outputs differ systematically across groups such as nationality, gender or age.
For the wider questions of policy, ownership and oversight, see governing AI at work.
What this means for the OD team
The skills an OD practitioner needs are shifting. Writing a first draft matters less. Writing a precise brief, reviewing output critically, spotting a plausible but wrong conclusion, and explaining a recommendation to an executive all matter more.
I encourage teams to treat this as a capability programme, not a tool rollout. Give practitioners real OD tasks to work on with agents, review the results together, and agree team standards for what good agent-assisted work looks like. Over time, this also changes how you plan the team's size and roles, which I discuss in workforce planning when part of the workforce is AI.
Where to start
- List the tasks in one live OD programme, such as a restructure or an engagement cycle, and apply the three questions to each.
- Assign each task a collaboration mode using the decision table as a starting point, and name the human owner for every task involving an agent.
- Pick two low-risk, high-volume tasks and run them with an agent for one cycle, reviewing quality against the previous approach.
- Write a one-page statement for employees on where AI is used in HR processes and what people still decide.
- Run a short working session with your OD team on briefing and reviewing agent output, using real work from their current projects.
At Humanyx we design OD programmes and the AI agents that support them together, so the people and the agents each do the work they are best suited to.