A consultancy introduces an AI assistant to help staff prepare technical reports. Experienced consultants use it to assemble a first draft, check sources, and refine the recommendation. A new employee receives the same tool but may not yet know which omissions matter. The technology is shared; the learning needs, review responsibilities, and risks are different.
Preparing an organisation for AI-assisted work means understanding those differences. Training people to operate a tool is useful, but it does not decide how jobs will change, how expertise will develop, or who is responsible for an accepted result. These questions need to be settled with the people doing the work and revisited as evidence from actual use becomes available.
Describe the Task Changes Before Renaming the Roles
A report contains several kinds of work: gathering evidence, interpreting the brief, analysing options, drafting, checking claims, and agreeing recommendations. An assistant might help with some of these activities while adding verification or coordination work elsewhere. Calling the whole job “augmented” does not explain the resulting role.
Map what employees currently do and what the proposed process expects them to do. Include informal coaching, discussions with clients, and the small checks that prevent errors. The article on tacit knowledge explains why documented procedures may not capture everything that supports competent performance.
Be specific about responsibility. A consultant might remain responsible for the report while the assistant proposes wording and retrieves candidate material. The reviewer needs to know which facts were independently checked and which questions remain open. Responsibility can be distributed across roles without leaving acceptance of the final deliverable ambiguous.
Use the task map to identify what changes for different experience levels. Senior staff may need support evaluating new failure modes; junior staff may need structured practice in the underlying analysis. Neither group should be treated as uniformly capable or incapable simply because of its familiarity with the tool.
Involve Employees While the Design Can Still Change
People performing a task can identify missing context, awkward handoffs, and incentives that a demonstration may conceal. Involve them before the workflow and procurement decisions are fixed. Ask what would make the tool useful, what evidence they need to check it, and which errors would be difficult to recover from.
Participation should have a route to influence decisions. If staff identify an unmanageable review requirement, someone needs authority to change the scope, staffing, interface, or release plan. Collecting comments without a response process can produce the appearance of consultation while leaving practical problems unresolved.
The OECD’s 2023 review of AI and job quality reports evidence on worker outcomes, including associations involving consultation and training. These findings support examining how AI is introduced, while not proving that consultation alone causes every desirable outcome.
Include employees who may be less visible in a pilot: staff on different shifts, part-time workers, people with accessibility needs, and those receiving escalated work from the new tool. A system that benefits the drafting team may increase the burden on reviewers or client-support staff. Their experience belongs in the design evidence too.
Train for Decisions and Errors in the Actual Job
Useful training gives staff practice with representative tasks and the evidence needed to complete them. For report writing, this might include identifying an unsupported assertion, checking whether a citation supports a claim, resolving conflicting source versions, and recognising that an apparently complete draft has not answered the client’s question.
Teach what the tool actually does with information. Staff should know whether it retrieves documents, calls a business service, retains conversation context, or submits an action for execution. A correction in a chat does not necessarily update the underlying record or become a training example. The distinction matters when someone believes they have fixed a problem.
Include permitted data use and access boundaries. Employees need an approved route for working with sensitive material, along with practical instructions for situations where the chosen tool is unsuitable. A broad instruction to “use judgment” leaves too much uncertainty if policies and supported alternatives are unavailable.
Provide protected time and support. Learning to check an AI-assisted report is work, and early attempts may be slower. Managers should allow for that rather than measuring trainees immediately against an assumed productivity gain. Training should also produce feedback to the implementation team when failures arise from the workflow rather than user understanding.
Preserve the Development of Domain Expertise
Routine work often provides examples from which people learn. Removing parts of it can change the route to expertise, but skill loss is not inevitable. AI can also support feedback, expose useful examples, or help employees attempt tasks with guidance. The organisation needs to design the learning path rather than assuming either automatic improvement or unavoidable decline.
For junior consultants, require selected exercises in building an argument from source material before comparing it with a generated draft. Discuss omissions and corrections with an experienced reviewer. The objective is not to prohibit assistance generally; it is to preserve opportunities to develop the judgment the role requires.
Keep access to specialists for unfamiliar cases. A tool can make employees feel able to work beyond their experience without providing the feedback needed to recognise subtle errors. Define when a report requires specialist review and make obtaining that review practical. The same principle applies when an experienced consultant enters a new domain.
Measure learning over time using relevant work samples, correction quality, and the ability to explain a decision with evidence. Tool usage and prompt fluency are incomplete indicators of competence. The organisation needs people who can perform the role under normal conditions and recognise when additional help is required.
Allocate Review Work and Authority Explicitly
An AI-generated report can move drafting effort into verification. Establish who checks facts, who evaluates the analysis, and who accepts the result. Avoid silently treating every employee as responsible for every possible model failure without the time, information, or authority to address it.
The human oversight guide examines these conditions in detail. For the consultancy, reviewers need direct access to relevant sources, a route to reject or return unsuitable work, and realistic time allowances. A sign-off field does not establish that these conditions exist.
Consider the experience of workers receiving the residual cases. If an assistant handles simpler enquiries, people may receive a higher proportion of ambiguous or emotionally demanding work. Evaluate workload and recovery time using that changed mix, rather than preserving the old volume target unchanged.
Keep error reporting separate from an automatic assumption of individual fault. Investigate the model, data, interface, workload, training, and policy as well as the person’s actions. This supports learning while leaving room to address actual misuse or negligent behaviour on the evidence. The organisation remains responsible for the conditions it designed.
Evaluate Collaboration Rather Than Adoption Alone
More logins and accepted suggestions do not establish better work. Compare accepted quality, completion effort, rework, client outcomes, and the experience of employees. Include discarded outputs and cases where the existing process remains more appropriate. The operations article follows those measures across the complete service.
Evidence from one role should not be treated as a universal effect. The published customer-support study found productivity gains that varied substantially with worker experience and skill. It illustrates why evaluation should examine who benefits and under which conditions, rather than assuming identical gains for all staff.
Observe whether employees can identify and correct deliberately included errors in a safe test setting. Ask them to explain what the assistant did and what their approval means. These checks can reveal misunderstanding without requiring harmful experiments on live clients.
Review results with the participants and agree the next change. A pilot may support expansion for one task, further training for another, or a decision to retain the existing method. Preparing for collaboration includes the ability to narrow or stop a use that is not meeting its requirements.
Communicate Workforce Consequences Without Inventing Certainty
Task automation does not translate directly into a proportional reduction in headcount. Demand, coverage, service standards, new work, and review effort all affect staffing. Equally, a programme described as collaboration can still lead to fewer roles. Employees deserve a clear account of what has been decided, what remains under evaluation, and how changes will be handled.
The ILO’s 2025 study of generative-AI occupational exposure evaluates tasks that could be affected by the technology. Potential exposure is distinct from observed displacement. Use such research to inform planning, alongside evidence about the organisation’s own workload and choices.
If responsibilities change, update role descriptions, training, performance expectations, and support arrangements. Where employment terms or staffing may change, follow the applicable legal and contractual requirements for that situation. A general technology forecast is not a substitute for those obligations or for direct communication.
The consultancy’s goal should be a working arrangement employees can understand: what they own, what the assistant contributes, how quality is checked, and how expertise develops. That makes preparation a continuing part of management and professional practice, supported by evidence about the work people are actually doing.
