Before deploying AI, unions and works councils should be able to examine the proposed work, its effects on employees and the decisions still open to change. Discuss the actual workflow, data and monitoring, responsibility for mistakes, workload and skills, and the conditions for a pilot or wider rollout. Record what changes as a result, who must act and what remains unresolved.
Start by identifying the applicable representation process. Consultation, collective bargaining and formal participation rights give representatives different roles. A presentation of a finished purchasing decision does not answer the same question as a proposal employees can still influence.
Establish the process before setting a rollout date
Ask the deployment owner and employee representatives to establish which law, collective agreement, works agreement and internal commitments apply to the proposed use. Bring the actual task and system capabilities to that discussion. The relevant process may depend on what the system does, the employees affected and the existing agreement, rather than the vendor's description of the product.
In UK guidance, Acas distinguishes consultation from collective bargaining. Consultation involves talking and listening about changes; collective bargaining involves an employer and recognized union negotiating agreements on matters such as pay and employment terms. Do not assume a consultation meeting replaces a negotiation required by an existing agreement.
Germany provides a specific example of planning obligations. Section 90 of the Works Constitution Act expressly includes AI in planned work processes and workflows. It requires timely information and necessary documents for the works council, with consultation on the measures and employee effects early enough for proposals and concerns to be considered in planning.
Section 87 separately provides for co-determination, subject to its stated conditions, including the introduction and use of technical equipment intended to monitor employee behavior or performance. These provisions do not establish a universal rule that every AI tool needs the same approval. Have someone qualified in the relevant jurisdiction assess the actual proposal and applicable agreements.
Set a process that leaves time for representatives to understand the proposal, consult the people affected and respond. Acas's UK consultation guidance describes that work as part of a genuine two-way discussion. A vendor demonstration alone cannot give representatives the missing workplace information.
Show the work and its consequences
Prepare a short proposal that someone outside the project team can read. Include the current task, the proposed change, the employees affected and a realistic demonstration. Make unresolved decisions visible, including whether the use should proceed at all.
Suppose a hotel group proposes AI-assisted prioritization of maintenance tickets. Facilities manager Alex Morgan brings two sanitized repair requests to a meeting with elected worker representative Claire Bennett and technician Sam Taylor. One concerns a water leak; the other asks for a paint touch-up.
The output ranks the tickets, but Sam asks how the proposal handles access restrictions, specialist skills and time spent moving between buildings. Claire asks whether timestamps will later be used to compare technicians' performance. Alex has to establish both the operational decision and the permitted reporting before the trial can be evaluated.
Use examples that reveal consequences as well as the tool's intended benefit. Show what happens when information is incomplete, an output is wrong or a worker disagrees. For the hotel, that means explaining who can change a priority and who remains responsible for safe maintenance. Do not make a technician's ability to correct the output depend on quietly working around the system.
Separate useful information from employee monitoring
A proposal to improve work allocation can also create records about individual employees. Review the inputs, outputs, reports and access permissions together. “We are only looking for bottlenecks” leaves too much unexplained if a dashboard also compares named workers.
In a July 2026 Reddit discussion, a Reddit user described a proposed pilot combining ticket, calendar and project records to identify overloaded teams, while apparently producing individual productivity scores. Another Reddit user involved in negotiating a similar agreement said they were seeking group-only analysis without tracing results back to individuals. These are anonymous accounts of proposals and negotiations, not verified outcomes.
That distinction gives the meeting a concrete question: which reports are needed for the stated purpose, and which uses are excluded? Group reporting also needs scrutiny. A small team, an unusual task or access to other records may make an individual recognizable. A group-size threshold alone does not establish anonymity.
The following discussion topics are a starting point to adapt to the deployment, not a substitute for its applicable legal process.
| Topic | Evidence to bring | Decision to record |
|---|---|---|
Purpose and scope | A real task, affected roles and the proposed output | Permitted uses and excluded decisions |
Data and reporting | Input fields, sample reports, access and retention arrangements | What is collected, who sees it and whether individual evaluation is allowed |
Worker authority | A wrong output and the correction path | Who can challenge, override or stop the use |
Workload and skills | Changed tasks, review work, staffing assumptions and learning needs | Time, support and responsibilities for the changed work |
Pilot and expansion | Proposed participants, duration, evaluation and system changes | Conditions for starting, stopping, changing or extending the use |
Ask for information that allows the representatives to test the proposal. If the deployment team cannot explain a report or a vendor cannot establish what data is retained, record that gap and its owner. Do not convert an unanswered question into an assurance.
Agree on authority and the work of transition
“Human oversight” needs a practical description. Ask:
- Who sees the output before it affects someone?
- Can they disagree without being penalized?
- Can an employee correct source information or challenge a decision?
- Who investigates an error, and what happens while it is being resolved?
A reported historical agreement illustrates how specific these terms can be. In a 2023 ver.di interview, IBM group works council chair Frank Remers described a 2020 framework agreement. For a skills-recommendation use, he described restricted access and an employee's ability to ignore the recommendation. He also described an ethics board reviewing applications. This is a representative's account of negotiated arrangements, not evidence that those terms apply everywhere or describe IBM's current practice.
Representatives may also seek binding commitments. Cornell ILR's account of a union toolkit for the port sector describes proposals for defined deployment scope, joint technology committees and stronger bargaining rights. These are proposals for negotiation, rather than general legal requirements or proof of successful implementation.
Discuss the work that remains after automation. Reviewing outputs, dealing with exceptions and maintaining information take time and judgment. If the hotel expects technicians to check priorities and explain every correction, the proposal should account for that work. Faster ticket sorting does not establish that fewer staff can safely cover the same buildings.
Keep transition questions distinct enough to resolve:
- Job content: Which tasks disappear, expand or become more demanding?
- Learning: What paid work time, support and practice are available?
- Staffing and hours: What changes are proposed, and which assumptions need examination?
- Employment terms: Which matters require negotiation or another applicable process?
Use the separate guide to communicating about job uncertainty when preparing employee communications. Clear messages should reflect the decisions and uncertainties established here.
Record changes, disagreement and pilot boundaries
Calling a deployment a pilot tells representatives little about its consequences. Establish who participates, what information is used, which outputs affect real work and what happens when the trial ends. Check the applicable process before using employee data or operational outputs; the pilot label should not stand in for that assessment.
For the hotel example, a discussion could produce a narrower proposal: test prioritization using sanitized requests first, examine the exception path with technicians, and resolve reporting permissions before using live employee records. That is a possible local decision, not a universally sufficient safeguard or approval process.
Keep a decision record that distinguishes three states:
- Agreed changes: The proposal changed, with a named owner and completion condition.
- Open questions: More information or a further decision is needed, with a response owner and date.
- Disagreement: The parties' positions remain different, with the applicable next step recorded.
Include what requires a further discussion, such as a new data source, performance report, affected role or use of the output. A time-limited trial should not silently become permission for a different deployment.
Acas advises explaining the decision after consultation, including how feedback was considered. Tell representatives which suggestions changed the proposal and why others were not adopted. Where an agreement or formal participation process is required, establish its outcome separately rather than treating a meeting note as completion.
The people and change guide places this discussion alongside learning, management and employee trust. Representative participation should be visible in the resulting work and decisions. Attendance at a presentation does not demonstrate agreement or a better rollout.
Questions and answers
Does every AI deployment need works council approval?
The answer depends on the jurisdiction, the system's actual use and applicable agreements. In Germany, planning consultation under Section 90 and co-determination under Section 87 are distinct provisions with different conditions. They do not justify assuming one universal approval rule for every AI product.
Bring the proposed task, monitoring capabilities and employee effects to a qualified assessment. The process discussion explains what to establish before fixing a rollout date.
What should representatives receive before an AI consultation?
Provide a readable proposal describing the current task, the proposed change, affected employees, data, outputs and responsibilities. Include a realistic demonstration, reporting examples and the decisions still open. Explain what happens when an output is wrong or challenged.
Ask what additional information and time representatives need to consult workers and respond. The discussion table connects each topic to evidence and a decision to record.
Can an AI pilot begin before consultation or negotiation is complete?
Do not assume that calling the use a pilot removes applicable duties. Establish which process and agreements apply to its data, outputs and employee effects before starting it. A demonstration using sanitized examples and a live trial affecting employee work can raise different questions.
Define participants, scope, authority and stopping conditions. The pilot discussion shows how to record unresolved matters without implying that a trial is automatically approved.
What if management and employee representatives disagree?
Record the specific issue, each position and what information could resolve it. Distinguish an unanswered question from a rejected proposal or an unresolved negotiation. Identify the next step under the relevant law, agreement or consultation process, with a responsible owner.
Explain how feedback changed the proposal and the reasons for decisions. Use the three-state decision record to keep agreed changes, open questions and disagreement separate. A generic claim that everyone was consulted does not resolve the underlying issue.
