AI adoption in a highly decentralized company needs shared boundaries and services that fit independently managed businesses. Agree what the group requires, what help it offers and which decisions remain with each operating company. Then let local leaders commit to a specific use, with the people and information needed to make it work.
A group-wide tool can support that arrangement. It does not make every business ready for the same workflow or give one company permission to use another's records. For the group adoption lead, the practical question is which parts are worth sharing and what a receiving business must contribute.
Start with the decisions the group can actually make
Decentralization can mean separate management teams, budgets, systems and customer relationships, even within one country. A group AI lead may be able to recommend a service without having authority to change a company's processes or purchase commitments. Establish that boundary with the relevant business and group leaders before announcing a rollout.
Halma's description of its operating model offers a concrete distinction. Its companies have autonomy and their own boards; sector teams connect them with the group, while group teams set risk appetite and provide expertise. The company also describes minimum controls for finance, legal and IT. This is Halma's account of its organization, not evidence about the effectiveness of its AI adoption. It shows why local autonomy should not be confused with an absence of common obligations.
For the proposed AI service, separate three kinds of commitment:
- Requirements that apply. Identify the authorized group or company owner for permitted data, security, purchasing and other relevant controls.
- Shared provision. State what the group will supply, maintain and fund, and which businesses the offer covers.
- Local participation. Confirm who can commit staff time, choose the business task and accept the resulting work.
Do not turn this into an assumption that participation is always voluntary. Some requirements or services may be mandatory under the organization's actual authority. Equally, calling something a group initiative does not create authority that has never been delegated. The central and local ownership guide explains how to assign the service interface once those decisions are clear.
Offer help an operating company can use
An operating-company manager needs to see what the shared offer removes from their workload and what it adds. Approved access, a maintained service, specialist help and reusable learning can be valuable. They still require local time for checking outputs, maintaining business information and helping colleagues change their work.
ABB's April 2026 account of employee AI use describes common provision that includes an approved-tools microsite, live and self-paced training, an employee community and an internal assistant alongside off-the-shelf tools. These are company-reported arrangements, not proof of a particular productivity gain or a model every group should copy. They illustrate a shared offer that extends beyond distributing a software license.
Make the offer answerable from the receiving company's perspective:
- Which task can staff try, using which permitted information?
- What access, support and reusable material will arrive with it?
- What checking, preparation and ongoing work must the company supply?
- Who pays for the trial and continuing use, and who can change that commitment?
Funding needs an explicit conversation. In an August 2026 LinkedIn discussion, adviser Tariq Munir explored central funding for learning and business funding once value is established. He called it an unfinished thought and raised a tension: moving to chargeback too early could discourage experimentation, while leaving it too late could blur ownership. A reply also challenged treating low-consequence drafting and payment approval alike. This is practitioner discussion, not a tested funding rule.
Agree the arrangement before businesses join. The cost-allocation guide distinguishes showing a company its attributed costs from assigning charges to its budget. Neither a license allocation nor an internal invoice establishes that the work is useful.
Let local conditions shape participation
Suppose a group owns independently managed industrial-parts distributors. Claire Bennett, a sales manager at one company, wants an approved assistant to draft explanations for customer quotations. Staff would supply confirmed product details, prices and stock information, then check the draft before sending it through the existing sales process. The assistant would not choose substitute parts, set prices or send messages.
Claire's company maintains product identifiers in a current catalogue linked to its stock records. Another group company uses different identifiers and records some substitutions in handwritten supplier notes. Both sell fittings. That similarity does not make their source material interchangeable.
The group could offer a maintained drafting service while each company keeps its own approved records and sales checks. It could also share the questions used to test whether a draft describes the right item. A successful demonstration at Claire's company would still leave the second company with work to do.
Compare the options against the receiving company's task:
| Participation choice | When it may fit | What must be resolved |
|---|---|---|
Use the shared service | Its approved inputs and output checks fit the local quotation task | Local access, source ownership, staff checking and continuing support |
Adapt the shared method | The task is similar, but product records or sales checks differ | The changed mapping and instructions, local testing and an owner for maintenance |
Retain an approved local alternative | A material requirement is unsupported by the shared offer | The reason, applicable controls, complete cost and when to reconsider |
Prepare before adopting | Neither method has reliable permitted inputs or enough checking capacity | An owner for the prerequisite, the current manual process and a return decision |
A local alternative should have an answerable rationale. “We prefer our own tool” is insufficient when the difference adds cost or leaves a required control unresolved. Conversely, insisting on identical software does not repair missing product records. Resolve the actual gap with its owner before expanding the task.
Transfer a method without copying its permissions
A useful local improvement should be easy for another company to examine. Share the task, permitted example material, checking method and limitations. Keep customer records, commercial terms and system access within their authorized boundaries. Membership of the same group is not, by itself, the permission decision.
In her January 2025 account of research at VINCI, HEC Paris professor Aluna Wang describes innovation support through Leonard, combining technical learning with practical operational projects. AI lead Bruno Daunay explained the motivation in human terms: employees “loved their work, but they were sick of doing it inefficiently and wanted to change it.” The account concerns a specific decentralized group and includes specialized engineering applications. It supports examining local work alongside shared expertise, rather than assuming its project results transfer to an ordinary office assistant.
For Claire's quotation method, involve a receiving manager before calling the work reusable:
Try the transfer in the receiving business
Inspect the difference
Compare product identifiers, approved source records and the checks that protect the receiving company's quotations.
Test a permitted sample
Try the method on that company's material, including a missing identifier or an unconfirmed substitution. Keep the existing sales check.
Agree the next commitment
Review the result and remaining gaps with the receiving owner. Decide whether to join, adapt, continue the approved alternative or resolve prerequisites first.
Record the scope of reuse
Name the receiving task, local owner and maintained shared component. Keep unresolved conditions visible.
The use-case library guide explains how to keep a promising example distinct from a reviewed reuse. Sharing the method should reduce rediscovery, while leaving the next company enough evidence to make its own authorized decision.
Review the portfolio without demanding the same pace
A group view should reveal commitments and unresolved decisions. For each participating business, retain the task, accountable sponsor, current service or alternative, known result and next dependency. A short explanation of why a company has not joined is more useful than marking it resistant without investigation.
Ask whether the work has actually improved. At Claire's company, that means checking whether quotation explanations describe confirmed products and whether staff can complete the task with acceptable checking and rework. At the other distributor, the next useful step may be resolving product identifiers before AI drafting can be tested. Those are different stages, with different evidence.
Also inspect the shared side. If several companies want the same supported capability, decide who funds its maintenance and what local variation it can carry. If a service has little useful uptake, examine task fit, access, local capacity and cost before buying more licenses or making participation a performance contest.
The enterprise strategy guide places these choices in the wider program. For a decentralized group, the immediate action is smaller: make one shared offer concrete, test it with one receiving business and record the commitment both sides accept.
Questions about decentralized AI adoption
Should every operating company use the same AI tool?
A decentralized group may standardize an AI tool where its authorized leaders have a clear service or control reason. The common tool still needs to fit each approved task and its information. A receiving company may need an adaptation or an approved alternative when a material requirement is unsupported. Compare the participation choices and record the reason before adding another purchase or requiring a switch.
Does a shared AI service require pooling every company's data?
No. Sharing AI provision does not by itself require unrestricted access to every company's records. Agree which information the intended task needs and obtain the relevant permissions. Companies can share a method or permitted example without sharing underlying customer or commercial records. Use the receiving-business test to identify the actual source and access requirements.
How should the group handle duplicate AI spending?
First establish what each purchase supports, who committed to it and whether an approved shared option meets the same need. Include checking, support and existing commitments in the comparison. Similar product names do not prove duplication, and different names do not prove useful differentiation. Start with the shared-offer questions, then use the linked cost-allocation guide to agree budget treatment with finance.
What if a small operating company has no AI specialist?
A small company can own a business task while drawing on shared technical expertise. It still needs a local person with time and knowledge to check the work, maintain business information and arrange its use. If those commitments are unavailable, narrow the task or delay participation instead of assigning unsupported responsibilities. Confirm the authority and local commitment before treating a named contact as a complete adoption plan.
