An AI tool can save time for one employee while making the wider process harder to understand. A draft is produced in a personal account, copied into a shared document and then treated as approved information. The task appears complete, but nobody else knows which sources were used, what was checked or where the final version belongs. This is how a useful experiment can become a shadow workflow.
Introducing AI tools successfully means making their place in the business visible. Staff need a practical route for trying ideas, clear boundaries for information and actions, and a shared way to review the results. The aim is to capture useful improvements without allowing essential work to depend on undocumented personal arrangements.
Find out where AI is already being used
Begin with a straightforward conversation about current tasks. Ask staff which tools help them prepare messages, summarise documents, organise research or draft internal material. Include AI features embedded in existing software, since people may not think of those as separate tools.
Make the conversation useful rather than accusatory. Someone may have found a better way to handle repetitive work that the official process does not support well. Understanding that difficulty gives the business a chance to improve the process and establish appropriate controls.
Record the task, tool, account type, input information and destination of the result. You do not need everyone's entire conversation history to map the workflow. You need enough context to identify which operating decisions require attention and who should make them.
Describe the approved task in plain language
Choose a limited starting use that the team can explain. For example, AI might help prepare an internal draft from approved service information. The final message would still be checked by the person responsible for it before being sent. Record those steps and the intended outcome.
Distinguish assistance from authority. A tool that suggests wording should not silently become the place where staff decide prices, resolve disputed policy or approve customer commitments. Make clear which decisions remain with an accountable person and which existing process records the decision.
Describe what lies outside the starting scope. Staff should know when to stop and ask for help, particularly if the material includes unfamiliar information or the tool will influence an operational decision. A boundary becomes useful when it includes realistic examples from the team's work.
Give each workflow an accountable owner
Name the person who maintains the working instructions and decides whether the process still fits the business. That person does not need to perform every task, but they should know how inputs, review and final actions connect.
Separate responsibilities where necessary. A subject expert may approve factual content, an administrator may manage tool access and a team lead may decide how the output enters the customer process. Record the handovers so that everyone understands which check they are performing.
An AI introduction needs staff skills, organisational policies and clear review and escalation arrangements. Discuss how these fit your business, rather than adopting a policy written for a different organisation without checking its assumptions. Our AI assurance and procurement guide provides the related checks.
Make account and information choices explicit
Identify the approved account and product configuration for the task. Features and information handling arrangements can differ between products and subscription types. Ask the responsible administrator to confirm what applies to the actual tool the team will use, rather than relying on a general statement about AI.
Define the information that may enter the workflow. Use examples of acceptable source material and examples that require advice first. A public service description and an internal customer file are different inputs, even if both can be pasted into the same text box.
Provide a workable route for requests that fall outside the approved arrangement. If staff cannot obtain help or access within a reasonable process, they may return to personal accounts to finish the job. Good governance includes a practical way to meet legitimate working needs.
Keep the source and final record in shared systems
Decide where approved source information lives and where the checked output will be stored. The AI conversation should not become the only place that holds a customer instruction or an important operating decision. Use the existing system of record where it remains suitable.
For a drafting workflow, a useful record might include the approved source reference, the reviewed document and the person who approved it. The level of detail should fit the consequences of the task. Avoid recording everything indefinitely without a clear reason or agreed handling arrangement.
Make status visible. Staff should be able to distinguish a proposed draft, a reviewed version and a final message that has been sent. If these states look identical in a shared folder, the business can accidentally reuse unfinished material even when the original employee followed the process correctly.
Train people through actual examples
Use representative tasks in training rather than a presentation of features alone. Show a good input, a plausible but incorrect result, a review decision and the final approved output. Ask participants to explain what they would check and where they would store the result.
Include cases where the right answer is to seek clarification. A prompt asking for a delivery promise may lack the information needed to make one. Staff should recognise that a confident sentence is not evidence of operational availability.
Teach the whole workflow, including the handover to another person. Someone receiving an AI-assisted draft needs to understand what has already been checked and what remains open. A short working example can reveal assumptions that a lengthy policy document leaves hidden.
Evaluate effort across the whole process
Measure the task from preparation to a usable outcome. Include the time spent gathering sources, checking facts, correcting the draft and transferring it to the proper record. A quick generation step can be attractive while adding substantial review work later.
Choose quality criteria suited to the task. For customer messages, these might include accurate service details, understandable wording and a clear next step. For an internal summary, they might include coverage of important points and honest treatment of missing information.
Compare representative examples with the existing process and record the assumptions. Avoid presenting one successful demonstration as a general productivity figure. The team needs evidence about the tasks it will actually perform and the review it will actually require.
Create a visible route for useful experiments
Let staff propose new uses with a short description of the task, expected benefit, information involved and proposed review. Give someone responsibility for responding. A simple register can show which ideas are being tested, which are approved and which need further work.
Keep experiments distinct from routine operation. A trial should have a limited scope, named participants and a review point. If it succeeds, document the working arrangement before expanding it. If it fails, retain the useful learning without leaving an unsupported tool in daily use.
Share findings with the team. A rejected use may reveal a missing business rule or poor source material rather than a problem with the idea itself. Explaining the reason helps staff understand how to improve the next proposal.
Review access, changes and departure arrangements
Check what happens when a tool changes, an employee moves role or a supplier relationship ends. The team should be able to locate the approved materials, maintain the process and remove access through its normal arrangements. Essential work should not disappear with one person's account.
Schedule reviews around relevant changes as well as ordinary operating feedback. New data sources, additional actions and wider audiences can alter the consequences of a workflow. Treat expansion as a decision requiring evidence, even when the original drafting task worked well.
Giraffe Digital's digital strategy service can help connect tools to customer journeys and operating responsibilities. AI adoption becomes easier to sustain when staff can see the approved process, understand their role and improve it through a shared route.
Frequently asked questions
How should we find out where AI is already being used?
Begin with a straightforward conversation about current tasks. Ask staff which tools help them prepare messages, summarise documents, organise research or draft internal material. Include AI features embedded in existing software, since people may not think of those as separate tools. Make the conversation useful rather than accusatory. Someone may have found a better way to handle repetitive work that the official process does not support well. Understanding that difficulty gives the business a chance to improve the process and establish appropriate controls. Record the task, tool, account type, input information and destination of the result. You do not need everyone's entire conversation history to map the workflow. You need enough context to identify which operating decisions require attention and who should make them.
How should we give each workflow an accountable owner?
Name the person who maintains the working instructions and decides whether the process still fits the business. That person does not need to perform every task, but they should know how inputs, review and final actions connect. Separate responsibilities where necessary. A subject expert may approve factual content, an administrator may manage tool access and a team lead may decide how the output enters the customer process. Record the handovers so that everyone understands which check they are performing. An AI introduction needs staff skills, organisational policies and clear review and escalation arrangements. Discuss how these fit your business, rather than adopting a policy written for a different organisation without checking its assumptions. Our AI assurance and procurement guide provides the related checks.
How should we keep the source and final record in shared systems?
Decide where approved source information lives and where the checked output will be stored. The AI conversation should not become the only place that holds a customer instruction or an important operating decision. Use the existing system of record where it remains suitable. For a drafting workflow, a useful record might include the approved source reference, the reviewed document and the person who approved it. The level of detail should fit the consequences of the task. Avoid recording everything indefinitely without a clear reason or agreed handling arrangement. Make status visible. Staff should be able to distinguish a proposed draft, a reviewed version and a final message that has been sent. If these states look identical in a shared folder, the business can accidentally reuse unfinished material even when the original employee followed the process correctly.
How should we evaluate effort across the whole process?
Measure the task from preparation to a usable outcome. Include the time spent gathering sources, checking facts, correcting the draft and transferring it to the proper record. A quick generation step can be attractive while adding substantial review work later. Choose quality criteria suited to the task. For customer messages, these might include accurate service details, understandable wording and a clear next step. For an internal summary, they might include coverage of important points and honest treatment of missing information. Compare representative examples with the existing process and record the assumptions. Avoid presenting one successful demonstration as a general productivity figure. The team needs evidence about the tasks it will actually perform and the review it will actually require.
How should we review access, changes and departure arrangements?
Check what happens when a tool changes, an employee moves role or a supplier relationship ends. The team should be able to locate the approved materials, maintain the process and remove access through its normal arrangements. Essential work should not disappear with one person's account. Schedule reviews around relevant changes as well as ordinary operating feedback. New data sources, additional actions and wider audiences can alter the consequences of a workflow. Treat expansion as a decision requiring evidence, even when the original drafting task worked well. Giraffe Digital's digital strategy service can help connect tools to customer journeys and operating responsibilities. AI adoption becomes easier to sustain when staff can see the approved process, understand their role and improve it through a shared route.


