What information does the model actually need?
Start by examining a typical enquiry before using AI with client data. A service manager wants to use nenna.ai to draft replies to complaints. A message contains a buyer’s name, an order number, a product line and a reference to an unpublished supply agreement. A useful reply requires the grounds for the complaint and the agreed procedure. Forwarding the entire conversation would disclose considerably more. Start with a fictional example and establish which details influence the quality of the answer. Ask sales to explain why a particular contract provision might matter even when the other party’s name can be removed.
Effective masking preserves relationships. Two distinct contacts should remain distinguishable in the draft. A date might determine whether a complaint arrived in time. Removing a name may be insufficient when the location, product line and contract date together identify a customer. Have someone familiar with your customers review the proposed input. That second reading reveals which details carry particular significance in your industry.
Which data belongs in the pilot?
Obtain three examples from a department handling frequent external enquiries. One should contain public information, another internal details and a third a mixture of customer context and commercial confidentiality. Describe the purpose of AI use for each in one sentence. “We want to save time” offers little guidance for data selection. “We want to acknowledge receipt politely without assessing the merits of the complaint” sharply limits the context required. Test which details you can remove entirely before applying masking.
For an internal tender, industry vocabulary may matter while the precise site and contacts do not. A technical complaint may present the opposite situation. A serial number might be essential for an internal lookup without belonging in an external writing process. Record the decision for each stage. Otherwise approval to draft an acknowledgement can easily be misread as approval to analyse all attachments.
Turn an approved-tool list into working guidance
A list of permitted applications helps only if staff can recognise their own case. Write short instructions for one approved workflow describing the task, permitted information and a reviewed result. Include a situation in which work should stop. Someone facing a technical report with confidential attachments needs an accessible decision maker. A general request to be careful leaves the original problem unresolved.
Return to users after several weeks. Which tasks do they still complete through personal accounts? Where was the approved alternative cumbersome? Which warnings did they understand and which did they routinely dismiss? Use the answers to improve training and rules. This turns indications of shadow AI into observable organisational changes. Success means that teams under time pressure know a dependable route and actually use it.
Who decides whether a complaint is suitable for AI?
Give sales and service the same fictional complaint. Sales wants to reflect the tone of the relationship. Service needs the year of manufacture and a description of the fault. Have them mark the details they need independently. Compare their choices with the text you intend to pass to the model. Remove any information for which nobody can give a business reason. Keep the original customer message and the trial draft in their respective approved locations. Someone who only drafts a reply may have no reason to read the full agreement.
Then try a harder case. Two customers ordered identical equipment but negotiated different maintenance terms. Can the person handling the request still find the right commitment using the reduced information? If not, a human should add that part from the authorised system. A polished reply offering free repairs to the wrong customer clearly fails this test.
Three tests for your pilot
Preserve names as distinct roles
Create a fictional complaint involving two contacts. One placed the order and the other authorises replacement. Read the masked input without seeing the original. Can responsibilities still be distinguished? Compare the draft against the original allocation of roles. Swapping the approver is a substantive error even if every name was concealed.
Check indirect identification
Use a case in which a small location, rare product and precise delivery date together point to one order. Remove details one by one and observe when the reply becomes unusable. This boundary gives the business and privacy teams something concrete to discuss. Document which class of enquiry should follow a different processing route.
Discuss shadow AI with the people doing the work
Ask two teams which tasks they currently perform with tools they chose themselves. The approved route may be slow or unfamiliar. Compare their experience with findings from your inventory. Develop an approved alternative for one frequent task and ask the same people whether it is practical during a normal working day.
Preparing your team to use the tool
Rehearse an actual complaints workflow with fictional cases
Start with the decision the service desk needs to prepare. Should the reply offer a replacement, ask for a missing detail or simply acknowledge receipt? Write a short incoming message and an expected reply for each situation. Invent customer names, order references and delivery dates. Sales and support can then review the same examples together. Keep the expected replies separate from the input. A reviewer should be able to tell whether the proposed decision follows from the complaint or whether the draft merely sounds plausible because the reviewer already knows what the answer ought to be.
Give a second colleague only the version intended for processing. Can they still tell which part is missing and when the customer needs it? Could they identify the customer from an unusual combination of delivery location, industry and product? Record those observations separately. Remove one detail and run the same case again. This comparison helps pinpoint where the change improves confidentiality and where it removes information that a service representative needs to write a useful reply.
Trace the route from inbox to workspace and back
Follow one complaint all the way to the outgoing reply. An email may include a signature, an earlier conversation and an attachment alongside the current request. Decide which of these the service agent needs. Start with plain text, then try a long email chain and a case with an attached order confirmation. Watch what the agent actually copies. If they habitually select the entire thread, you have identified a specific issue to address in the workspace setup and in staff training.
Inspect the outgoing email too. A well-written draft can still contain the wrong greeting, a deadline copied from a previous order or a placeholder left in the text. Have the sender check those three points against the original message. Ask them to show where they switch between windows while doing so. A check that takes several awkward steps is easy to skip under time pressure. That observation gives you a practical basis for deciding what to simplify before expanding the pilot.
Address shadow AI where the approved workflow stalls
Ask about the last task that made someone want to open another tool. Perhaps they needed a translation, struggled with a long attachment or could not access the approved workspace. Have them reproduce that step with a fictional example. Once you understand the result they needed, agree on an approved way to achieve it. Write down who can help with a blocked task and where a colleague can find that person's contact details during a shift.
Rehearse an urgent complaint near the end of the working day. A customer expects a reply, the approved workspace is unavailable and the message includes contract details. What can the agent write immediately from their own knowledge? Which question can wait until tomorrow? Repeat the exercise with someone covering the shift. If they struggle to find the current instruction or the right colleague, you have a specific improvement to make. You also have a training example that reflects a situation staff recognise.
Decide which cases to take beyond the pilot
Bring three items to the review: the fictional complaint, the draft from the pilot and the approved reply. Mark changes to the facts separately from changes to tone. Adding a friendlier closing is a different intervention from correcting a promise to ship a replacement. Include cases where the team deliberately chose to write the reply themselves. Those examples help define which requests belong in the workspace and which ones need a person to take responsibility from the outset.
Compare total handling time, including preparation and review. Drafting might save time on routine questions while a complex contract issue takes longer once the checks are included. Approve a clearly described group of cases and agree who will revisit that decision after the first few weeks. A new attachment type, a different recipient or a change in contract terms may alter the assessment. Keeping the examples with the decision gives the next person enough context to repeat the check when the real situation changes.
Finish by preparing a short exercise for a new member of the service team. Give them a fictional message and ask them to decide what to transfer, when to discard a draft and how to handle an urgent follow-up. Have them explain their reasoning aloud. Where their reasoning differs from the agreed procedure, the onboarding material may be missing a concrete example. Add that example next to the relevant instruction and try the exercise again with the next colleague joining the team.
Discussing nenna.ai and ChatGPT in one selection meeting
When the team mainly wants help writing proposals, start by studying the drafting process. If confidential customer details are essential to the draft, you also need a decision about how those details may be processed. Examine both questions separately using the same fictional brief. You can then see what writing assistance covers and where control over the input becomes the deciding factor.
Read the ChatGPT guide