Picture an office manager at a contractor, turning rough job notes into a customer email. The engineer has moved on to the next site. The contracts manager wants an update. Rather than wrestle with the wording, she opens an AI assistant on her own account.
That is the sort of everyday workaround I mean by shadow AI: AI used for business work outside the organisation's agreed tools or ways of working. This is an illustrative situation; the firm is made up.
I'd hold off on the warning email and start with a question: what made that workaround useful enough to adopt?
Shadow AI is a workflow question before it is a tool question
There is evidence behind the concern, but be careful about applying it to your own business. Microsoft and LinkedIn's 2024 Work Trend Index reported employees bringing their own AI tools to work and reluctance among some users to admit their use. This is vendor-backed research about knowledge workers globally, so it tells you little about what your own engineers or office team are doing.
Find out rather than assume.
I'd look at the jobs people are trying to finish: rewriting emails, organising meeting notes, drafting tender answers or making sense of a supplier's document. Ask where the work starts, what goes into the assistant and where the result ends up.
The distinction matters. An experiment with made-up text is not the same operating decision as making an AI-written customer update part of your normal reporting process. Treating both as simply “using AI” misses the work that needs managing.
My view is that a useful response has to do more than name an approved product. It needs to make the task, the checking and the ownership clear.
Ask for a walkthrough
Start with a straightforward conversation with the people doing the admin. I would say: “Show me a task where you've used AI, or wanted to. What were you trying to finish, and what happened next?”
Make clear that you are establishing how work gets done. Nobody should be demonstrating experiments with real customer material. Use a made-up example for the walkthrough. Keep any existing restrictions in place while you investigate.
Record the essentials in a simple working list:
- The task and the person responsible for it.
- The tool and whether the account is personal or company-managed.
- The type of information the task involves, without copying that information into the list.
- What the assistant produces and who checks it.
- Where the checked result is stored or sent.
- What happens if the assistant is unavailable or the answer is unusable.
This is an operational discovery exercise. Nobody should be collecting passwords or inspecting personal accounts. Ask people to describe the work and demonstrate only what they are authorised to show.
Then ask the less comfortable question: what would make the agreed way easier than this workaround? A list of tools alone will not answer that.
Separate useful drafting from decisions that need judgement
I would distinguish between helping someone express an answer and deciding what the answer should be.
In the illustrative contractor, an assistant might be tested on rewriting a made-up progress update into clearer language. I would not treat that test as evidence that it can decide whether work is complete, approve a variation or establish the cause of a defect. Those are different tasks, with different consequences.
There is a technical reason to resist polished wording as a quality test. NIST's Generative AI Profile describes confabulation: generated content that is wrong or false but presented confidently. In plain English, an answer can sound settled without being correct.
My recommendation is to keep operational decisions with a person who understands the underlying job. Give the draft a named reviewer and define what that reviewer checks. “Read it over” is too vague.
For a customer update, my checklist would include the job reference, work actually completed, outstanding items and any promised next action. Compare each against the original record. If the source notes do not establish a completion date, the draft should not quietly supply one.
The aim is a usable message that still says what actually happened, even if the prose is plainer.
Agree a working route people can follow
After discovery, I would write a short operating note for the specific task you want to test. Keep it specific to that task. A generic declaration that AI is encouraged helps nobody, and no tool is suitable for every kind of business information.
Name the task, the approved route, the person responsible and the review step. State what is outside the trial and where people should take a new request. Have the relevant specialists confirm information-handling and contractual questions before expanding into real business material; this article covers organising the work and leaves those questions to them.
There is a useful example of explicit ownership in the Home Office's engineering standard for AI use. It requires appropriately experienced human review of AI-assisted outputs before production and keeps accountability with human teams. That is the Home Office's own engineering standard, and it doesn't apply to your contractor business.
The principle I would borrow is simple: assistance does not remove ownership.
For your office team, spell out who can accept the finished update and where it belongs. If the responsible person is absent, identify a substitute or use the existing manual route. Do not let the trial depend on the only person who knows which prompt to paste.
And give people somewhere to ask for a better method. I'd rather receive a request for a new use than leave the team guessing whether the old agreement covers it.
Fix the friction that made the workaround attractive
Return to the illustrative office manager. Why does a routine customer email require so much interpretation?
Perhaps the engineer's notes mix completed work with proposed work. Perhaps the office is searching WhatsApp for a missing detail. Perhaps the job sheet does not ask for the information the customer expects.
I would test changes there before buying another subscription. Give the engineer a clearer job-note structure. Separate completed work, unresolved issues and next actions. Make the office's source record obvious. Try an agreed email template against the same task.
These are proposed process improvements, with no promised results. Judge them against what happens in your own workflow. If a template removes the writing problem, keep the template. If AI still earns a place after that, test it on the remaining task rather than making it compensate for every upstream omission.
That is also the starting point in our AI readiness assessment guide: understand the work before choosing the technology. Mise's Assess stage is about identifying that friction and deciding what deserves attention.
A workaround is worth examining. Whether it deserves to become a permanent system is a separate decision.
Test the whole task, then keep what works
This week, choose a recurring admin task and document the current route. Use fictional material to compare the existing method, a clearer template and an AI-assisted draft. Arrange any further approvals before trying real material.
For each attempt, record preparation, drafting, checking and correction effort. Look at the final result as well as how quickly the first draft appears. Ask the next person in the workflow whether it gives them what they need.
Before continuing, agree what would make you stop: invented details, unclear ownership or more correction work than the original task. Keep a workable manual route and set a review point rather than allowing an experiment to become permanent by default.
My position is simple. Do not reward hidden use, but do not mistake discovery for the solution either. Bring the work into view, give it an owner and remove the friction you can. AI should earn its place in the process. Otherwise it becomes another workaround everyone depends on.
If you want a clearer view of the admin behind your team's workarounds, we can talk it through.