A client of twenty years sent one line in a supplier questionnaire: "Do you use AI in delivering our work?" By Friday, three people in the same firm had given three different answers.
Everything in this article is an illustration, with hypothetical numbers. The firm is invented, the client is unnamed, and none of it is a client story. The method, though, is the one I use when a firm asks me how to answer this question honestly.
The question arrives on a Tuesday
The questionnaire comes in on a Tuesday, as one more item among sixty. The managing director forwards it to whoever usually handles supplier forms. That person ticks "No" in about five minutes, because nobody has ever told them otherwise and "no" feels like the safe word.
Over the next three days the client's contact asks around informally, and the firm's own people answer. The sales director says no, we do not use AI on their work. An operations lead says only for drafting internal notes. Someone in finance says yes, for reconciliation.
All three are sincere. Nobody lied, and nobody looked. Each person described the part of the firm they could see, and the form says something that none of them can confirm.
A wrong "no" is the expensive answer
In my experience, a large client is rarely alarmed to learn that a supplier uses AI carefully. What damages trust is a pair of answers from the same firm that do not match, or a statement that is later shown to be wrong. The first happens within a week, as above. The second happens when a questionnaire answer is repeated in a tender, a contract schedule and a sales deck, and then someone on the client side asks a follow-up that reaches a person who knows better.
So a false "no" is the expensive answer. It is repeated, and it cannot be defended once someone checks. An honest "yes, here is how" costs you a paragraph, and it tells the client that the firm knows what its own people do.
The risk lives in the gap between the form and the floor. Closing that gap is ordinary management work and takes about a week. Duties under contracts and data-protection rules differ by client and by country, so have a lawyer read your final sheet against your actual agreements. The method below makes your answer consistent and checkable. It does not satisfy any regulator on its own, and it does not guarantee how a client will respond.
Build the Client Answer Sheet in five steps
This is my working method, not research. Five steps, in this order:
- Collect the question as clients actually ask it. Pull every questionnaire, tender section, contract schedule and email from the last twelve months that touches AI. Note who answered and what they said. You will often find the answer already differs between documents.
- Run a one-week discovery, on client work only. Each team lists every place AI touched client material in the last 30 days: the task, the tool, whose account it was (the firm's or a personal one), and what client material went in. Say at the start that this is fact-finding, not discipline. If people expect blame, they under-report, and the sheet inherits the gap.
- Sort each use into State, Move or Stop. State means it is fine as it is. Move means a real need in the wrong place, such as a personal free account, so shift it to a firm-paid tool. Stop means client material that should not go into AI at all, or output that nobody checks. Fix Move and Stop before you write the answer, so the sheet describes today and not a hope.
- Write the sheet: five lines, plain words, one page. The five lines follow below.
- Name one keeper and date the sheet. One person looks after it, every reply uses it word for word, and it is re-checked on a fixed date and on three triggers: a new tool on client work, a new kind of client data, or a new question from a client.
The five lines of the sheet:
- Line 1. Where we use AI on your work: the tasks, not the brand names.
- Line 2. What of yours goes into it, and in which kind of tool, firm-paid or not.
- Line 3. Who checks the output before you see it, and what that check consists of.
- Line 4. What we do not use AI for on your work.
- Line 5. What we will tell you if this changes, and how soon.
Line 3 is where firms most often write something vague. I have argued separately what "a human reviewed it" has to mean in practice. Here it is enough to say that the line must describe a check someone really performs.
A worked example: a 45-person Beirut firm
Here is the method applied to an invented firm. Every figure below is a hypothetical number for illustration. The firm is a 45-person services business in Beirut with annual revenue of $3,000,000. The client who sent the questionnaire accounts for 12% of that revenue: 0.12 x $3,000,000 = $360,000 a year.
Before the exercise, three answers had circulated in a week, and the form had gone back with "No" ticked. Discovery then ran for one week across six teams, looking back over the previous 30 days, and found 14 uses of AI that touched client material. Five of them are shown below as samples, followed by the totals for all 14.
| Use | Tool type | Sort | Action |
|---|---|---|---|
| Drafting notes from client calls | Firm-paid, business terms | State | List on the sheet |
| Translating client documents | Firm-paid, business terms | State | List on the sheet |
| Summarising a client's long report | Personal free account | Move | Redo in a firm-paid tool, then close the habit |
| Reconciling client invoices | Personal free account | Move | Move to a firm-paid tool |
| Analysing a confidential client file | Free tool, no reviewer | Stop | Stop; any future use needs a named reviewer |
| All 14 uses | 5 State, 6 Move, 3 Stop | 5 + 6 + 3 = 14 |
After the fixes, the three Stop uses are gone, so the sheet lists 14 - 3 = 11 uses, and all 11 run in firm-paid tools. Before the fixes, 6 of the 14 uses sat in personal accounts: 6 / 14 = 0.4286, or about 43%. The form had said "No" to all 14.
State
fine as it is (5)
Move
real need, wrong tool (6)
Stop
should not be in AI (3)
On the sheet (11)
The sheet for this firm then reads, one sentence per line:
- We use AI to draft meeting notes, summarise and translate documents, and reconcile invoices on your work.
- Your documents go only into firm-paid tools under business terms, never into personal accounts.
- A named colleague reads and corrects every output before it reaches you.
- We do not put your confidential files into AI without a named reviewer.
- We will tell you within 30 days of any change to the above.
The 30-day promise in line 5 is an assumed commitment for this example, not a standard.
The cost of the exercise: two people for one day each is 2 x 8 = 16 hours. At an assumed internal cost of $25 an hour, that is 16 x 25 = $400. Set that beside the $360,000 account whose questionnaire started it. I am not predicting that the account was at risk. The comparison only shows the scale of what depends on the answer.
The 12% share, the $3,000,000, the $25 an hour, and the 14, 6, 5 and 3 split are assumptions, not research. Your own discovery will produce its own numbers.
Keeping it true
A sheet is accurate on the day it is dated and starts to age the next morning. That is why step 5 asks for a keeper and three triggers. The keeper is a named person whose name is on the page and who is asked first whenever something changes. The triggers are a new tool on client work, a new kind of client data, and a client's new question.
I have written before that a tool approved in March is not always the tool running in July. The same applies here. A tool whose terms or behaviour have changed counts as a new tool on client work, and it triggers a re-check.
Monday action: find the last client questionnaire that asked about AI. Read your answer aloud to two colleagues from different teams and ask each what they actually do. If what they say differs from the form, you have the first finding for step 2. If you want a second pair of eyes on your answer sheet, write to me.
Related evidence: Wikipedia describes shadow AI as a subset of shadow IT: employees using AI applications, including large language models and AI productivity tools, without the knowledge or approval of the organisation. (Wikipedia's entry on shadow IT and shadow AI)
Microsoft and LinkedIn's 2024 Work Trend Index reports that 78% of AI users bring their own AI tools to work, 80% at small and medium-sized companies, often without guidance from leadership. (Microsoft and LinkedIn's 2024 Work Trend Index on employees bringing their own AI to work)
