Voice AI in Procurement
Where it helps, and where it does not
Voice is being added to procurement software fairly indiscriminately. It solves one problem extremely well and several others badly, and telling the difference before you buy is worth doing.
Voice AI is most useful in procurement where the person providing information is not at a desk or is working in a second language, which makes supplier onboarding its clearest application. It is least useful for tasks requiring precision, review or comparison, such as approving high value purchases or reading a contract, where reading and clicking are faster and safer than speaking.
Three situations
In each of these, speaking removes a real barrier rather than adding a novelty.
| Situation | Why voice helps |
|---|---|
| The person is not at a desk | Field contractors, site crews and small business owners complete things by speaking that they would postpone indefinitely if a laptop were required |
| The person is working in a second language | Speaking is substantially easier than writing in a language you read imperfectly, and compliance questionnaires are the hardest kind of writing |
| The task is providing information, not reviewing it | Answering questions works well spoken. Comparing options or checking a document does not |
Four cases worth resisting
Approving anything material
Approval needs the approver to see what they are approving. Spoken summaries of a purchase encourage approval without review, which is the opposite of the control.
Reading or comparing documents
Documents are scanned, not read linearly. Having a contract read aloud is slower and worse than reading it, and comparison is essentially impossible.
Anything requiring precision
Account numbers, tax identifiers and quantities are error prone spoken and need confirming visually anyway, which removes the saving.
Tasks somebody does at a desk all day
If the user is already sitting at a screen, typing is faster. Voice added here is a feature demonstration rather than a benefit.
Five questions
These separate voice that does a job from voice that exists to be mentioned in a product announcement.
| Question | What a good answer looks like |
|---|---|
| What comes out at the end? | Structured data mapped to specific fields, not a transcript somebody has to read and re-enter |
| What happens to required fields? | They stay required. Voice should not be a route around validation that typing enforces |
| Can the user switch to typing? | Yes, at any point, with progress carried across. Forcing one mode suits some users and excludes others |
| Which languages? | The languages your suppliers actually use, not a demonstration language plus a roadmap |
| What happens to the recording? | A clear answer about retention and purpose. Vague answers here are worth pressing on |
It fits every criterion
| Criterion | Supplier onboarding |
|---|---|
| Is the person away from a desk? | Frequently. Field contractors and small suppliers often have no back office at all |
| Is a second language involved? | Very often, and the questionnaire is exactly the hardest kind of document to complete in one |
| Is the task providing information? | Entirely. Onboarding is answering questions, which is what speech is good at |
| Does the output need structure? | Yes, and that is achievable because the questions are known in advance |
| What is the cost of not doing it? | A supplier who abandons onboarding silently, which you never learn about |
