Prototype a purchasing workflow
How can you connect generative AI to purchasing tracking without allowing an unverified proposal to modify management data?
3500 € per session, excl. VAT
In short
Two days, comprising 14 hours excluding breaks, to design a prototype linking generative AI to a purchasing process. Participants structure data exchanges, build a validation workflow, test errors and prepare a measurable pilot. The work remains in a test environment, with no industrialisation or writes to production.
Why this training?
Supplier confirmations often circulate separately from recorded purchase orders. Re-entering them can conceal a discrepancy in an item reference, quantity or delivery lead time; automating without checks risks propagating these discrepancies into procurement. Generative AI can propose structured extraction, but it replaces neither deterministic checks nor authorisation by the relevant business function. The training brings together source data, verified interfaces, human validation and traceability of proposed changes. Continuing case study: The fictitious organisation « Ondrelis Robinetterie » receives supplier confirmations with delivery lead times that sometimes differ from those on purchase orders. Using fictitious documents and an export from a management system, the team prototypes a proposed update that must be approved by a buyer. Transfer to the workplace: Present the dossier to the purchasing manager and the IT contact to authorise a limited pilot: the purchase orders concerned, those responsible, quality thresholds, monitoring and stopping conditions must still be validated before any live connection.
Objectives of this training
- Map the flow between supplier confirmations and purchase orders
- Define a data contract and the necessary permissions
- Build a prototype with human validation
- Measure errors using an annotated test set
- Draft a pilot plan with stopping conditions
Who is it for and what are the prerequisites?
Roles and tasks concerned
- Purchasing and procurement managers leading an automation project
- Management tool specialists, process analysts and IT project managers
Prior skills
- Have practical experience of generative AI and structure extraction instructions
- Know the order–confirmation–receipt cycle and its controls
- Be able to work with exports and understand the principles of an API
What you will work on
Define the process boundaries
Objective: Select the scope that can be automated. Events, stakeholders, decisions and exclusions.
Prepare the data
Objective: Define the expected data exchanges. Identifiers, types, units, dates and data minimisation.
Choose an architecture
Objective: Justify the integration method. Files or a documented API; minimum permissions and secrets.
Structure the extraction
Objective: Produce output that can be checked. Instructions, expected schema, missing values and evidence.
Assemble the prototype
Objective: Run an initial workflow. Reading, extraction, checking and validation queue.
Safeguard validation
Objective: Establish controls for business decisions. Approval, rejection, separation of roles and logging.
Build the tests
Objective: Establish a reference for evaluation. Standard cases, ambiguous documents and embedded malicious instructions.
Measure, then correct
Objective: Quantify the quality achieved. Accuracy by field, omissions and incorrect matches.
Test incidents
Objective: Prevent inconsistent processing. Duplicates, unavailability, recovery and the absence of duplicate processing.
Prepare the pilot decision
Objective: Define the conditions for experimentation. Scope, acceptance thresholds, supervision and shutdown.
What you take away
- Flow map and data contract
- Extraction and validation prototype with a log
- Annotated test set and quality report
- Pilot dossier and recovery procedure
Method and assessment
- Short instructional sessions followed by immediate practice on the running case study.
- Individual work followed by work in pairs, with peer review of the outputs.
- Guided group discussion: what works, what needs correcting and what remains a human decision.
- Preparation: Describe a purchasing workflow and its approval rules, without confidential data. Check the authorised tools, test access and available documentation; plan for file-based exchanges if necessary.
Observable criteria: Each proposal retains its reference to the source document; Mandatory fields and invalid references are checked; No proposed change bypasses human approval; Extraction errors are counted by field and compared with the annotations; The pilot specifies responsibilities, thresholds, and conditions for stopping and resuming.
Frequently asked questions
Is an API essential?
No. Files can be used to prototype the exchanges; a connection depends on the interfaces actually available.
Can we use our purchase orders?
Only in anonymised form and within authorised tools; the fictitious case study is sufficient for the exercises.
Does the AI modify purchase orders?
No. It proposes values; the prototype enforces checks and human approval.
Will the solution be operational in production?
No. The two days produce a tested prototype and a pilot plan that still requires authorisation.
Last updated: