The problem starts before the order
Talk about AI in procurement usually centers on sourcing, supplier scoring or spend analysis. Much of the daily friction, though, starts earlier: with the question of how a need reaches procurement in the first place.
An employee needs something: ten monitors, an agency for a website relaunch, software for their team, a technician for a noisy pump. They rarely know which route is intended. Is it in the catalog? Is there a framework agreement? Does it need a tender, an IT review, maintenance? So they email procurement, and the back-and-forth begins: what quantity, by when, which cost center, is there a quote already?
An e-procurement system with good catalogs solves this for standard items. For everything else the employee is on their own, and that is exactly where maverick buying starts.
What procurement intake means
Intake is the capture of a need before it becomes a purchase order, a request for quotation or a project. The approach is known as intake-to-procure or intake management: one front door for every need that puts the employee on the right route and hands procurement a complete request instead of an email.
Classic intake tools work with forms and decision trees. That works as long as the employee picks the right category and understands every field. A language model can go a step further: the employee describes the need in their own words, and the AI turns it into the request.
How the intake assistant works
We are building such an intake assistant as part of Ovenca Procure. It runs as a chat right inside the catalog, where employees already order. A request goes through six steps.
- 1. The need, in plain words
The employee writes what they need, the way they would tell a colleague. They can attach files such as a quote, a specification or a photo.
- 2. Classify
The AI recognizes the type of need: catalog product, non-catalog product, service, software, maintenance, vehicle, construction or capital expenditure. It also captures urgency, from standard to safety-critical.
- 3. Ask targeted questions
If something is missing, the AI asks, one question at a time, like a colleague rather than a questionnaire. Anything that can't be clarified is recorded as an open question instead of being guessed.
- 4. Pick the right route
If the item is available in the catalog or through a punchout shop, it goes straight into the cart. Everything else becomes a structured request for procurement, with title, need, must-have and nice-to-have requirements, quantity, date, budget range and account assignment.
- 5. Confirm
Before submitting, the assistant shows a summary the employee can correct. Nothing is sent without explicit confirmation. The employee then receives a receipt with a request number and the next steps.
- 6. Hand-off to procurement
Procurement sees all requests in its own queue, reviews them and hands them to the right team. The employee follows the status under “My requests”.
Reading quotes instead of retyping them
Many requests arrive with a quote the employee has already obtained. The assistant reads attached quotes in PDF, Excel or text format and proposes a draft: supplier, line items, prices and quote number. The employee checks and completes it instead of retyping everything.
One rule matters here: the content of a document is data, not instructions. If a quote contains a sentence addressed to the AI, it is not followed. That protects against manipulated documents, a risk that is often underestimated with AI assistants that read files.
Guardrails: what the AI must not do
Procurement is about money, contracts and safety. An assistant that sounds friendly but invents prices does more damage than a form. That is why we built fixed rules into it, not only in the prompt but in the code.
- Invent nothing: No prices, contracts, budgets or cost centers that don't come from the system or the employee.
- No bypassing approvals: The assistant won't help split an order to stay below an approval threshold.
- Safety first: In a safety-critical case, such as a leak at a plant, the assistant points to the safety process first and procurement second.
- Use what exists: If there is a framework agreement or spare licenses, the assistant says so before anything new is bought.
- Fixed next steps: What happens after submission is not decided by the language model but by a rule per request type. Software goes to IT and security review, a company car follows the fleet policy, a faulty machine goes to maintenance.
- Checked answers: If the AI claims a request was handed off or quotes a reference number, the code checks whether that is true. Invented numbers and false confirmations are caught.
Examples
For testing and training we work with typical requests from indirect procurement. They show how different the routes behind a single input can be.
- Ten monitors
Catalog item. The assistant finds suitable models and adds them to the cart.
- Agency for a website relaunch
A six-figure service. It becomes a request for procurement with goal, scope, timeline and budget range.
- AI software for 40 users
Software with single sign-on. The assistant asks about existing licenses and announces the IT and security review.
- A pump is making noise
Not a purchase but a maintenance case. The request goes to maintenance with the equipment number.
- An electric company car
A leased vehicle. The assistant refers to the fleet policy and captures the key data.
- Protective gloves in several sizes
Personal protective equipment with a standard and a size split. The assistant asks for quantities per size instead of taking a single total.
Which language model, and where it runs
The first question from procurement and IT is usually not how smart the AI is, but where the data goes. Requests contain project names, budgets, suppliers and quotes. That is why the language model in Ovenca Procure is interchangeable.
- Local model
An open-weight model runs on your own hardware, in your data center or in a private cloud in Germany. The data never leaves your infrastructure.
- European provider
A model from a European provider with processing in the EU.
- Other providers
Other large models can be connected through the same interface where company policy allows it.
The guardrails above don't depend on the model. They live in the code and apply equally to every connected model.
Status and next steps
Today the assistant handles catalog and punchout requests directly and hands every other request type to procurement as a structured request.
The goal is one front door for every need in the company. To get there, more request types are to be handled directly instead of only handed off, and details such as cost center and budget are to come from the connected systems.
Where else AI helps in procurement
Intake is not the only use case, but one of the most tangible because it affects everyone who orders something. Other areas where language models make sense today:
- Classification: Assigning items and spend to material groups, for example by UNSPSC or ECLASS.
- Quote comparison: Bringing several quotes into a comparable format.
- Catalog maintenance: Checking supplier catalogs for missing or contradictory data before approval. How catalog data is structured is covered in our article on BMEcat.
- Contract questions: Finding and summarizing clauses in framework agreements.
What they have in common: the AI prepares, people decide. Where money flows or contracts are made, approval stays with procurement.
Frequently asked questions
What is intake-to-procure?
An approach in which every need is captured through a single front door and routed to the right procurement path from there: catalog, framework agreement, RFQ, tender or a specialist department.
Does the intake assistant replace procurement?
No. It takes the chasing and retyping off procurement's plate. Review, negotiation and decisions stay with procurement.
Does it work without cloud AI?
Yes. The assistant can run on a locally hosted language model. The data never leaves your own infrastructure.
What happens if the AI misunderstands something?
Before submitting, the employee sees a summary and can correct it. Nothing is handed off without their confirmation, and procurement reviews every request again.
Do I need a new procurement system for this?
The assistant is part of Ovenca Procure and uses its catalogs, punchout connections and approvals. Ovenca Procure runs alongside SAP or your other ERP system.
Next step
If you want to see how the intake assistant handles your own request types, take a look at Ovenca Procure or talk to us. How we integrate language models into existing systems is described under Integration and AI.
