In short
AI governance in procurement defines which tasks AI may take on, which data it works with and where people approve. With agentic AI this belongs in the architecture: autonomy per process step, set by reversibility and materiality, approval points, data sovereignty and traceable decisions. Without an approved path, employees use their own tools, and control is lost.
Data: third-party research with source. Our view: the joint position of H&Z and Elevatry, indicative observations from project work. Model: calculations and examples, not customer results.
01Argument
Autonomy is not one policy for everything. It is a dial per process step.
Two criteria do most of the work: how easily an error can be undone, and how much is at stake. A misclassified order is corrected with one click. An award binds the company for years. Both errors cost something, but not the same thing.
That gives a dial with three levels. In autonomous mode a Companion completes an entire process where errors are cheap and reversible; people approve the result or check samples. In guided mode the Companion prepares and the team confirms every step; that applies to evaluations, scenarios and anything material and hard to undo. In orchestrated mode, currently in development, several Companions work together. Awards and terminations are decided by a person at every level.
Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027. A common reason is that nobody defined what the machine may decide. The pilot then stays a pilot, or scales without control. Both are expensive.
View of Elevatry and H&Z. Typical starting levels for a first deployment; each organisation calibrates its own thresholds.
02Argument
Autonomy is earned with evidence, not granted by decree.
Measure Companions the way you measure suppliers. For each process step, track acceptance, override and error rates. A step that runs a quarter in guided mode with few overrides switches to autonomous mode. If overrides rise, it goes back to guided mode.
The second risk is quieter: polished output approved without challenge under time pressure. The design response is evidence behind every claim, a named person responsible for challenging, and sampled reviews even at higher levels. Whoever only approves what they could no longer produce themselves is no longer reviewing. They are signing.
Elevatry records every recommendation, approval and correction with source, person and time in the Event Trail. That is the basis for justifying autonomy, and it allows EU AI Act classification per use case rather than one verdict for the whole platform.
A step that ran a quarter with almost no overrides has earned autonomous mode.
03Argument
Rent the model. Own the mandate.
According to Microsoft, 71% of AI users in German companies use their own, non-approved tools. Contracts and supplier prices then end up in private accounts with no data processing agreement. Bans only push that use into channels nobody sees. What helps is an approved path that works better day to day.
Sourcing that path calls for procurement logic. No procurement team would single-source a critical component without a contingency. Many AI strategies do exactly that with their intelligence. Models are converging and available to every competitor. Equating strategy with model choice means competing on what everyone can buy.
What you should own is method, memory and mandate. The Model Gateway selects the model per task. The Company Brain stays separate from the model in the EU data zone, with no training of shared models on your data. If the provider changes, the context stays.
Companions for this topic
What Elevatry does for you.
Companions work around the clock in your data, with the H&Z method from more than 5,000 procurement projects. They prepare decisions, work in your experts' feedback and, once you approve, carry out the next steps. Your team decides.
Worked example
ModelIndex clauses: which level for which step?
A company has 240 contracts with index clauses covering €40M of spend. Today they are checked by hand twice a year. Index reductions are claimed on average five months after they arise, if at all.
- 01Break the process down
Matching index movements against contracts is cheap and reversible. A letter to a supplier leaves the building and cannot be recalled. An escalation or termination is material.
- 02Start cautiously
In the first quarter every step runs in guided mode. Buyers review each calculation and each letter.
- 03Collect the evidence
Of 410 calculations, 6 are overridden, 1.5%. All trace back to contract amendments missing from the system. The data source is fixed.
- 04Turn the dial
Calculation switches to autonomous mode with a 10% sample. Letters run in autonomous mode with approval. Escalations are still decided by people.
- 05The person decides
The head of procurement sets and owns the levels and thresholds. Every change is justified in the Event Trail.
Index reductions are claimed within one month instead of five. With a 2% index fall on €40M, that is about €0.3M a year in the model. No letter leaves the building without approval.
Illustrative, not a customer result. Assumptions: €40M spend under index clauses, 2% index fall, four months gained (2% × €40M × 4/12 ≈ €0.27M).
For your next leadership meeting
Questions worth asking.
- 01
Which procurement decisions may a machine take today, and where is that written down?
- 02
How much of our AI use runs through private accounts, and how do we know?
- 03
Could we switch our AI model tomorrow without losing context and decision history?
- 04
Who in our organisation notices when a poor AI recommendation is approved anyway?
FAQ
What procurement leaders ask us.
Is AI in procurement compatible with GDPR and the EU AI Act?
Who is liable if an AI recommendation in procurement is wrong?
Can buyers put contracts and supplier prices into public AI tools?
How much can AI decide in procurement without approval?
What happens to our procurement data if we switch AI models?
Read on