Skip to content
Exclusive partner BME Symposium18–19 November · Berlin · Booth A16Book a meeting
Contact

Efficiency hits its limits. Decisions don't.

Procurement AI so far has made tasks faster. The larger value lies in better decisions across the whole spend, and those depend on method, memory and mandate, not on the model.

In short

AI in strategic procurement prepares decisions: it combines spend, contract and market data, applies procurement method and puts reasoned options on the table. The decision stays with the procurement team. Its value lies less in hours saved than in better decisions across the entire spend. That takes more than a model: codified method, a company memory and a clear mandate.

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.

Automation can save at most what the process costs.

E-procurement, catalogues, guided buying and invoice matching made transactions cheaper. That work was right and remains the foundation. But it has a built-in limit: automation can never save more than the automated process costs. In most companies the entire procurement function costs well under one percent of the spend it is responsible for.

Decisions act on something else: the spend itself. Which supplier, which index formula, which notice period, which bundling across plants. In our model, better decisions across the whole spend are worth 5 to 10% of volume: €25M to €50M a year on €500M. That is many times what automation realistically saves. McKinsey puts the difference at roughly elevenfold: better sourcing decisions deliver around eleven times more value than automating the procurement function. Measuring AI by hours saved means optimising the smaller of the two numbers.

This explains why so many pilots stall. They speed up summarising a contract or drafting an email. A faster task rarely changes a decision, and without measurable business value there is no case for scaling. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027.

Figure 1ModelEfficiency ends at €3.5M. Better decisions move €25M to €50M.

Limit of efficiency

€2.5M
€3.5M
€25M
€50M
Efficiency: process 70% automatedEfficiency: procurement process eliminatedDecision: 5% better outcome on the spendDecision: 10% better outcome on the spend

Model. Assumptions: €500M addressable external spend; full procurement process cost of 0.7% of spend (€3.5M), realistically up to 70% automatable. Decision lever 5–10% of spend per the H&Z and Elevatry lever model (twelve levers, net of overlaps). Values per year, before implementation cost.

Where AI sounds most convincing, its mistakes cost the most.

AI pushes the cost of analysis towards zero. Market scans, bid comparisons and contract summaries take minutes. That moves the bottleneck: what is scarce is no longer the analysis but knowing which question matters and which lever works in this category.

A field experiment with 758 consultants shows how unevenly AI helps. On tasks within its capabilities it delivered markedly higher quality. Just outside them, the rate of correct answers fell by 19 percentage points, with fluent answers. In procurement, an index clause can be summarised flawlessly while the conclusion about negotiating leverage is still wrong.

What separates the good answer from the plausible one is method: the levers, assessment criteria, cost logic and negotiation frameworks of experienced category managers. At Elevatry this method, drawn from H&Z's project work across more than 5,000 procurement projects, is built into the Companions. They apply it step by step and show their reasoning.

−19 ppfewer correct answers when AI was used on a task just outside its capabilities.DataDell'Acqua et al., Harvard Business School working paper 2023, field experiment with 758 BCG consultants
A model without method scales the average, only faster.

Rent the model. Own method, memory and mandate.

Capable language models are available to every company, often from several providers. An organisation that equates its AI strategy with its choice of model competes on what any competitor can rent tomorrow. The advantage builds up in the three layers above it: method, memory and mandate. If any one is missing, the impact does not materialise, however good the model.

Memory means that contracts, prices, negotiation history and every decision with its reasons stay in the company, separate from the model. At Elevatry that is the Company Brain, connected through a provider-agnostic Model Gateway, with an EU data zone by default and no training of shared models. Today procurement knowledge often has the half-life of a job change. With memory, each decision starts where the last one ended.

Mandate means it is defined what a machine may decide, on which evidence and under whose accountability. Where errors are cheap and reversible, it acts after approval. Where much is at stake, it prepares and people decide. Without that mandate, employees turn to their own tools, and nobody controls what happens there.

71%of AI users in German companies use their own, non-approved AI tools.DataMicrosoft Work Trend Index 2024

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.

All Companions
Model

A framework agreement that would have renewed itself

A manufacturer buys maintenance services worth €8M a year for three plants from one provider. The framework agreement renews automatically for 24 months, with six months' notice. It ends in eight months, and it is on nobody's list.

  1. 01
    Work the deadline backwards

    The Companion reads the contract and works the notice period backwards. The decision window closes in eight weeks, not eight months.

  2. 02
    Put the spend side by side

    The three plants pay different hourly rates for the same services. Two plants with €5M of combined volume pay on average 6% more than the cheapest.

  3. 03
    Ask the memory

    The Company Brain holds the notes from the last negotiation. The provider had agreed a discount tier above a volume that has been exceeded for two years. It was never applied.

  4. 04
    Put a price on the options

    Recommendation: give notice on time to keep the option open, then renegotiate or tender with bundled volume. Harmonised rates are worth about €0.3M, the discount tier another €0.16M a year.

  5. 05
    The category manager decides

    Whether to negotiate or tender depends on service quality and the relationship with the provider. That trade-off is a human call.

About €0.46M a year of room to act, and a deadline that would otherwise have passed unnoticed. No process step got faster; one decision got better.

Illustrative, not a customer result.

Questions worth asking.

  1. 01

    What share of our spend received a decision of expert quality last year, and how do we know?

  2. 02

    Do we measure our AI initiatives in hours saved or in better decisions, in euros?

  3. 03

    Where do our decisions and their reasons live, and do they belong to us or to a vendor?

  4. 04

    What may a machine decide in our organisation today, and who decided that?

What procurement leaders ask us.

How do I know whether AI in procurement pays off?
By whether decisions improve in euros. Hours saved are the wrong measure. Automation can never save more than the process costs; in our model with €500M of spend, that ceiling is €3.5M a year, even if the procurement process disappeared entirely. Better decisions on supplier choice, index formula and bundling are worth 5 to 10% of spend in the same model, €25M to €50M. Measure realised value against a baseline, and the share of spend that has seen a full decision.
Is a general-purpose AI assistant enough for strategic procurement?
No, because it lacks method, memory and mandate. It can summarise an index clause flawlessly and still know nothing of your negotiation history or the rebate tier that was promised and never applied. It becomes dangerous just outside its capabilities: there, consultants using AI were 19 percentage points less often correct in a field experiment. A model without method scales the average. With codified method, every category gets the same quality of reasoning, not only the large ones.
How is an AI platform for strategic procurement different from purchase-to-pay software?
Purchase-to-pay software makes transactions cheaper; an AI platform for strategic procurement makes decisions better. One acts on process cost, in our model at most €3.5M a year on €500M of spend. The other acts on the spend itself: which supplier, which index formula, which notice period, which bundling across plants. The two complement each other, in that order. Elevatry brings the method from more than 5,000 H&Z projects to every category, not only the large ones.
Will AI mean we need fewer category managers?
No. The problem is not too many people but too little coverage. In H&Z and Elevatry project work, only about one supplier in five is actively managed, and 50–80% of expert time goes into knowledge-intensive, complex preparation work. The rest is renewed and re-ordered. When Companions take on this work, the same expertise reaches categories that were only administered so far. The team stays the same size; the share of spend actively shaped grows. Judgment, negotiation and award stay with people.
Why do so many procurement AI pilots stall?
Because they speed up tasks without changing a decision. A faster contract summary saves minutes and moves no euros; without business value there is no reason to scale. According to EFESO, 75% of procurement organisations are still in the AI learning phase. The way out is to start with a decision quantified in euros, such as a framework agreement that would otherwise renew silently. What then scales is judgment across the full spend, not text production.

Friday notes on the topic

Next step

See it on your own procurement.

We show you on one of your processes how Companions do the work from analysis to contract. With your data and your approval rules, if you like.

See the platform