Procurement teams are being asked to do more with less. The Hackett Group expects procurement workloads to grow 8% in 2026 while headcount and budgets shrink, and AI is the obvious answer. The adoption numbers look impressive: 71% of procurement organisations use generative AI and 56% are experimenting with AI agents. But look closer and the picture changes. Only 12% run generative AI at large scale, and EFESO's 2026 CPO Pulse Report finds that just 5% have successfully industrialised it. This guide shows where AI really works in procurement, what savings are realistic, which agents SAP, Coupa, Anthropic, OpenAI and Mistral offer, and how companies in Croatia, Germany and Austria can move from pilot to production.
Adoption: wide, but shallow
The Hackett Group's 2026 Procurement Agenda and Key Issues study is the best single picture of where procurement stands. Generative AI and AI agents have spread quickly, but almost all of it is still at the pilot stage. 59% of organisations are piloting generative AI and only 12% use it at large scale; for agentic AI the split is 43% to 13%. Contract management is the leading use case: 49% are actively piloting or implementing AI there.
Source: The Hackett Group, 2026 Procurement Agenda and Key Issues study (press summaries, March 2026). Shares of procurement organisations at pilot or large-scale implementation.
Other surveys tell the same story with different numbers. In the 2026 HFS Research and Art of Procurement survey, 15% say they have reimagined procurement with AI at the core, while 38% only run pilots. GEP counts 17% of organisations with agentic AI deployed and 44% piloting it. EFESO is the most sceptical: 5% have industrialised generative AI, and only 34% are satisfied with the value it has delivered relative to scope and investment.
Sources: EFESO CPO Pulse Report 2026; The Hackett Group 2026 Key Issues study; HFS Research and Art of Procurement, State of AI in Procurement 2026; GEP 2026 Procurement Executive Insight Report. Surveys use different samples and definitions, so the bars are not directly comparable. The grey bar shows Hackett's total adoption including pilots for contrast.
Where AI works in procurement today
Procurement is full of text and repetitive decisions: requests, specifications, quotes, contracts, invoices and supplier emails. That makes it a natural fit for language models. The value differs a lot by task, and so does the risk.
| Task | What AI does | Typical tools in 2026 | Maturity |
|---|---|---|---|
| Intake and guided buying | Turns a free-text request into a correct purchase requisition, picks the catalogue item or the right buying channel | Coupa, SAP Ariba with Joule, Zip, custom assistants on Claude or ChatGPT | Production-ready |
| Contract review and management | Extracts terms, compares clauses with your playbook, flags risk, tracks renewals and obligations | SAP Ariba contract features, Icertis, Ironclad, Claude with long context | Production-ready, human review needed |
| Spend analysis | Classifies spend, finds maverick buying, duplicate suppliers and price variance | Suite analytics, text-to-SQL agents on ERP data | Production-ready if data is clean |
| Sourcing and bid analysis | Drafts RFQs, compares bids, summarises supplier answers | SAP Joule Bid Analysis Agent, Coupa, Keelvar | Early production |
| Tail-spend negotiation | Negotiates small, low-risk purchases with suppliers within set limits | Pactum and similar autonomous negotiation tools | Pilots, selected large buyers |
| Supplier risk and research | Monitors news, financials, sanctions and ESG signals; researches new suppliers | Risk platforms, deep research agents, Manus | Production-ready for research |
Classification by AI Workshop based on vendor documentation and the surveys cited in this article. Tool lists are examples, not recommendations.
What savings are realistic?
Vendor brochures promise 20 to 30% cost reductions. The consultancies that measure it are more careful. McKinsey estimates that agentic AI in strategic sourcing, deployed well, adds 3 to 6% savings on external spend. Bain reports 3 to 7% incremental savings from effective AI deployment, with productivity gains of 60% or more. McKinsey's broader estimate for generative AI at scale is 5 to 15%, mostly from less value leakage and better contract compliance, not from tougher negotiation.
Sources: McKinsey, "Agentic AI in strategic sourcing: Hype or hero?" (3–6% on external spend); Bain via Procurement Magazine (3–7% incremental savings); McKinsey estimate for generative AI at scale (5–15%, mainly value leakage and compliance). Ranges, not guarantees.
A few percent sounds modest until you apply it to real spend. For a company with €50 million of addressable external spend, 3 to 6% is €1.5 to 3 million a year, usually far more than an AI programme costs. The chart below shows the same arithmetic for three company sizes.
Illustrative calculation by AI Workshop using McKinsey's 3–6% range on addressable external spend. Actual results depend on spend mix, data quality and how much of the spend AI touches.
The tools: SAP, Coupa and build-your-own with Claude, GPT and Mistral
The big source-to-pay suites have turned AI into agents. SAP relaunched SAP Ariba in March 2026 as an AI-native platform with Joule built into procurement workflows, and plans 11 new Joule agents across Ariba and Fieldglass during 2026, including a bid analysis agent. Coupa launched Coupa Compose at Inspire in May 2026, a platform for building, deploying and orchestrating agents without migration or new code, alongside outcome-based pricing options. If your company already runs one of these suites, start there: the agents see your catalogues, contracts and approval flows.
Many mid-sized companies in Croatia and DACH do not run a full suite. For them, a custom assistant is often faster. Anthropic's Claude handles long contracts and supplier documents well and connects to ERP, document management and e-mail through the Model Context Protocol. OpenAI's GPT models and ChatGPT Enterprise are common for drafting RFQs and supplier communication. Mistral's open models can run on your own servers when pricing and contract data must not leave the company (see our Mistral guide). General agents such as Manus are useful for supplier research and market scans, but should never get purchasing authority or ERP write access.
Why most pilots stall
The surveys agree on the obstacles. In the ProcureCon CPO report, 67% name data privacy, security and compliance as the top barrier to scaling AI. GEP's leaders point to data quality: AI cannot classify spend or compare prices if supplier master data is full of duplicates. Gartner adds a general warning that over 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear value or weak risk controls. Our guide to AI-ready data covers the groundwork.
There is also a gap between pilots and process change. A chatbot that answers policy questions is easy. An agent that creates purchase orders, changes supplier records or commits the company to a price needs approval limits, audit logs and clear legal authority. That is where most pilots stop.
A safe rollout ladder for procurement AI: Level 1, assistants that read and summarise (policies, contracts, supplier answers). Level 2, drafts reviewed by a buyer (RFQs, requisitions, clause comparisons). Level 3, the agent prepares a transaction and a human approves it in the ERP. Level 4, autonomous action only for low-value, reversible purchases within hard limits set outside the model, for example tail-spend orders below a fixed amount from approved suppliers. Track savings and error rates per level before moving up. Our article on measuring AI ROI shows how.
What Gartner expects
Gartner predicts that by 2027, 50% of organisations will support supplier contract negotiations with AI-enabled contract risk analysis and editing tools. Hackett already sees 49% piloting or implementing AI in contract management, so that forecast looks within reach. Fully autonomous procurement is further away: in Gartner's 2026 Hype Cycle for agentic AI, only 17% of organisations had deployed AI agents at all, even though more than 60% expect to within two years.
The EU angle: AI Act, GDPR and public procurement
- EU AI Act. Most procurement AI, such as spend analysis, contract review and RFQ drafting, is not high-risk. It becomes sensitive when AI evaluates people, for example freelancers or contractors as individuals, or when it is used in hiring through staffing suppliers. Check use cases against Annex III. See our EU AI Act compliance guide.
- GDPR. Supplier contacts, freelancer contracts and e-mails contain personal data. Use a data processing agreement with every AI vendor and prefer EU-hosted or self-hosted models for sensitive data.
- Contract authority. An agent that negotiates or accepts an offer acts for the company. Define in writing what it may commit to, up to which amount, and log every offer and acceptance.
- Public procurement. Public buyers in Croatia publish tenders through the EOJN RH portal and must treat bidders equally. AI can help draft documentation and check bids, but evaluation decisions must remain traceable and human. When public bodies buy AI, the EU's model contractual clauses for AI procurement (MCC-AI) are a useful starting point.
What this means for companies in Croatia and DACH
- Start with contracts and intake. They have the most text, the clearest value and the lowest risk.
- Clean the supplier master data first. Duplicates and wrong categories make every AI result worse.
- Measure savings, not usage. Count avoided price variance, enforced contract terms and hours saved per buyer, not prompts or licences.
- Use the suite if you have one. SAP Ariba and Coupa agents already see your data. Without a suite, a Claude or Mistral assistant connected to your ERP is a faster first step.
- Keep humans on commitments. Let AI prepare and recommend; let people sign until the agent has a track record on your own purchases. Our article on agentic commerce shows where autonomous buying is heading, and our guide to AI vendor selection helps when you buy AI itself.
The bottom line
AI in procurement has passed the curiosity stage: most teams use it, and contract management and intake already deliver. But only a small minority has scaled it into the way procurement actually works. Realistic savings are 3 to 7% of spend, which is a lot of money, but only for companies that fix their data, set clear limits for agents and measure results. The rest will stay in the 59% that are still piloting.
Want AI to save real money in purchasing?
AI Workshop builds procurement assistants with Claude, MCP and your ERP, from contract review to guided buying, with the approval limits, audit trails and GDPR safeguards EU companies need. Our Claude Certified Architects start with one process and a measurable savings target.
Talk to an AI consultant