The 2026 data says procurement AI pays off inside document-heavy steps, not in chat windows. Most teams are still piloting, most AI use runs through general tools, and trust breaks when people cannot check an answer. The pattern that works is checkable AI: a suggestion inside the step, citing its source, that a human accepts or overrides.
We read the analyst reports, the surveys and the user reviews before deciding how AI should work in iProcure. This is what we found, with every source named. Where a survey was paid for by a vendor, we say so.
Where is procurement AI in 2026?
Mostly in pilots. Three sources, one of them vendor-sponsored, point the same way.
- Gartner placed generative AI for procurement in the Trough of Disillusionment in July 2025, citing "uneven ROI or falling short of expectations" and noting that "fragmented data and platform integration challenges are slowing progress."
- The Hackett Group's CPO Agenda 2026 found 59% of procurement organisations piloting generative AI and 12% running it at scale. 69% get their AI through a platform they already use. Where it runs at scale, measured gains were about 9.7% in productivity and 9.3% in cycle time, as summarised by Ironclad.
- ProcureCon 2026, a study sponsored by Icertis, found 65% piloting and 11% "implementing with measurable impact". The blockers were privacy and compliance (67%), data quality (54%) and resistance to change (51%).
Two more warnings are worth keeping in view. Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, and called "agent washing" common. And in May 2026 Gartner noted that without redesigning roles and processes, AI gains "remain confined to the individual level."
Where does AI actually earn its keep?
In steps that are heavy with documents or typing. BCG reported in April 2025 that generative AI made RFx drafting and bid comparison about 50% faster, and that the value splits roughly "10% algorithms, 20% data/technology, 70% people."
Ranking use cases by the evidence of value we could find, our own synthesis puts them in this order:
- Invoice capture and matching
- Contract extraction and summaries
- Spend analytics
- RFx drafting
- Bid comparison
- Intake that fills in and routes a request
Fully autonomous negotiation and autonomous tail-spend sourcing have strong single case studies, but most of that evidence is reported by the vendors themselves.
Why are chat assistants struggling?
Because people do not use them where the work happens. A 2026 benchmark by Suplari, a vendor, across 121 procurement teams found about 90% of procurement AI use goes through general tools (ChatGPT at 62%, Microsoft Copilot at 61%) and only 8% through the AI inside the procurement platform.
The large suite assistants have almost no public user reviews. Secondary reviews describe them as fine for lookups and weak on analysis, and advise treating answers "as a draft, not a source of record."
The pattern is not unique to procurement:
- Recon Analytics counts Microsoft 365 Copilot at 30 million or more paid seats on 450 million or more commercial seats, under 7%. Of users who tried it and stopped, 44.2% cite distrust of the answers.
- Salesforce retired Einstein Copilot, its chat assistant, in favour of Agentforce, which acts inside the workflow.
- a16z argues that chat "creates an alternative set of work that disrupts an otherwise native workflow," and that the ideal is for the workflow step "to become a click of a button."
- Nielsen Norman Group calls it the articulation barrier: "Articulating your needs in writing is challenging." It recommends hybrid interfaces over a blank prompt box.
When users complain about procurement tools, the complaints are about accuracy and plumbing: poor suggestions, misrouted approvals, broken integrations. They are not about missing AI.
What makes AI checkable?
A citation is not enough on its own. Research presented at AAAI 2025 (Ding et al.) found that citations raise trust even when they are random, and that trust falls once users check them. So the source has to be real, and it has to be one click away.
Checkable AI, as we define it, meets five tests:
- It lives inside the step. A draft request appears in the request form. A suggested score appears beside the score box. There is no separate AI screen to visit.
- It cites the source. A suggested score quotes the supplier's own sentence. A drafted field points to the words in your text it came from.
- A human decides. Every suggestion can be accepted, edited or overridden, with a reason, and each choice is logged.
- It never awards or signs. Awards, approvals and signatures stay with people.
- It is measured by use. Success is how often suggestions are accepted, edited or overridden, not how many chat sessions were opened.
What does this mean for how iProcure uses AI?
We made three decisions from this research. AI goes inside the steps, with no chat assistant and no named bots for now. The demo only shows AI that is real. And a chat assistant waits until it can answer from your own data.
Here is the honest status:
- Live: draft evaluation questions from your requirement in the eRFx module. You still set every weight by hand.
- Roadmap: in-step suggestions that cite their source, such as suggested scores in evaluation and draft requests in Intake.
- Parked until a customer asks: a negotiation bot, autonomous agents and forecasting.
Suggestions sit inside the step, citing the supplier's words; a human accepts or overrides; the AI never awards or signs. If you want to see why that matters for fairness, read about evaluation consensus, or see the evaluation flow live in 10 minutes.