In short
Supplier performance tracking used to mean a quarterly scorecard: on-time delivery, quality defects, price variance, a traffic-light rating. It still does in most companies. What changed between 2024 and 2026 is the evidence that AI can do something more useful: predict which purchase orders will arrive late, give a likely range rather than a single score, and explain a rating in plain language.
The evidence also shows the limits: off-the-shelf language models score suppliers poorly without domain data, and most procurement teams are still piloting. This guide covers what changed, the research, adoption, tool types and a five-step playbook for 2027.
What changed for supplier performance tracking
Disruption signals went up in 2025
Resilinc, a supplier-risk vendor, reports that its monitoring service issued 26,225 disruption alerts in 2025, up 38% on the year. Bankruptcy notifications rose 20%, profit warnings 50% and corporate restructuring 31% (Resilinc data, Illumination: 2025 Annual Supply Chain Report). It is one vendor's data, but the direction matters: financial stress shows up before missed deliveries, and a quarterly scorecard sees it too late.
Procurement teams have responded by keeping options open. In Deloitte's 2025 Global CPO Survey, the top risk-mitigation strategy was "maintaining active alternative sources (74%)" (Deloitte, 19 August 2025).
EU due-diligence rules moved
Corporate Sustainability Due Diligence Directive: the Omnibus I amendments, published in February 2026, narrowed scope to EU companies with more than 5,000 employees and net worldwide turnover above 1.5 billion euros. Member states transpose by 26 July 2028, and companies comply from 26 July 2029 (Latham & Watkins, 27 February 2026).
Forced Labour Regulation: this one has a 2027 date. It "will apply as of 14 December 2027" and bans products made with forced labour from the EU market, whether made inside or outside the EU (Linklaters).
A supplier that ships on time but cannot document its own supply base is now a performance problem too.
AI agents arrived, with a warning
Gartner predicts that 40% of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% in 2025 (Gartner, 26 August 2025). It also predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, "due to escalating costs, unclear business value or inadequate risk controls" (Gartner, 25 June 2025). Both are predictions, not results.
What the research says AI can do
Predict late deliveries from purchase-order history
The most direct evidence comes from a furniture manufacturer. Mohamed Aziz Zaghdoudi and colleagues used decision trees, random forests and XGBoost to classify which orders would be delayed, validated "using a real 3.5-year dataset, including 139 suppliers, 7,934 products and 89,080 purchase orders", and fed the predictions into supplier selection. They report lower delays and lower total procurement cost (preprint submitted to Computers & Industrial Engineering, December 2025). It is a preprint, so treat the results as promising rather than settled.
The data needed (order, promised and receipt dates) is usually already in the ERP.
Forecast a range, not a single score
A traditional scorecard gives one on-time percentage. A 2025 automotive study by Gabellini and colleagues compared traditional, machine learning and deep learning models for probabilistic forecasts of supplier punctuality over short and long horizons, judged on how well the predicted intervals covered what happened (IFAC-PapersOnLine, 2025). A supplier whose on-time rate swings widely needs a different plan from one that is steady at a slightly lower level.
Retrain the model, and accept some churn
A second paper, co-authored by Alexandra Brintrup of the University of Cambridge, fed machine-learning delay predictions into an order-allocation model for an automotive company. Periodic retraining "reduces prediction errors and total costs more effectively than models without continuous training", but "with increased order allocation instability" (Gabellini et al., Production & Manufacturing Research, 2025). Decide how much reallocation your supplier relationships can take before you switch it on.
Use language models to explain, not to guess
Language models can help evaluate suppliers when they are grounded in the right documents. An ICIS 2025 paper combined retrieval-augmented generation with step-by-step reasoning to rate the qualifications of suppliers for complex production lines, using data from a large Chinese manufacturer, and reports that the method "effectively generates accurate rating levels and interpretable management insights" (Li et al., ICIS 2025).
Without that grounding, results are weak. Researchers testing general-purpose models on multi-criteria decisions, including a supplier evaluation dataset, found they reached "around 60% accuracy rate compared to the evaluation ground truth". Better prompting raised this to around 70%; fine-tuning on domain data took it to around 95% (Wang, Zhang & Mu, arXiv 2502.15778, 2025). A generic chatbot is not a supplier scorecard.
Where adoption stands
The 2026 surveys agree that intent is running well ahead of practice:
The Hackett Group found 43% of procurement organisations actively pursuing AI deployment, nearly double the year before, but only 12% at large scale. Sixty-nine percent access AI through capabilities embedded in the procurement platforms they already use (Hackett Group, 17 March 2026).
The University of Mannheim and ISM found that 80% of organisations are still in the exploration or pilot phase, with no respondent reporting AI scaled and embedded in core processes (State of the Procurement Profession 2026).
Gartner found 63% of senior procurement leaders expect AI to significantly improve procurement performance over the next three years, yet only 25% have redesigned jobs or roles as a result (Gartner, 8 October 2026).
"Successful AI implementation at the functional level is fundamentally a role redesign challenge," said Fareen Mehrzai, Senior Director Analyst in Gartner's Supply Chain practice.
Where the work is done well, the gains are concrete. McKinsey describes a pharmaceutical company whose AI agents "track supplier delivery performance and automatically check invoices and POs against contract terms", cutting value lost through leakage by 4% (McKinsey, 5 February 2026). The same article estimates that procurement functions "use less than 20 percent of the data available to them to support decision-making".
What practitioners say
A small share of suppliers can hold back a whole business. Honeywell Aerospace found roughly 2% of its 3,000 suppliers underperformed in the first half of 2026, enough to stop output keeping pace with orders. "Frankly, I underestimated how long it would take to implement and see traction from the corrective measures we had taken and are taking," President and CEO James Currier said on the company's August earnings call. The company added sources for more than 50 constrained parts (Supply Chain Dive, 18 August 2026).
Buyers want answers they can rely on. "Make sure your tool is focused on really, really being good with the data and allowing for really good inputs, and there's no hallucinations," Nordstrom CPO Karoline Dygas told a Manifest 2026 panel (Supply Chain Dive, 18 February 2026).
They also want a reason before they spend. "You have to have a clear business case for what you want to achieve, and clear objectives, and then you can identify whether the juice is worth the squeeze," said Chris Martus, VP of procurement at The Mosaic Company, at ISM World 2026 (Supply Chain Dive, 21 May 2026).
Alan Amling of the University of Tennessee, Knoxville, asks one question of any AI tool: would you leave it running unsupervised over a weekend? "If the answer is no, and it has been no for months with no credible path to yes, you are not running AI," he said. "You are running an expensive suggestion engine" (Supply Chain Dive, 20 July 2026).
Tool types for supplier performance tracking
Tools differ more by job than by brand. Match the category to the question you are asking.
Category | What it answers | Examples | Watch for |
|---|---|---|---|
Embedded analytics in ERP and procurement suites | How did each supplier perform on delivery, quality and price, and is it getting worse | AI features inside the ERP, SRM or source-to-pay suite you already run | Whether it predicts or only reports; how often models retrain |
Supplier risk management platforms | Which suppliers show financial, operational or regulatory warning signs right now | Gartner published its first Magic Quadrant for this category in April 2025 (Everstream) | Coverage beyond tier 1, false-alert rate, source transparency |
Delay and OTIF prediction models | Which open orders are likely to arrive late, with a likely range | Built in-house on PO history or offered as an analytics module | Enough clean history per supplier; allocation churn after retraining |
LLM-assisted evaluation | Why a supplier scored the way it did, in plain language | Assistants grounded in your audits, contracts and scorecards | Answers without cited documents; generic models scoring without your data |
AI supplier discovery and vetting | Who else can make this when a supplier falls behind | Speya: describe the need in plain language; AI agents search millions of supplier profiles across 250+ data sources and return AI-vetted shortlists with sources beside every name | Whether every claim comes with a source you can check |
Most teams need more than one. A common 2027 setup: embedded analytics for the scorecard, risk sensing for early warning, and AI discovery to have alternatives ready before a supplier fails.
A 2027 playbook
Clean the delivery data first. Promised date, actual receipt date and quantity per order line are the minimum for any delay model. McKinsey's estimate that procurement uses under 20% of its available data suggests most of it already exists.
Predict on a few critical categories. Start with the suppliers that can stop production. McKinsey reports organisations "can go from prototype to pilot in weeks, and from pilot to scale in under a year" with a few key datasets and defined use cases.
Report ranges and explanations. Replace a single on-time percentage with a likely range, and require any AI rating to cite the records behind it.
Line up second sources before you need them. Use AI discovery to build vetted alternatives for single-sourced parts, and add forced-labour evidence to the qualification file ahead of 14 December 2027.
Redesign the role and the KPIs. Gartner's numbers show where projects stall. Decide who acts on a predicted delay, and measure that decision rather than the time spent on the scorecard.
Events to watch in 2027
North American Procurement Executive Summit (NAPES), 21–23 April 2027. The agenda includes "From Automation to Advantage: How AI Is Transforming Procurement Performance" by Heleen Du Toit, VP Procurement at PepsiCo (agenda).
ISM World 2027, 16–18 May 2027, Washington DC (ISM).
Gartner Supply Chain Symposium/Xpo, 24–26 May 2027, Barcelona (Gartner).
Frequently Asked Questions
Can AI predict late supplier deliveries?
Yes, within limits. A 2025 study using 89,080 purchase orders from 139 suppliers showed tree-based models classifying likely delays well enough to improve supplier selection. Results depend on clean order history and regular retraining.
Can a general AI chatbot score my suppliers?
Not reliably. In published tests, general-purpose models reached around 60% accuracy against an expert ground truth, and around 95% only after fine-tuning on domain data. Use language models to explain ratings drawn from your own records, with sources.
How is AI changing supplier performance management in 2027?
Mostly through tools teams already own. Hackett found 69% of organisations access AI through capabilities embedded in existing procurement platforms, and only 12% have scaled it. The 2027 task is turning scorecards into forecasts and deciding who acts on them.
Does the CSDDD still apply after the Omnibus?
Yes, but to fewer companies and later. It now covers EU companies with more than 5,000 employees and over 1.5 billion euros in net worldwide turnover, with compliance from 26 July 2029. The EU Forced Labour Regulation applies earlier, from 14 December 2027, and covers all products on the EU market.
Sources
Every source above is linked inline. Key references: Gartner press releases (25 June 2025, 26 August 2025, 8 October 2026); The Hackett Group 2026 Key Issues release (17 March 2026); McKinsey, "Redefining procurement performance in the era of agentic AI" (5 February 2026); Deloitte 2025 Global CPO Survey; University of Mannheim and ISM, State of the Procurement Profession 2026; Resilinc, Illumination: 2025 Annual Supply Chain Report; Latham & Watkins on Omnibus I (27 February 2026); Linklaters on the EU Forced Labour Regulation; Zaghdoudi et al. (preprint, 2025); Gabellini et al., Production & Manufacturing Research (2025) and IFAC-PapersOnLine (2025); Li et al., ICIS 2025; Wang, Zhang & Mu, arXiv (2025); Supply Chain Dive (February, May, July and August 2026).