Autonomous AI in Procurement Fails Mainly on Data, Processes, and Accountability

BCG published the results of its 2026 technology-procurement study. According to the report, companies gain the first benefits of autonomous systems inside operations: they classify requests faster, prepare supporting material, compare bids, and check contract performance. The best results do not come from isolated pilots, but where a company combines deployment with process redesign, skills development, and firm accountability rules. The authors identify four main obstacles. The first is varied inputs and uneven data quality. The second is difficult integration with legacy business systems. The third is the need to change procurement while keeping operations running, and the fourth is restrictions arising from risk management. A product may evaluate one document correctly, but commercial value will not emerge if it cannot access contracts, supplier records, purchase orders, and actual prices. A former chief information officer at a European media company described a common disappointment: “Many expect the tool itself to solve the problem quickly.” Without integration into the application environment, workflows, and data foundation, value does not appear. A pilot should therefore assess more than answer accuracy. It must test whether the solution can be connected, maintained, monitored, and transferred safely into everyday operations. BCG recommends starting with the commercial outcome, not supplier selection. The chief procurement officer should first decide whether the goal is to shorten tendering time, reduce price variation, limit off-contract buying, or improve negotiating power. Only then should the process be redesigned and tasks divided between the system and people. Released capacity creates no saving if employees keep performing the same checks and the company leaves work allocation unchanged. Independent decision-making is the most sensitive area. A system may recommend a supplier ranking, draft a counterproposal, or flag a suspicious price. The company must define when it may only recommend and when it may carry out an action. Each boundary requires a financial limit, a dual-control rule, a record of the data used, and the ability to stop the intervention immediately. These measures protect against both erroneous purchases and disputes over authority. Legal risks concern bid confidentiality, personal data, trade secrets, and competitive conduct. If a tool processes prices from several suppliers or uses an external model, the contract must state whether data is stored, used for further training, or shared with third parties. The software supplier’s liability must not replace internal control. The final selection decision should remain explainable and documented in the procurement file. The article concludes that the main constraint is no longer model capability, but organisational design. The highest returns come to companies that turn procurement from administrative transaction processing into a source of market and supplier intelligence. “This is not merely a more efficient version of today’s procurement,” BCG stresses. The technology makes sense when it improves commercial judgement, not when it only produces the same forms faster. Terms explained Autonomous procurement system: Software capable of preparing or executing connected procurement tasks with limited human intervention. Legacy business system: Previously installed software that is often difficult and costly to connect with new tools. Dual control: A rule requiring a significant action to be independently checked or approved by at least two authorised people. Off-contract buying: Purchasing goods or services without using an approved supplier or agreed commercial terms.

Autonomous AI in Procurement Fails Mainly on Data, Processes, and Accountability

The full article is available to users registered at Hard Skills