AI in Procurement: What Agents Can Run Alone and What Humans Must Own

Almost every procurement function is now touching AI, but very few have changed how work is done. McKinsey's 2025 state of AI survey found that 88% report regular AI use in at least one function and 62% are at least experimenting with AI agents. Yet only 23% say their organisations are scaling an agentic AI system, and in any given business function no more than 10% are scaling agents. Only 39% report any enterprise-level EBIT impact from AI.

Procurement-specific data tells the same story. The Hackett Group's 2025 Key Issues Study reported that about 49% of procurement teams piloted generative AI use cases in 2024, while only 4% reported large-scale deployment. A ProcureCon Insights survey of North American leaders, sponsored by Icertis, found that 90% of procurement leaders have considered or are already using AI agents. Considering is not deploying, and the gap between the two numbers is the real state of the market.

For a chief procurement officer, the implication is that peer behaviour is a poor guide. Most competitors are running pilots, which means the advantage still sits with whoever converts one pilot into a governed production workflow. It also means vendor claims deserve scrutiny: Gartner estimates that only about 130 of the thousands of agentic AI vendors are real, and predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear value or inadequate risk controls.

Regional starting points differ, though the global pattern holds. Deloitte's survey of over 250 CPOs spans 40 countries and shows technology budgets rising across markets, which suggests the constraint is execution rather than appetite. In our judgement, US and UK functions tend to lead on tooling, European groups face the heaviest regulatory design burden, and Gulf organisations, where procurement is often concentrated in a few large buyers, have an opportunity to design agent-ready processes from the outset instead of retrofitting them. These are views, not measured findings, and each leadership team should test them against its own estate.

What to do: stop counting pilots and start counting workflows in production. Pick a single measure, such as the share of tail-spend purchase orders handled without a buyer touching them, and report it monthly. If the figure is zero after two quarters, the programme is a demonstration, not a transformation.

The evidence for value is consistent across the major advisers, and it is more modest than the vendor narrative. Bain's 2026 analysis states that AI-enabled procurement can lift return on investment up to five times while boosting productivity by 60%, enabling incremental savings of 3% to 7%. BCG's 2025 analysis says up to 15% savings is achievable in the medium term, that AI and GenAI can automate up to 75% of procurement tasks, and that negotiation results improved by approximately 2-3% where AI was applied.

Case evidence points the same way. Bain describes one tool that delivered 3% to 5% savings and reduced the time to develop category strategies by 90%, and a global bank whose agentic procurement solution is projected to save up to $180 million. These are projections and single cases, not guarantees, but they show where the leverage sits: in analysis, preparation and coverage rather than in replacing the negotiator.

The capacity argument may matter more than the savings argument. The Hackett Group projected that procurement workloads would rise 10% in 2025 while budgets grew just 1%, creating a 9% efficiency gap. Deloitte's 2025 Global CPO Survey of over 250 CPOs across 40 countries found leaders allocating up to 24% of their budgets to technology, and reported that its Digital Masters achieved an average 3.2x return on GenAI investment. Spending is rising; returns depend on how it is deployed.

Be wary of how savings are counted. Agents tend to be deployed first on spend that was previously unmanaged, so the first savings often come from coverage rather than from better prices on managed categories. That is legitimate value, but it should be reported separately, because blending the two makes the managed-category result look better than it is and sets an expectation that cannot be repeated. Finance should agree the baseline method, the attribution rules and the treatment of cost avoidance before the first number is published, and internal audit should be invited to test the method once.

Peer comparison needs care. The published studies differ in sample, definition and date: McKinsey surveys organisations across functions, Hackett reports on procurement teams, Deloitte on CPOs and Bain and BCG on engagements and modelled potential. Treat the numbers as a range of plausible outcomes. Before presenting a business case, rebuild the estimate from your own spend profile, category mix and labour cost, and show the board both the central case and the downside if adoption is slower than planned.

What to do: frame the business case around released capacity and coverage of previously unmanaged spend, and use the savings range as a sanity check rather than a target. A programme promising double-digit savings from agents alone is asserting more than the published evidence supports.

The strongest proof that an agent can negotiate without a human is also the oldest. Harvard Business Review reported that Walmart's negotiation chatbot was negotiating and closing agreements with 68% of suppliers approached. EY's analysis of AI chat negotiation reported that one retailer gained 1.5% in savings on negotiated spend and extended payment terms by an average of 35 days. These deals suit automation because the value per supplier is small, the terms are bounded and the supplier would not have received a senior buyer's attention anyway.

That pattern generalises into a test. An agent can act alone where the category is low in strategic value, the outcome can be verified quickly, the commitment is capped and a mistake can be reversed. Catalogue purchases, repeat quotations, renewals of standard terms, invoice and contract matching and price benchmarking meet that test. BCG's estimate that AI and GenAI can automate up to 75% of procurement tasks is plausible precisely because so many tasks are this kind of bounded work.

The implication is that the first agent should be given a tier, not a task list. A mandate defined by spend band, supplier type and contract value is easier to govern than a mandate defined by technique. It is also easier to explain to suppliers, finance and internal audit, all of whom will ask who authorised the agent to commit the company.

Supplier experience deserves explicit design. Autonomous negotiation works when the supplier understands the rules: a clear brief, a bounded range of acceptable terms, a stated deadline and a named human to call if something goes wrong. The Walmart result is often quoted as proof of the technology, but it is equally proof of design, since the supplier population was chosen because the deals were simple and the alternative was no negotiation at all. Where an agent displaces an existing human conversation, the commercial risk is higher and the evidence base thinner.

Contract intelligence is the quieter companion to autonomous buying. Extracting dates, price mechanisms, renewal terms and liability clauses from a repository gives an agent the context it needs to act, and gives buyers a view of exposure they rarely have. It is also low in risk because the agent reads rather than commits. Many functions should begin here, prove accuracy against a sample reviewed by lawyers, and only then connect the same data to renewal and reorder workflows.

What to do: segment spend into three tiers, define the ceiling for each, and allow an agent to complete the transaction alone only in the lowest tier. Require a human to approve anything above the ceiling, any new supplier and any deviation from standard terms. Exhibit 4 sets out one way to structure this.

Autonomous negotiation breaks down where value comes from things a mandate cannot capture. A strategic supplier relationship involves capacity commitments, joint innovation, resilience under stress and the willingness to give and take over years. Gartner predicts that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024. That leaves a large majority of decisions with people, and the highest-value ones are the least likely to move.

The measured gains from AI in negotiation support this reading. BCG reports improvements of approximately 2-3% in negotiation results, which is meaningful at scale but is a gain from better preparation, benchmarks and scenario analysis, not from handing over the table. Hackett found that 64% of procurement leaders expect AI and GenAI to transform their roles within five years, which describes changed roles rather than removed ones.

There is a second-order effect that CPOs should plan for. Gartner forecasts that AI agents will intermediate more than $15 trillion in B2B spending by 2028. If buyers' and suppliers' agents increasingly negotiate with each other, advantage will shift to organisations that have clearer objectives, better data and sharper walk-away positions. Those are human decisions that must be made before an agent is given a brief.

A practical test helps leaders decide where a category sits. Ask four questions: would we be comfortable explaining this award to the board, would losing this supplier damage our operations, does the price depend on non-price terms such as capacity or innovation, and could a mistake be corrected within a quarter? Two or more answers in the strategic direction should put the category in human-led territory regardless of how capable the available agent appears. The test is deliberately simple so that category managers and finance can apply it consistently without specialist help.

What to do: redefine the category manager's job as owning objectives, trade-offs and relationships, with agents providing preparation, monitoring and follow-up. Protect time for the top 20 or so strategic suppliers in each category by moving the routine work elsewhere. That reallocation is the real return on the programme.

Most failures will trace back to the organisation rather than the algorithm. Deloitte's 2025 survey asked CPOs about barriers to value delivery: 57% cited siloed ways of working, 46% competing priorities diluting focus, 40% the organisational or technology capability to support execution and 34% the talent gap. None of these is a model problem. They are problems of ownership, sequencing and skills.

The data position is equally stark. In the ProcureCon Insights survey, 88% cited integration issues and 75% data quality issues as detracting from procurement confidence in AI. Bain cautions against two beliefs that stall progress, that AI will fix messy processes and that an organisation needs perfect data before it can start. Both are wrong in practice: agents amplify whatever process they are given, but a single well-scoped data domain can be cleaned in weeks.

BCG offers a useful sense of proportion. It estimates that a typical digital transformation breaks down into 10% from AI, 20% from technology and data, and 70% from people, processes and organisation. For a CPO, the implication is that the budget line for the agent is the smallest part of the investment. Change management, intake redesign, supplier communication and policy rewriting will consume most of the effort and determine most of the outcome.

Skills are the quieter constraint. Deloitte's finding that 34% of CPOs cite the talent gap understates the shift in the work itself, because supervising agents requires buyers who can specify objectives precisely, interpret exceptions and challenge outputs. That is closer to commissioning and quality assurance than to traditional purchasing. Leaders should plan role definitions, training and incentives for it now, since a team that treats the agent as a black box will either over-trust it or ignore it, and both outcomes destroy value.

What to do: before selecting a platform, audit the chosen tier on four points: one intake route, one supplier master, one approval matrix and one contract repository. Where any of those is fragmented, fix it first. If a vendor demonstration works only on clean sample data, treat that as information about your readiness, not about the product.

An agent that can commit company money is a delegated authority, and should be governed like one. That means a written mandate, a financial ceiling, a defined list of permitted counterparties, a duty to escalate and an audit trail that a finance director or regulator can follow. Gartner's forecast that over 40% of agentic AI projects will be cancelled by the end of 2027 cites inadequate risk controls alongside cost and unclear value, so governance is a determinant of survival, not an afterthought.

Regulation is tightening around the edges. The European Commission states that the EU AI Act became applicable on 2 August 2026, with some exceptions, and that high-risk AI systems in certain areas face obligations from 2 December 2027. Most procurement agents will not be classed as high-risk, but any that screen suppliers in ways affecting people, or process personal data, will need documented risk assessment. Organisations operating in the UK, EU and Gulf should expect scrutiny of automated decisions to rise.

Supplier relationships need governance too. A supplier told at the last minute that it is negotiating with software will ask who is accountable for the terms. Transparent disclosure, a named human contact and an easy route to escalate preserve trust and reduce disputes. Walmart's reported 68% closing rate suggests suppliers will engage with automated negotiation when the process is clear and the terms are fair.

Security and confidentiality complete the picture. An agent that reads contracts, price lists and supplier correspondence handles commercially sensitive information, and any external model used must be configured so that this material is not retained or reused for training. Procurement should therefore work with information security from the first week, agree which data classes may leave the environment and require vendors to document where processing occurs. Gulf and European data-residency expectations make this a design constraint rather than a late-stage compliance check.

What to do: write the agent's mandate before configuring it, test it against five adverse scenarios, such as a supplier who disputes terms, a price spike or a sanctions flag, and name a human owner for every agent. Review logs monthly in the first year and report exceptions to the audit committee.

A sensible programme has four phases and a deliberate pace. In the first quarter, choose one tier and one category family, fix the data and intake for it, and define the mandate. In the second quarter, run the agent in recommend-only mode alongside buyers, measuring accuracy, cycle time and supplier response. In the third, grant autonomy within the ceiling for the proven tier. In the fourth, extend to adjacent categories and begin agent-assisted preparation for strategic negotiations.

The measures that matter are few. Track the share of tier-three spend handled autonomously, exceptions per hundred transactions, cycle time from request to order, savings against baseline and buyer hours redeployed to strategic suppliers. The Hackett Group reports that AI-driven procurement tools delivered up to 10% improvements in productivity, quality and cost savings in 2024, so set expectations at that order of magnitude in year one and let the evidence move them.

Funding should follow the same logic. Deloitte reports that CPOs are devoting up to 24% of budgets to technology; the question is whether the spend is concentrated on a few governed workflows or scattered across tools. Concentration is cheaper to govern and easier to measure, and it produces the internal evidence that unlocks the next tranche of budget.

Communicate the programme honestly inside the function. Hackett found that 64% of leaders expect their roles to be transformed, and teams will read the arrival of agents through that lens. Leaders who explain which work moves to agents, which stays with people and how released time will be used earn cooperation and better exception handling. Those who stay silent find that the people best placed to spot an agent's errors are the least motivated to report them.

Finally, decide what success looks like at the end of the year in terms the board will recognise. A credible statement would be that one tier of spend is managed by a governed agent, exception rates are stable, buyers spend measurably more time on strategic suppliers, and the function has a pipeline of the next three workflows ranked by value and readiness. That is a less dramatic claim than autonomous procurement, but it is one the evidence supports and one that compounds.

What to do: agree the twelve-month plan with finance and risk before the first vendor contract is signed, and set a stop rule. If exception rates or supplier complaints exceed an agreed threshold, autonomy is withdrawn until the cause is fixed. That is the discipline that turns an experiment into an operating capability.

Karter McKinley supports procurement leaders through its AI & Technology implementation service: scoping the first agent tier, rebuilding intake and data foundations, writing the mandate and governance model, and measuring results against an agreed baseline. We work alongside your procurement, finance and risk teams and prefer the simplest tooling that fits your existing systems.