← All articles
- Should-Cost Analysis
- Strategy
- Category Management
- Supply Chain Risk
Procurement has been optimising the wrong end of its own spend base
The profession's technology investment has concentrated on transaction speed. The argument for putting it into analysis instead, and what that position commits us to.
By EXOS Research · · 11 min read
A position, stated before the evidence
Most of what procurement technology has achieved over the past decade has been achieved at the transactional end: faster requisitions, cleaner approvals, better invoice matching, less administrative friction per purchase order. That work is real and we are not disputing its value.
Our position is that the profession's investment priority has been misallocated, and that the returns available from analysing a small number of significant decisions properly exceed the returns available from processing a large number of insignificant ones faster. This is a claim about where the next euro of technology budget should go, not a claim that the previous one was wasted.
We publish it as a position because we act on it. It determines what we build and, more usefully for the reader, what we decline to build.
The efficiency case is easier to make, which is why it wins
There is a structural reason the transactional end attracts the investment, and it has nothing to do with where the value is.
Transaction cost reduction produces a business case that fits on one slide. Volume multiplied by cost per transaction, before and after. The baseline is in the system, the improvement is countable, the payback period is calculable, and the finance function accepts the arithmetic without argument.
Strategic analysis produces no comparable artefact. The value of having modelled a renewal properly is a contract signed at terms that would otherwise have been worse — a counterfactual, unmeasurable by construction, and invisible in any report. The disruption that did not occur has no line item.
So the profession invests in the thing it can prove and underinvests in the thing that matters more. Not through poor judgement, but because one of them survives contact with a business case and the other does not.
The loss happens before execution, which is not how the figure is usually read
There is one measurement of this, and it says something more specific than it is usually quoted as saying.
McKinsey's analysis of more than 340,000 transformation initiatives found that the average procurement savings pipeline loses one third of its estimated value in the planning stages and a further 20% in execution. Late additions to the pipeline typically recover only about 15% of the original projected value. (McKinsey — Aim higher and move faster for successful procurement-led transformation)
The figure is generally cited as an argument for execution discipline — tighter tracking, better contract compliance, a system to stop savings leaking after the deal is done. Read as published, it does not say that. The larger share of the loss occurs in planning, before execution begins. That is a pre-award finding, and the tooling that addresses execution discipline cannot reach it.
The RFQ bot is a strong narrative and a weak diagnosis
The most confident version of the transactional argument currently on offer is the autonomous agent that issues requests to suppliers, collects bids, and awards on defined logic. It is a genuinely impressive capability, and within its range it is the most efficient thing available.
Its range is the point. Optimisation of that kind requires a settled specification, a supplier field large enough for the outcome to be contested, and enough event volume for automation to repay itself. Read against the Kraljic matrix, those are the leverage and non-critical quadrants, and there the technology earns its keep. (Where the hidden cost actually sits — the Kraljic reading)
Ask any experienced practitioner where their difficult categories fail, and the answer is not that the requests went out too slowly. It is upstream of the request: whether the specification was right, whether the total cost across the term was understood, whether the dependency on a single supplier was ever classified as such, whether the walk-away point existed before the meeting began. Automating the issuing of requests addresses none of these. In a low-frequency high-value deal, establishing the parameters is the work, and an agent operates on the parameters it is handed. (Autonomous sourcing, defined)
This is the iceberg argument in a specific form. Sourcing events are the visible portion. The cost is determined below them.
The two ends of the spend base fail differently
| Tail spend | Strategic spend |
|---|
| Failure mode | Accumulation — many trivial inefficiencies aggregating into drag | Single events — one renewal, one substitution, one unclassified dependency |
| Visibility | Gradual and visible in reporting | Frequently invisible at the time |
| Remedy | Process and automation | Analysis before the decision |
| Magnitude | Proportional to volume | Capable of exceeding a year of tail-spend savings on its own |
| Who addresses it | Procure-to-pay, intake, guided buying, autonomous sourcing | Largely unaddressed by systems; done in spreadsheets or not at all |
McKinsey Global Institute's modelling across 23 value chains gives the order of magnitude on the right-hand column: disruptions lasting a month or longer occur roughly every 3.7 years and cost the average company in the region of 45% of one year's EBITDA over a decade. Sector variation is wide, with consumer goods nearer 30%. No improvement in invoice-matching throughput appears anywhere in that figure. (MGI — Risk, resilience, and rebalancing in global value chains)
The current memory shortage is what this looks like in practice
Generative AI workloads require high bandwidth memory, which consumes substantially more wafer capacity per unit than conventional DRAM. Manufacturers have reallocated advanced process nodes and new capacity toward server and HBM products, and TrendForce reports the effect propagating into legacy components as Taiwanese mature-node suppliers attempt to fill the DDR4 and DDR3 gaps.
Date-stamped, because it moves quickly. DRAM prices were up around 172% year-on-year by the end of Q3 2025. Conventional DRAM contract prices rose approximately 93–98% quarter-on-quarter in Q1 2026, lifting industry revenue 81% to some $97bn. TrendForce's July 2026 forecast has a further 13–18% quarter-on-quarter rise in Q3 2026, with NAND at 10–15% — moderating, from a high base, with supply still short of demand. (TrendForce, 3 July 2026 · TrendForce, 9 July 2026)
An organisation with an efficient transactional stack and no analytical capability meets this situation with nothing. It cannot negotiate with an allocation-constrained supplier, cannot trace wafer reallocation onto a component two tiers down, and cannot model what a redesign or a buffer would cost against what the shortage will. The requirement is analytical and it sits before the transaction.
Most organisations cannot see far enough to do it. McKinsey's 2025 survey work found the majority understand supply chain risk only to tier one, with tier-2 awareness having fallen in 2023 and 2024. The exception is instructive: tariff compliance obligations produced a 22 percentage-point increase in tier-two visibility. The constraint has been priority, not feasibility. (McKinsey — Supply chain risk pulse 2025)
The credibility problem is downstream of the same choice
Procurement's difficulty in having its savings believed is usually treated as a reporting issue. It is closer to a consequence of what the function has been equipped to measure.
Protiviti's study of 440 finance and procurement executives found 48% of finance leaders believed 20% or less of reported procurement savings reached the bottom line, and only 16% of respondents said procurement was seen internally as a profit centre. That study is from 2017. (Protiviti — Bridging the gap between finance and procurement) The more recent picture is similar in kind: a 2025 survey of over 100 CFOs by Fine Tune and CFO Leadership found 38% expressing high confidence that procurement savings reach the P&L, falling to 29% at companies above $250m revenue. (CFO Leadership / Fine Tune, September 2025) Eight years apart, the finance-side view has not moved.
Transaction-level efficiency generates activity metrics. Analysis generates positions that can be defended in front of finance: a reconstructed cost base, a modelled alternative, a quantified dependency. Bain's published estimate is that a strong procurement function can reduce the purchasing cost base by 8–12% with a further 2–3% annually, and those figures only matter if they survive that conversation. (Bain — Unearthing the hidden treasure of procurement)
What should-cost reconstruction actually involves
Of the four practices below, should-cost is the one most often described and least often shown, so it is worth setting out the mechanism rather than asserting the value.
A supplier announcing an increase is making a claim about their own cost structure. That claim decomposes. Build the supplier's cost base from its components — materials, energy, labour, logistics, overhead, margin — establish the weight of each, and apply the movement in the relevant published index only to the component it affects.
The arithmetic is unremarkable and that is the point. Where materials account for roughly 61% of total cost, a 15% rise in material prices justifies about 9% at product level, not 15%. The remaining 6% is a claim about something else, and the supplier can be asked what.
A model built on clean data and category knowledge can be expected to fall within roughly ±10–15% of actual cost. That is an accuracy band rather than a savings figure, and it is deliberately stated as one: a stated error margin is what makes the model arguable in front of a supplier, and we have not found a credible primary source for any of the savings percentages attached to should-cost in vendor material.
Models go stale. Annual revision is the minimum, quarterly for volatile categories.
What we are committing to, and what we are not
Four practices sit at the analytical end. None is novel; the intended reader has met all of them. Each requires an object the transactional systems do not produce: an assembled view of one decision.
| Practice | What it produces | What it requires |
|---|
| Should-cost reconstruction | The share of a claimed increase that the supplier's own cost structure justifies | A component-level cost base and the relevant published indices |
| Explicit walk-away definition | The best alternative and the resulting zone of agreement, established before the negotiation | An assessment of how many credible alternatives exist in the category |
| Stress-testing against recovery objectives | High and low risk ratings converted into quantities: weeks of buffer, alternative capacity required, time to qualify a second source | A defined failure scenario and a stated recovery objective |
| Savings definitions agreed with finance in advance | A baseline both functions accept before the negotiation rather than after | A methodology chosen and documented up front — historic, budget, market benchmark or technical |
On the second of these, the only quantified evidence available comes from public procurement, and it is worth stating with its limitation attached. The OECD finds single bidding associated with prices higher by 9.6% (OECD — Maximising the benefits of effective competition in public procurement), and the European Court of Auditors records the EU single-bid rate rising from 23.5% in 2011 to 41.8% in 2021 (ECA Special Report 28/2023). Peer-reviewed analysis of more than 17,000 Finnish invitations to tender adds that the price effect of competition is larger in low-competition environments — the value of one credible alternative is highest exactly where alternatives are scarcest (Journal of Public Procurement — Anatomy of competition in public procurement). This transfers to private negotiation by analogy rather than by measurement.
The commitment this implies is narrow. EXOS is a scenario-based procurement analytics system: it models total cost of ownership, negotiation scenarios and supplier risk before a contract is awarded. It reasons in projects and categories rather than transactions, and it runs upstream of procure-to-pay and ERP platforms rather than in place of them.
It does not address tail spend, and we do not intend it to. That work belongs to the systems built for it, which do it better than we would. Our argument is not that those systems should not exist. It is that they have absorbed a share of the profession's attention out of proportion to what they can change, and that the next increment of effort returns more at the other end of the spend base.
The full classification of procurement AI systems → · Six kinds of procurement AI, as we see them →
Sources:
Savings pipeline: McKinsey — Aim higher and move faster for successful procurement-led transformation (analysis of 340,000+ transformation initiatives)
Disruption economics: McKinsey Global Institute — Risk, resilience, and rebalancing in global value chains (August 2020) · McKinsey — Decoding disruption to reshape manufacturing footprints (January 2026) · McKinsey — Supply chain risk pulse 2025
Memory market: TrendForce press release, 3 July 2026 · TrendForce press release, 9 July 2026
Savings credibility: Protiviti — Bridging the gap between finance and procurement (n=440, 2017) · Fine Tune and CFO Leadership CFO survey (September 2025) · Bain & Company — Unearthing the hidden treasure of procurement
Competition and price: OECD — Maximising the benefits of effective competition in public procurement · European Court of Auditors, Special Report 28/2023 · Journal of Public Procurement — Anatomy of competition in public procurement (Finnish tenders, 2010–2017)