Should-Cost Modelling and Price Benchmarking — AI for Strategic Sourcing and Supply Risk

What you'll learn

Build an AI-assisted should-cost model whose assumptions are visible and individually challengeable
Identify which cost drivers move together, so a model does not overstate its own precision
Distinguish a benchmark that reflects a genuinely comparable transaction from one that does not
Use a should-cost model in negotiation without overstating what it proves

A should-cost model estimates what an item ought to cost by building it up from inputs — materials, labour, energy, overhead, logistics, margin — rather than accepting the supplier's price as given. It converts a negotiation from a contest of positions into a discussion about a specific number, which is why it is the most powerful analytical tool in strategic sourcing.

AI has made should-cost modelling dramatically cheaper. Models that once took an engineer weeks can be assembled in hours from commodity indices, wage data, and energy prices. That accessibility is genuinely valuable and it introduces a specific hazard: a model built quickly from public data, presented to three decimal places, whose assumptions nobody can name.

Assumptions Must Be Individually Challengeable

A should-cost model is a set of assumptions with arithmetic on top. Its credibility rests entirely on whether each assumption can be stated, sourced, and argued about separately.

This is where AI-generated models most often fail. Asked to estimate the cost of a machined component, a model will produce a plausible build-up: so much material at a given price per kilogram, so much machining time at a given rate, an overhead percentage, a margin. Every figure is reasonable. Very few are sourced, and some are effectively invented — plausible values drawn from the general shape of such models rather than from anything about this part, this process, or this supplier's region.

Taken into a negotiation, that model fails on first contact. A supplier who actually makes the part knows their cycle time and scrap rate, and one unsupported assumption discredits the whole exercise, including the parts that were right.

The requirement is therefore structural rather than a matter of care: every line must carry its source and its basis. Material price from a named index on a named date. Labour rate from a named wage dataset for the specific region and skill level. Cycle time from a named engineering estimate, a benchmark part, or a supplier disclosure. Where a figure is an assumption rather than a source, it must say so and carry a range.

A model where three lines are sourced and two are assumptions with stated ranges is more useful than one where all five look authoritative and two are fabricated, because you know where to concentrate the conversation.

Critical

A should-cost model is a negotiation instrument, and its power comes from being defensible line by line. One unsupported assumption that the supplier can refute discredits the entire model, including the well-founded parts. Never take a model into a negotiation containing a figure you cannot source.

Correlated Drivers and False Precision

Cost models present a total, often with a confidence range. That range is usually calculated as though the input uncertainties were independent, and in most supply chains they are not.

Energy prices, transport costs, and many commodity prices move together, because energy is an input to production and freight for nearly everything. A model treating a 10 percent energy uncertainty, an 8 percent freight uncertainty, and a 12 percent material uncertainty as independent will combine them into a narrower total range than reality supports, because independent errors partially cancel while correlated ones compound.

The practical consequence is that a should-cost model is most likely to overstate its precision exactly when input markets are volatile — which is exactly when you most want to rely on it.

Two defences. Stress the correlated set together: rather than varying each input separately, model an adverse case where energy, freight, and energy-intensive materials all move in the same direction, and see what happens to the total. Present a range, never a point. A should-cost output of "between 8.40 and 9.90 per unit, driven mainly by the alloy price assumption" is honest and negotiable. A point estimate of 9.14 invites a challenge to the third digit, and loses.

Benchmarks and Genuine Comparability

Price benchmarking compares what you pay against what others pay. AI makes it easy to assemble comparisons from market data, published contracts, industry surveys, and internal spend across business units. The difficulty is that a benchmark is only meaningful if the transactions are genuinely comparable, and price differences usually reflect real differences rather than negotiation failure.

Dimensions that commonly differ and that are frequently ignored:

Volume and commitment. A unit price against a committed annual volume is not comparable to a spot price.
Specification. Nominally identical parts differ in tolerance, material grade, testing, and certification, and those differences are frequently the whole price gap.
Incoterms and logistics. Ex-works and delivered-duty-paid prices are not the same number.
Payment terms. Sixty days against fourteen is a financing cost embedded in the unit price.
Service scope. Included engineering support, quality documentation, or inventory holding.
Contract duration and currency. A price fixed for three years in a volatile currency carries risk the supplier has priced.

An AI-assembled benchmark showing you paying 14 percent above market is a hypothesis. The work is establishing whether the comparators match on those dimensions, and often they do not — which is itself useful, because "we pay more because we buy in small quantities on short lead times" identifies a demand-management opportunity rather than a negotiation one.

Knowledge check

An AI-generated should-cost model for a machined component produces a total of 9.14 per unit with a confidence range of plus or minus 4 percent. Reviewing it, you find the material price is sourced from a named index, the labour rate from regional wage data, but the cycle time and scrap rate carry no source. What is the correct action before using it in a negotiation?

Select one answer.

A model that won the argument by conceding two of its five lines

Strategic Sourcing Manager, consumer goods

Context

A sourcing manager was preparing to challenge a 7 percent price increase on an injection-moulded packaging component. She built a should-cost model using an AI tool, which produced a detailed five-line build-up and a total suggesting the current price was already 11 percent above cost-justified levels, before the proposed increase.

Action

Rather than taking the model into the meeting as generated, she worked through each line for its source. Resin price came from a named polymer index and was solid. Energy and freight were sourced from published regional data. Cycle time and cavitation were not sourced at all — the tool had produced typical values. She contacted the supplier's technical account manager, framed it as wanting to understand the process, and obtained the actual tool cavitation and approximate cycle time. Both differed materially from the model's assumption, which reduced her calculated cost gap from 11 percent to about 4 percent.

Outcome

She took the revised model to the negotiation, opened by stating the two lines the supplier had provided and that she had used their figures, and concentrated the discussion on the resin index, which had fallen 9 percent since the last price review. The supplier accepted the index argument and withdrew the increase. The manager's assessment afterwards was that conceding the two lines was what made the resin argument land: the supplier could see the model used their own data where they had it, so the parts they could not refute carried weight. The original 11 percent version would have been refuted on cycle time in the first five minutes.

Quick check

Why does this lesson argue that a should-cost model's stated confidence range is usually narrower than reality supports?

Select one answer.

Exercise

~30 min

Your Task

Build or take an existing should-cost model for a component or service you buy. For every line, record the source and classify it as sourced, estimated from a comparable, or assumed. For any line classified as assumed, either obtain a basis or attach an explicit range. Then run a correlated stress test: instead of varying inputs independently, move energy, freight, and any energy-intensive materials adversely together and record the effect on the total. Finally, restate the output as a range with the dominant driver named, and write the one sentence you would use to open a negotiation with it.

Success looks like

Every line is classified by source type, with assumptions explicitly labelled rather than blended in
The correlated stress test moves related inputs together rather than independently
The output is expressed as a range naming its dominant driver, not as a point estimate
Any figure that could be obtained from the supplier is identified as worth asking for before the negotiation

Watch out for

Taking a model into a negotiation containing plausible but unsourced process figures the supplier knows precisely
Reporting a narrow confidence range calculated as though correlated input uncertainties were independent

Key takeaways

A should-cost model is only as strong as its weakest line. Every figure needs a source or an explicit assumption label with a range, because one refutable number discredits the sourced lines alongside it.
AI-generated cost models produce plausible values for process parameters such as cycle time and scrap rate that are not grounded in the specific part or supplier. These are precisely the figures the supplier knows exactly.
Energy, freight, and many commodity prices are correlated, so models treating input uncertainties as independent understate the true range — and do so most in volatile markets, when the model is most needed. Stress the correlated set together.
Present should-cost output as a range with the dominant driver named. A point estimate invites a challenge on precision that the model cannot survive.
A benchmark gap is a hypothesis, not a finding. Volume, specification, incoterms, payment terms, service scope, and contract duration explain most price differences, and establishing that you pay more for a structural reason often identifies a demand-management opportunity rather than a negotiation one.