Appendix
Appendix: Procurement and Sourcing
An optional appendix on procurement finance. Category management and the purchasing portfolio; the should-cost model as an independent estimate that changes a negotiation; evaluating competing vendor bids on best value rather than lowest price; and where AI helps total, rank, and narrate while the buyer owns the strategy and the negotiation.
~110 min6 sections16 questions5 tools
Learning objectives (8)
Learning Objectives
By the end of this chapter you should be able to:
- 1Recap category management, the purchasing portfolio (the Kraljic matrix), the should-cost model as an independent bottom-up estimate, and best-value (not lowest-price) award, before any AI is introduced.
- 2Design the sourcing workflow so the deterministic math (bid totals, the gap to should-cost, and annualized savings) sits in a template, and the model performs only the non-deterministic recommendation narrative.
- 3Assemble the minimal context folder that briefs a model for a sourcing event: the vendor bids, the should-cost buildup, the brief, and one worked example of the destination memo.
- 4Run the red-lines check for commercially sensitive bid data before starting, and keep the AI-assisted recommendation inside the normal procurement review chain.
- 5Validate a drafted recommendation by tying each number to a bid or the should-cost, confirming each bid total foots, and confirming savings are stated against a clear baseline.
- 6Recognize the three common failure modes of an AI sourcing recommendation (a lowest-price-only pick, an invented external benchmark, and a total that does not foot) and correct them.
- 7Reach a best-value judgment that weighs the should-cost gap and the cost elements rather than price alone, and name the negotiation levers the should-cost reveals.
- 8Frame the sourcing recommendation in a category manager's voice, leading with the best-value pick, the annualized savings, and the levers and open questions.
Part One: Category Management, the Should-Cost Model, and Best-Value Award. Section 1 of 6.
Part One · Category Management, the Should-Cost Model, and Best-Value Award
Category Management, the Should-Cost Model, and Best-Value Award
Part One
Category Management, the Should-Cost Model, and Best-Value Award
Cost takeout in procurement runs on four ideas: category management, a purchasing portfolio that decides how hard to work each category, a should-cost model that sets an independent estimate, and a best-value award rather than the lowest sticker.
The sourcing event and category management
You are in procurement finance at Meridian Components, and a machined component is up for re-sourcing. The naive way to buy is to react to each purchase requisition as it arrives, taking whatever quote is in front of you. Category management, as the Chartered Institute of Procurement & Supply (CIPS) frames it, is the disciplined alternative: treat the spend category strategically and end to end, so you understand the internal demand, the shape of the supply market, the total cost of the part, and where your leverage sits, rather than handling purchases one at a time.
For this component that means a few concrete questions before a single bid arrives. How many suppliers can realistically make the part to spec? How large is the annual spend, so you know how much analysis the category can justify? And what does the part actually cost to produce, independent of what any vendor chooses to quote? Category management turns "get three quotes" into a structured event with a defensible recommendation at the end.
The purchasing portfolio: how hard to work a category
Not every category deserves the same effort, and Peter Kraljic's 1983 Harvard Business Review article, "Purchasing Must Become Supply Management," gives the map that is still in daily use. The purchasing portfolio, often called the Kraljic matrix, segments what you buy along two dimensions: the profit impact of the item (its share of spend and its effect on the business) and the supply risk (how scarce or complex the supply market is). Those two axes create four quadrants, each with a different sourcing posture.
- Non-critical (routine) items have low profit impact and low supply risk. The play is efficiency: simplify and automate the buying so it costs little to transact.
- Leverage items have high profit impact but low supply risk, because several capable suppliers compete. Here the buyer holds the stronger hand, so competitive bidding and a should-cost analysis tend to pay off.
- Bottleneck items have low profit impact but high supply risk. The priority is continuity of supply, so you secure the source and reduce dependence.
- Strategic items have both high profit impact and high supply risk, so the posture is a deeper partnership with a limited set of suppliers.
The machined component in this event is a classic leverage item: it is a meaningful share of annual spend, and several vendors can make it to spec. That is exactly the setting where an independent should-cost estimate and a competitive bid earn their keep.
The should-cost model: your independent yardstick
A vendor's quote tells you what the vendor wants to charge. It does not tell you what the part should cost. A should-cost model, sometimes called cost breakdown analysis, is the buyer's own bottom-up estimate, built from the ground up: the material content at current prices, the labor to machine and finish it, a reasonable overhead allocation, and a fair supplier margin. CIPS treats this kind of cost buildup as a core sourcing capability, because it changes the conversation.
The should-cost is not there to name a single "right" price and refuse anything above it; it gives you an independent anchor. When a bid sits well above the should-cost, the buildup shows you which element is out of line, an inflated overhead or a rich margin, so the negotiation has a specific lever rather than a vague "can you sharpen your pencil." A should-cost is an estimate, so it is only as good as its input assumptions, which is why the assumptions travel with the number.
Best value, not lowest price
The award decision is the finish line, and the standard is best value, not lowest quoted price. Best value weighs the should-cost gap and the cost elements alongside the things a unit price hides: quality and defect history, on-time delivery and lead time, capacity and continuity risk, and the total cost of ownership over the life of the part. A bid that is cheapest on the sticker can still lose if its low price rests on an element that looks unsustainable, or if the vendor's delivery record would cost you more downstream than it saves per unit.
Once the bids are in and the should-cost exists, most of the remaining work is structured: total each bid from its elements, compute the gap to the should-cost element by element, annualize the savings against a stated baseline, and draft a recommendation in a consistent shape. That work is repetitive and rule-bound, the kind of task AI assistance handles well. The judgment, which vendor is the best value, which risks matter, and which lever to raise first, is where the category manager adds value. That split, deterministic arithmetic on one side and judgment plus narrative on the other, is what the rest of this appendix is built around.
Check Your Understanding
Knowledge Check 1
Procurement & Sourcing
A machined component has several capable suppliers and represents a sizable share of annual spend. In Kraljic's purchasing portfolio, which quadrant does it fall into, and what does that imply for the sourcing approach?
Part Two
Where AI Fits: Map It, Split It, Fuel It
With the sourcing work and its best-value standard clear, this part places AI where it earns its keep and keeps it away from where it does not. Then it runs the setup moves as concrete decisions and wraps them in reusable scaffolding.
What AI is good at here, and what it is not
The tool has to fit the task. For a sourcing recommendation a language model is good at a narrow band of work: totaling each bid from its cost elements, computing the gap to the should-cost element by element, laying the comparison out in a consistent table, drafting a best-value recommendation in the house shape, and naming the levers the should-cost points to. That is real leverage on the slow, repetitive part of the event.
It is a poor fit for three other parts, and the workflow keeps it out of them. It should not do the arithmetic in prose, because a model asked to add and multiply in a sentence can drift, so the deterministic math belongs in a template it does not touch. It should not invent an external market benchmark, because a confident "roughly 10% below market" figure the folder did not contain is a real risk, so the analysis reasons only from the bids and the should-cost on hand. And it does not award the business: the model drafts, and procurement reviews, ties out, and approves. In this workflow the model is a fast analyst, never the approver.
Map it: last event's memo as the worked example
The most useful thing you can give a model for this task is an example of the destination. The recommendation memo from the last sourcing event is exactly that: it shows the house style, the level of detail, and how a category manager leads with the best-value pick and the savings before the supporting table. With that memo in the folder, the model works from a known shape instead of guessing at what good looks like. You start with bids and a should-cost, you are going to a defensible best-value recommendation, and the worked example anchors the far end.
Split it: the math is deterministic
Three calculations sit at the center of this event, and each has one correct answer for a given input, so each is deterministic work. A bid total is the sum of its elements: material plus labor plus overhead plus margin. The gap to should-cost is the bid less the should-cost, read both in total and element by element. And the annualized savings is the per-unit gap times the annual volume, stated against a named baseline. Asking a language model to produce those numbers in a sentence invites small, hard-to-catch errors.
So the workflow puts the math where it belongs. A template totals each bid, computes each gap, and annualizes the savings, and the model narrates only the results. The model performs the non-deterministic work, the part with many acceptable phrasings: how to describe the best-value case, how to word each negotiation lever, and how to sequence the memo. When the template owns the math there is one source of truth, and validation becomes a tie-out rather than a recalculation.
One bid, end to end
A single vendor makes the split concrete. Vendor Beta bids $78.50 per unit, the independent should-cost is $75.00, and the annual volume is 120,000 units. Here is how the workflow carries that bid from raw elements to one line of the recommendation, and who owns each step.
- Template, deterministic. Beta's element buildup foots to the $78.50 bid. The gap to should-cost is $78.50 minus $75.00, or $3.50 per unit. Annualized against the highest bid, and against the should-cost, the template does the multiplication so the model does not.
- Compare, judgment plus data. Beta is the lowest of the three bids, and its element buildup sits closest to the should-cost, so the should-cost confirms Beta as the best value rather than the buyer simply assuming the cheapest sticker is best. A low bid whose overhead sat well below the buildup would get a second look instead.
- Lever, judgment. Even the best bid carries a $3.50 gap to should-cost. Reading which element drives that gap is the non-deterministic part, and it is where the negotiation lever comes from.
- Narrate, non-deterministic. The model turns the computed figures and the confirmed lever into one measured line: "Vendor Beta is the recommended best value at $78.50 per unit, $3.50 above the should-cost of $75.00; the gap sits in overhead, which is the lever to test before award."
Fuel it: the minimal folder
The lab folder is deliberately small: the vendor bids broken into cost elements, the should-cost buildup, the brief, and the one worked example. Nothing else. A larger pile, the full supplier master or unrelated categories, would bury the elements that actually differ and raise the chance the model latches onto something irrelevant, and it would widen the exposure of commercially sensitive data. Minimum context is the fuel: give the model exactly what the event needs, plus the destination example so it knows the shape of the answer.
The worked example does work here that written instructions tend to miss. It encodes the tacit house style, the level of detail, and the sequencing a category manager would struggle to spell out as a rule, and the model tends to match a shown example more reliably than it follows an abstract description of one. That is why the folder carries a finished memo rather than a longer style guide.
Scaffolding: a reusable sourcing skill and a best-value checklist
The moves are not a one-off. Because sourcing events recur and share a shape, the prompt that carries these instructions is worth saving as a reusable skill or prompt template: the destination example, the deterministic-math rule, the minimal-folder list, and the best-value guidance, packaged so the next event starts from the same known shape rather than a blank page.
Pair it with a short best-value checklist the analyst runs before sign-off: confirm each bid total foots, tie each figure to a bid or the should-cost, confirm the recommendation weighs the gap and the elements rather than price alone, confirm the savings name their baseline, and confirm no external benchmark has crept in. The skill makes the draft repeatable; the checklist makes the review repeatable. Together they keep the human checkpoint from becoming a rubber stamp.
Check Your Understanding
Knowledge Check 2
AI Workflow Design
In a should-cost sourcing workflow, which task is best placed in a validated template rather than left to the language model?
Part Three
The Pattern
The whole sourcing workflow fits on one diagram. Each step expands to the prompt template, what a good result looks like, and the ways the step tends to fail. This is the shape you will run in the lab.
Reading the pattern
The pattern moves left to right through four kinds of step: the inputs you gather, the AI step that drafts, the human checkpoint where you validate, and the finished artifact. The gray input node holds the minimal folder of bids and the should-cost. At the green AI node, the model totals the bids, computes the gaps, and drafts a best-value recommendation with levers. The amber human node is where you confirm the totals foot and the best-value logic holds beyond price, a step that is not optional. What comes out the far end is the sourcing recommendation itself.
The human checkpoint sits between the draft and the artifact, not after it. Nothing becomes a recommendation you would take into the room until a person has confirmed the arithmetic and the value case. Open each step below to see the prompt and the failure modes before you run it.
Check Your Understanding
Knowledge Check 3
AI Workflow Design
In the sourcing pattern, where does the human checkpoint belong in the sequence of steps?
Part Four
Guard It, Then Run the Lab
Thirty seconds of governance before you open the folder, then the lab itself.
The red-lines check for bid data
Vendor bids are commercially sensitive: they expose supplier pricing, cost structure, and terms that a competitor or the other bidders would value. Before you point any tool at bid data at work, confirm the instance is approved for that data class, keep the folder scoped to this sourcing event rather than the whole supplier master, and make sure the recommendation still flows through your normal procurement review. The lab below uses a fully synthetic company and fictional bids, so its data is cleared for any tool. Running the check anyway is the habit you are practicing.
The lab
Download the folder and run the sourcing pattern in whatever AI you use. The folder holds Meridian's three vendor bids broken into cost elements, the independent should-cost buildup, the brief, and a worked example memo as your style guide. Let a template hold the math (total each bid, compute the gap to should-cost, annualize the savings against a stated baseline), have the model draft a best-value recommendation with the negotiation levers, and then come back for validation. The company and every number are fictional.
Check Your Understanding
Knowledge Check 4
AI Governance
You want an AI tool to draft a sourcing recommendation from live vendor bids. Which step best reflects sound data governance before you begin?
Part Five
Validate the Output
A fluent recommendation is not a finished one. Validation is where an AI-assisted sourcing memo earns the right to go into the room, and where its three characteristic failure modes get caught.
The three failure modes
An AI sourcing recommendation tends to fail in three recognizable ways, and knowing them turns validation from a vague read-through into a targeted search.
The first is the lowest-price-only pick. The model sees three numbers, names the smallest, and calls it done, with no reference to the should-cost gap or the cost elements. Sometimes the cheapest bid is the best value, but a recommendation that reaches that conclusion by looking at price alone has skipped the analysis that would catch a low bid resting on an unsustainable element. Best value is a reasoned call, not the minimum of a list.
The second, and the most dangerous, is the invented external benchmark. The model supplies a confident figure the folder did not contain, "the winning bid is about 10% below the market rate," when the folder holds only three bids and a should-cost. Fluent and specific is not the same as sourced. The third is a total that does not foot or a savings figure with no baseline: an element buildup that does not sum to the quoted bid, or an annualized saving stated with no answer to "saved against what." Each makes careful work look careless and, worse, undermines the number the reader is meant to trust.
Tie-out as the core discipline
The heart of validation is the tie-out: take each figure in the recommendation and follow it back to a bid or the should-cost. Confirm that each bid's elements sum to its quoted total, that the gap to should-cost is computed the same way for every vendor, that any annualized saving names its baseline (for example, versus the highest bid), and that no figure appears that the folder does not support. A single untied number, or one invented benchmark, is enough to send the memo back. Work the checklist below against your draft before you call it done.
Check Your Understanding
Knowledge Check 5
AI Validation
An AI-drafted recommendation picks the cheapest of three bids and states, "this vendor is roughly 12% below the market rate," though the folder contains only the three bids and the should-cost. Which failure mode does the quoted claim show?
Part Six
Debrief: A Best-Value Recommendation
A finished, defensible recommendation for the Meridian machined-component event appears below, annotated to show why each choice was made. Compare it against your own draft, then score your work.
The three bids and the should-cost
Three vendors bid the machined component, and Meridian built an independent should-cost of $75.00 per unit. The annual volume for the part is 120,000 units. Read the bids against the should-cost rather than only against each other.
- Vendor Alpha: $82.00 per unit. Gap to should-cost of $7.00 per unit.
- Vendor Beta: $78.50 per unit. Gap to should-cost of $3.50 per unit, the smallest of the three, and its element buildup tracks the should-cost most closely.
- Vendor Gamma: $84.50 per unit. Gap to should-cost of $9.50 per unit, the highest bid in the event.
Vendor Beta is the recommended best value. In this event the best value and the lowest bid coincide, but the should-cost is what lets you say so with confidence: Beta's elements sit closest to an independent buildup, so its low price rests on a defensible cost structure rather than on an element cut below what the part can be made for.
The recommendation exemplar
The exemplar below is what you would take into the award review. It leads with the best-value pick, states the savings against a named baseline, and points to the lever the should-cost reveals.
Recommendation. Award to Vendor Beta at $78.50 per unit. Beta is the best value of the three bids: it is the lowest price, and its element buildup sits closest to the independent should-cost of $75.00, so the price rests on a sound cost structure rather than an unsustainable cut.
Savings. Against the highest bid, Vendor Gamma at $84.50, awarding Beta saves ($84.50 minus $78.50) times 120,000 units, or $720,000 a year. Beta still sits $3.50 above the should-cost, which annualizes to ($78.50 minus $75.00) times 120,000, or $420,000 a year of remaining gap to the independent estimate.
Negotiation lever. The $3.50 per-unit gap to should-cost sits mainly in overhead. Before award, test that element against the buildup; closing even part of the gap on 120,000 units is worth pursuing, and the should-cost gives a specific, defensible number to anchor the ask.
Open questions. Confirm Beta's quoted lead time and quality record support the volume, since a best-value award weighs delivery and continuity alongside the price.
The recommendation names best value rather than defaulting to lowest price, states each saving against a stated baseline (the highest bid, and the should-cost), points the lever at a specific element, and leaves the non-price risks as open questions rather than burying them.
Where this breaks in the real world
The lab is clean by design: three bids, one should-cost, a stated volume, and elements that foot. Real sourcing events are messier, and the workflow tends to strain in three specific places.
- Should-cost quality. The should-cost is only as good as its input assumptions. A stale material index or an off labor rate can make a fair bid look inflated, or make an aggressive bid look reasonable, so the buildup assumptions travel with the number and get a second look before they anchor a negotiation.
- Non-price factors and total cost of ownership. The lowest gap to should-cost can still lose the award if a vendor's quality history, on-time delivery, or tooling and switching costs would cost more downstream than the per-unit saving. Best value weighs the total cost of ownership, not the sticker alone.
- Supply-risk and single-source drift. Awarding all volume to one vendor for a modest per-unit saving can raise supply risk, nudging a leverage item toward the bottleneck quadrant. A small premium to keep a qualified second source is sometimes the better value once continuity is priced in.
On top of these, the model will sometimes produce a recommendation that is fluent, specific, and wrong, which is exactly why the tie-out and the best-value check are not optional. The workflow does not remove the category manager's judgment; it removes the mechanical drudgery so the judgment has more room.
Score your work
Rate your own recommendation against the rubric below. Marked honestly against the rubric, the score shows where the sourcing workflow is strong and where it still needs work. Your scores roll up to the workflow maturity dashboard on the course hub.
Check Your Understanding
Knowledge Check 6
Best-Value Analysis
A recommended vendor's price is $4.25 per unit below the highest bid, and annual volume is 96,000 units. What is the annualized savings versus the highest bid?