AI in Strategic Sourcing: RFPs, Bid Analysis and Should-Cost

By Tara Minh · Updated Oct 11, 2026

AI in strategic sourcing is the use of language models and optimization or prediction models to run the analytical parts of a sourcing event: drafting the request for proposal (RFP), turning messy supplier bids into comparable numbers, estimating what a product or service should cost, and testing award scenarios before a person decides. The buyer still chooses suppliers and signs the contract. What changes is how much time goes into reading, reformatting and spreadsheet work, and how many award options get examined before the decision.

This article covers the sourcing event itself. For the wider picture, start with the AI in procurement overview, and if the category strategy behind the event is still fuzzy, read strategic sourcing for manufacturing first. AI speeds up a sound sourcing process. It doesn't repair a weak one.

Key Facts

• Sears Logistics Services was the first buyer of trucking services to use a combined-value (package-bid) auction, and the paper on it reports savings of 13 percent in 1993 (Ledyard et al., Interfaces 2002).
• A published Procter & Gamble case describes sourcing more than $3 billion over two and a half years with an optimization engine, with recommended savings of $294.8 million, or 9.6 percent (Interfaces 2006). These are the authors' reported figures for one company and one method, not an industry average.
• US federal rules define a should-cost review as a form of cost analysis that does not accept a supplier's past costs as proof of efficiency and instead examines how the work could be done for less (FAR 15.407-4).
• Competitive federal solicitations must state the evaluation factors and their relative importance (FAR 15.203). It's a good habit for any buyer, because a model can only score bids against criteria you wrote down.
• Both published savings cases above predate modern language models and rest on optimization over real bids. Language models add the reading, extraction and drafting around that core.

Where AI Fits in a Sourcing Event

A sourcing event has five stages, and AI helps differently in each.

The pattern is the same as in most operations work: the model prepares, the person decides. For negotiation specifically, see AI supplier negotiation.

Drafting RFPs With AI

A good RFP is mostly reuse. Most categories have been bought before, so past RFPs, contracts and specifications hold most of what you need. A language model is well suited to this: give it the previous RFP, the new volumes and the changes, and it produces a structured draft in minutes.

What to use it for:

• First drafts of requirement sections from existing specifications and contract terms.
• Gap checks. Ask the model what a typical RFP for this category covers (service levels, sustainability, data security, business continuity, exit terms) and compare against your draft.
• Question quality. Convert vague questions ("Describe your quality approach") into specific, scoreable ones ("State your defect rate in parts per million for the last 12 months and how it is measured").
• A consistent pricing template with separate lines for unit price, freight, tooling, minimum order quantities, price-adjustment clauses and payment terms. Separate lines are what make bids comparable later.
• Plain-language summaries of the RFP for internal stakeholders who need to sign off.

What not to delegate: the evaluation criteria and their weights. Decide these before suppliers respond, write them into the RFP, and keep them fixed. If the weights change after bids arrive, the result looks rigged to suppliers and to your own auditors. Also have legal review any clause the model drafted or altered.

A practical trick: have the model write the RFP, then ask a second pass to answer it as a supplier would. Ambiguities show up fast, because the answers diverge.

Comparing Bids: Normalization Comes First

Bids rarely arrive in the same shape. One supplier quotes ex-works, another delivered. One prices per unit at your forecast volume, another gives a table of break points. One quotes a three-year fixed price, another allows annual adjustment tied to an index. Comparing the headline numbers is how teams pick the "cheapest" bid that turns out to cost more.

Normalization is the unglamorous core of bid analysis, and it's where AI saves the most hours.

1. Extract. A language model reads each response (PDF, spreadsheet, email) and fills a standard template: unit price, volume tiers, incoterms, lead time, payment terms, validity period, exceptions to your terms.
2. Flag. It lists missing items, inconsistent figures (a total that doesn't match the line items) and exceptions to your terms.
3. Convert. Normalize everything to the same basis: same currency, same delivery point, same volume, same period. Landed cost is usually the right basis.
4. Check by hand. A buyer verifies the extracted figures for the finalists against the source documents. Extraction errors are rare but not zero, and a single wrong digit can reverse a ranking.

Illustrative normalization example

Three suppliers quote the same part for an annual volume of 10,000 units. All figures below are made up to show the method.

Supplier B had the lowest headline price and ends up the most expensive once freight and tooling are added. Payment terms would shift the picture further if you put a cost on the 30-day gap. A model builds this grid in seconds across 40 suppliers and 200 lines, which is the point: nobody has to retype it.

Scenario-Based Award and Sourcing Optimization

Once bids are comparable, the next question is who gets what. Awarding every line to the lowest bidder looks simple, but real awards carry constraints: no supplier above 60 percent of volume, at least two suppliers per site, minimum order quantities, capacity limits, incumbent preferences, diversity or local-supplier goals, and discounts that only apply if you buy a bundle.

This is a mathematical optimization problem, and it predates modern AI. The published record is useful here because it shows the size of the effect with real money:

• Sears Logistics Services used package bids for trucking lanes, letting carriers offer a single price for a bundle of lanes that fit their network, and reported 13 percent savings in 1993 (Ledyard et al.).
• Procter & Gamble described an approach in which suppliers submit offers that express their capabilities, P&G states its constraints and preferences, and an optimization engine allocates the business. The authors report more than $3 billion sourced over two and a half years and recommended savings of $294.8 million, or 9.6 percent (Interfaces 2006).

Notice what these have in common. Suppliers could say what they do best, and the buyer could state its rules, then software searched the combinations a person never could. Today's tools put a language model in front of this: you describe constraints in plain words ("cap any single supplier at 50 percent, keep the incumbent on at least 20 percent of the East region"), the model turns them into formal rules, and the optimizer runs scenarios.

A scenario table is the output worth looking at:

The numbers differ every time. What matters is the price of each constraint. If "keep the incumbent at 20 percent" costs 1.5 percent more, leadership can decide whether that's worth it, with the trade-off in front of them rather than hidden in a spreadsheet.

Should-Cost Modeling

A should-cost model estimates what a part or service ought to cost if the supplier were efficient, built up from its parts: materials, labor, machine time, overhead, scrap, logistics and a fair margin. The idea is old and formal. US federal acquisition rules describe a should-cost review as an examination of a contractor's workforce, methods, facilities and management to find ways to lower cost, rather than accepting past costs as proof of efficiency (FAR 15.407-4).

AI helps in three ways:

• Reading drawings and specifications to pick out materials, dimensions, tolerances and processes, so the cost build starts from structured inputs.
• Estimating cost drivers such as cycle time or scrap rate from similar past parts, using prediction models trained on your own purchase history and cost data.
• Explaining the gap between a quote and the model, line by line, so your negotiation focuses on the rows that differ most.

Illustrative should-cost breakdown

The figures below are invented to show the structure of a should-cost build for a machined bracket.

The value isn't the $7.81. It's the conversation it enables: "Your scrap allowance is more than twice what we model for this process. What yield are you running?" A should-cost estimate is a hypothesis for the negotiating table, not a verdict. Suppliers know things your model doesn't, such as a machine in a different state or a tool they already own.

Treat it as weakest where data is thin: custom engineered parts, new materials, services priced on judgment rather than hours. And since commodity inputs move, tie raw-material lines to a published index instead of a one-time number.

How to Implement AI in Strategic Sourcing, Step by Step

1. Pick one category and one event. A repeat category with 10 to 40 suppliers and a clear spend baseline is ideal. Avoid your most strategic, most politically charged category as the first test.
2. Write the criteria first. Evaluation factors, weights, must-haves and constraints go on paper before any model is involved.
3. Standardize your templates. One RFP structure and one pricing sheet make extraction and comparison far easier. This is half the benefit.
4. Use AI for the first RFP draft. Feed it past RFPs and specifications. Have category, legal and quality review it, as they would any draft.
5. Run extraction on live responses, with a spot-check rule. For example, a person verifies every finalist and a random sample of the rest against source documents.
6. Normalize to landed cost. Agree the basis with finance beforehand.
7. Run award scenarios. Start with three: lowest price, policy-compliant, and risk-weighted. Record the cost of each constraint.
8. Build should-cost for the top spend lines only. Pick the items where a 5 percent gap would matter, and compare model to quotes.
9. Decide, document, and keep the audit trail. Record which inputs the model used, who checked what, and why the final award differs from the optimizer's if it does.
10. Measure. Compare the awarded cost against the baseline and against the median bid, and track how many hours the event took versus the last one. Feed outcomes into supplier performance scorecards so you learn whether the winning bids delivered.

Examples by Function

These are illustrative scenarios, not case studies.

Manufacturing procurement. A buyer reruns a components RFP. The model drafts the document from last year's, extracts nine supplier responses, normalizes freight and tooling, and flags one supplier whose total doesn't match its line items. The buyer spends the saved time on supplier visits.

Logistics and transportation. A logistics team bids out freight lanes. Carriers offer bundle prices for lanes that fit their network, and an optimizer allocates lanes under capacity and service rules, the same idea as the Sears case above.

Facilities and indirect spend. Cleaning, security and maintenance contracts come back as long narrative proposals. A model summarizes scope, exclusions and service levels in a common format so the evaluation panel compares like with like.

IT and software. Responses to a software RFP differ in how they package modules and users. The model builds a requirements coverage matrix and flags where an answer is "roadmap" rather than "available today".

Finance. Finance uses the scenario table to see what each policy constraint costs, then signs off on the trade-offs rather than a single number.

Marketing and professional services. Agency pitches are scored against a rubric agreed in advance, with the model checking that each pitch answered every required question.

Pitfalls to Avoid

Changing the criteria after bids arrive. It undermines trust and can create legal exposure. Fix weights before you send the RFP.

Trusting extraction blindly. Models misread tables, skip footnotes and confuse units. Verify the numbers behind any award.

Comparing headline prices. Without normalization to landed cost, the lowest quote is often not the lowest cost.

Treating should-cost as truth. It's a negotiation aid. Share it carelessly or defend it rigidly and you'll lose supplier goodwill.

Optimizing only for price. An award that ignores quality, delivery performance, capacity and concentration risk will cost more later. Put those on the scorecard and in the constraints.

Sending confidential data to the wrong tool. Bids contain prices and terms that suppliers expect to stay private. Use tools under your company's data agreement, and never feed one supplier's bid into a workflow that could expose it to another.

Letting the optimizer's answer be the decision. The solver returns what the constraints say, nothing more. Review it for relationships, transition effort and things nobody modeled.

No audit trail. If you can't show who decided what and on what basis, you can't defend the award to a losing bidder or an auditor.

What to Do Next

Choose one upcoming sourcing event and run it two ways in parallel: your usual process, and an AI-assisted version of the reading, normalization and scenario steps. Compare hours spent, errors caught, and the number of award options examined. Keep what saved time without adding risk, and scale it to the next category. If a negotiation follows, the AI supplier negotiation guide covers that step. If you want the procurement function around all this, see procurement management.