AI for Procurement in Semiconductors: Buying Against a Process Window

A buyer arriving in a semiconductor business from almost any other sector brings one assumption that has to go early: that a specification is a starting position. It is not. In a fab, an assembly house or a fabless company buying wafers and substrates, the specification was written by the person who owns the process step, and it exists because a measured result depends on it. The buyer does not get to widen a purity limit to open up the bid list. The specification is the contract, and everything commercial happens inside it.

Three things follow, and they shape the whole desk. A substitution is a process change, so it carries engineering work and fab time rather than a vendor form. Lead times on the equipment side outrun the forecasts that justify the spend. And an unqualified alternative is not an alternative, so the second supplier on your approved list may be no help on the day you need one.

Agents are useful across this in a specific register: reading documents, tracing an item through your own data, comparing what a supplier sent against what you approved, and chasing the gaps. They do not approve materials and they do not sign capital.

The tool slot closes before the capital request opens

Long lead capital equipment is the clearest case of the timing mismatch. Lead time is counted in quarters, slots are held in a supplier's build schedule, and the configuration is frozen at order. Meanwhile the capital approval chain runs through the process owner, the fab manager, finance, and often a group committee that meets on its own calendar. The slot closes before the capital request opens, so the business is regularly deciding whether to reserve a position against a number nobody has signed yet.

The configuration is the part number

A tool quote is not a price for a thing. It is a price for a configuration: chamber count, source, handling options, metrology integration, software modules, facilities requirements, installation and start-up scope, warranty terms, and a spares and consumables package that is easy to under-order. Quotes revise several times during the approval cycle, and the version that goes to committee is often not the version that was technically agreed.

An agent can hold this honestly. Given the quote revisions, it can produce a difference list between the current quote and the configuration the process owner last signed off, flag options that were dropped to make a budget fit, pull the facilities requirements out into the site readiness list, and keep the capital request packet assembled as the inputs change. It can also watch the obligations that come with a reserved slot: progress milestones, documents due to the supplier, site readiness dates, and the point after which the configuration cannot move.

Forecasts that turn into commitments the moment you submit them

Wafer and assembly supply arrangements run on rolling forecast submissions, and the forecast is not an opinion once it is in. The near window is firm, the middle window flexes within a stated band, and the far window is planning information the supplier loads against. Submissions go out on a cycle, usually built from a demand plan that moved after it was approved.

The administrative weight here is reconciliation. What did we submit last cycle, what changed, which lines moved outside the flex band, what is the supplier now loaded to, what have we acknowledged, and where does our committed volume sit against what the business will actually take. An agent can do all of that mechanically: build the submission, diff it against the previous one, mark the lines that breach a window, and keep a running exposure view with the reasons attached. It gives the planner and the buyer an argument to take into the supplier call rather than a spreadsheet to rebuild first.

Gases, chemicals and substrates are bought against a process, not a catalogue

Specialty gases, wet chemicals, photoresists, polishing slurries, sputtering targets and substrates are ordered to a document that reads like a test plan: purity grade, trace metal limits, particle counts, moisture, viscosity, thickness and flatness tolerances, container type, shelf life, and the test methods used to prove each of them. Approval is frequently narrower than the supplier's name suggests. It is this grade, from this fill or finishing plant, within this lot age.

So the recurring work is comparison. A supplier returns a data sheet or a certificate of analysis template, and someone has to read it against the incoming specification line by line: a limit stated wider than ours, a method that is not the one we approved, an absent parameter, a plant that is not the plant on the approved list. That is exactly the work an agent does well and humans do unevenly, because it rewards never getting bored. The same comparison can be run as a gate on requisitions, so an order for an unapproved grade or plant is caught before it is placed rather than at goods receipt.

An unqualified alternative is not an alternative

Technical qualification in a fab is narrow by design. A material is qualified for a process step, on a tool or tool set, often for a product family, against electrical and reliability results from split lots. It is not qualified for the company. An approved vendor list can therefore be both accurate and misleading: the alternate source is real, but only on one layer, one tool and one product, and the part you need it for is none of those.

This changes the shape of a savings case. The cost of moving to a cheaper material is mostly not the material. It is engineering time, metrology, fab capacity consumed by split lots, reliability stress duration, and in some cases a notification obligation to a customer who qualified your product as it stands. Procurement cannot responsibly book a saving without that cost in the number.

Agents help by making qualification state legible. They can hold the plan status for each candidate, record which lots and measurements exist and which sign-offs are outstanding, keep the approved materials matrix current against the actual sign-offs rather than against memory, and answer the question the desk asks under pressure: for this step, on this tool, what else is actually usable today.

Notifications that force a decision you did not schedule

Suppliers send letters. A process change notice, a site or fill plant transfer, a subcontractor change, a data sheet revision, an end of life announcement with a last order date attached. Each one arrives in an inbox as an attachment and forces a decision that was not in anybody's plan: requalify, buy a final quantity, redesign, qualify an alternative, or accept.

The hard part is not the decision, it is the tracing that has to precede it. Which internal part numbers does this affect, which assemblies and products contain them, which process steps consume the material, which customers were told about the current configuration, and what response window does the agreement give us. Done by hand, each letter costs an afternoon of lookups, which is why the pile grows.

An agent can take the letter, classify what is changing, trace it through the item master, bills of material and approved materials list, and return a scoped impact summary with references back to the records it used. It can track the response deadline, draft the acknowledgement, and put the demand horizon question to planning in the form it needs to be answered. The judgement stays with engineering and the business. The lookups stop being a reason the judgement is late.

Sole source is a design decision you inherited

Most single source positions are not procurement's choice. They were created when a designer selected a component, or when a process owner wrote a specification that only one supplier's material meets, and procurement discovers the exposure when supply tightens or a notice arrives. By then the options are narrow and slow.

Mapping that exposure is a data exercise on information you already hold. Which purchased items have exactly one approved source, which of those have an alternate listed but no qualification evidence behind it, which of those sit in the products carrying the most revenue, and which have a supplier who has already changed something once. An agent can build that map from purchase history, approved source lists and qualification records, keep it current, and screen candidate alternatives against the specification before an engineer spends time on them. Sequencing the qualification programme, and deciding which exposures are worth closing, is a human call informed by the map.

When the supplier has less than everyone asked for

Allocation changes what a buyer's job is. Price stops being the lever, and evidence becomes the lever. A supplier distributing scarce output looks at historical offtake, forecast credibility, and whether your order book has been clean or full of dates nobody meant. Teams that can show consumption and honour their commitments do better than teams that shout.

That evidence is retrievable rather than creative. An agent can assemble the offtake history by part and site, reconcile the open order book against supplier acknowledgements, surface the acknowledged dates that drifted without anyone noticing, and maintain a daily view of which products are exposed and in what sequence they run dry. It can draft the escalation with the numbers in it. The relationship, and the decision about which product line gets starved, are not retrieval problems.

Origin and end use evidence collected at the quote, not at the crate

Export control in this sector reaches sourcing, not just shipping. Tools, parts, substrates and design services carry classifications, country of origin and manufacture declarations, end use and end user statements, and conditions attached to licences already held. Collecting that at the point of shipment is how shipments stop. Collecting it during sourcing, as part of the supplier response, is cheaper and quieter.

An agent can attach the right questions to the sourcing event based on what is being bought, chase the declarations, read what comes back, file it against the supplier and the item, and flag the gaps and the declarations that have gone stale. Classification and licence determination remain with the compliance owner. What the agent removes is the chase and the filing, which is where the omissions come from.

Where AI fails here

Be direct about the limits, because the failures are expensive ones.

Qualification approval belongs to a named engineer. An agent can show that every required result exists and is consistent. It cannot declare a material or a tool qualified for a process step, and nothing it produces should read as if it has.

Capital commitment is a signature, not an output. An agent can assemble the request, trace the configuration and flag what changed. Reserving a slot and committing the money stays with the people accountable for it.

Export classification is advisory, always. An agent can collect declarations and point out that a classification looks inconsistent with the item description. Determining the classification and the licence position is a compliance decision with personal consequences attached.

A trace is only as good as the bill of materials. Impact analysis depends on the item master, the approved source list and the product structure being right. Where those are stale, the output will be confident and wrong, and it will be wrong in a way that reads convincingly. Expect the first weeks of any deployment to be a data quality discovery exercise.

Supplier promises are commercial statements, not facts. A quoted lead time or a committed capacity figure is a position. Agents will treat it as data unless you tell them how it has historically behaved, and your own receipt history is the correction.

Writing into controlled records changes the compliance picture. An agent that reads systems and drafts for humans is a light proposition. One that writes into records your customers or auditors rely on pulls itself inside your change control and qualification obligations. Decide which side of that line you are on before you build, not after.

Take the change notification backlog

Start with the supplier letters. The reasons are practical rather than strategic: the inputs are already sitting in a mailbox, the work is genuinely unowned, nothing in it requires writing to a controlled record, and there is an existing pile you can measure against. Run the agent over notices you have already dispositioned and compare its impact traces against what the team concluded. Disagreements are the useful output, because most of them will turn out to be a data problem you did not know you had.

After that, the approved materials matrix, for the same reason it is worth fixing anywhere: the value is simply that it stops being wrong. Then requisition screening against it, which is where the saved rework shows up in receipts rather than in a report.

If it would help to talk through where the boundary between drafting and deciding should sit for your own buying desk, write to hello@adoptaipartners.com.