How Engineers Select Components With AI in 2026

What the workflow looks like now

Ask a design engineer whether AI has changed their job and you get a careful answer. Ask whether they used an assistant this week and the answer is almost always yes.

Weidmuller and EETech Research surveyed 400 North American engineers between April and May 2026 on how they design boards and evaluate components. More than nine in ten had used AI-based tools somewhere in their PCB workflow. But 75% described those tools as a productivity and acceleration layer inside existing processes — faster iteration, not delegated judgement.

That distinction is the whole story. AI has achieved near-universal adoption and near-zero authority. Engineers use it constantly and trust it conditionally, and the condition is verification.

Where does AI actually enter the selection process?

Note what dominates the third column. Very little of it is the model being unintelligent. Most of it is the model being handed a PDF.

Why the format matters more than the model

The strongest evidence on this arrived in June 2026, when researchers published PCB-QA: a dataset of 480 question-answer pairs derived from real KiCad design files, covering component connections, datasheet enquiries and SPICE simulation behaviour, evaluated against four frontier models.

The headline is not which model won. It is what happened when the same model was given the same information in different formats:

The best model on the JSON representation, Gemini 3 Flash Preview, reached 93%.

Read that as a supply-side result rather than a model result. Doubling accuracy did not require a better model, a bigger context window or a fine-tune. It required giving the machine the same facts in a form it could parse. Every manufacturer whose parametrics live only inside a PDF datasheet is running its parts at something close to the 43% row.

This is also why "the models will get better so we can wait" is a weaker argument than it sounds. Model improvement does not recover information that the format destroyed.

Why do engineers still open the datasheet?

Because of what a datasheet is, not what it contains.

A datasheet is a controlled, revision-managed document that a manufacturer stands behind. When a part fails qualification, the datasheet is the artefact in the argument. No engineer is going to substitute a chat transcript for that, and no manufacturer should want them to.

The failure is in expecting a machine to derive the parametrics from it. A PDF is a rendering of your data, not your data — and the specific things that get destroyed in rendering are exactly the things that cause respins: the footnote that the 2 A rating applies at 25°C, the asterisk indicating a preliminary specification, the note that the part is not recommended for new designs.

There are three distinct failure modes here, and conflating them leads teams to the wrong fix:

1. 01Staleness. The model recites a figure from training data that has since changed. Fixed by being reachable live, not by better prompting.
2. 02Extraction failure. The value exists but only inside a PDF table. Fixed by publishing structured parametrics.
3. 03Absence. No machine-readable record exists at all, so the model interpolates from neighbouring parts. Fixed by coverage.

Only the third is hallucination in the strict sense. The first two are things the manufacturer is doing to itself.

Why the errors compound at bill-of-materials scale

A single wrong parameter is a nuisance. The same error rate applied across a bill of materials is a different class of problem, and this is where AI-assisted selection stops resembling a chat and starts resembling a supply chain.

Consider a modest 200-line BOM. An assistant that is 95% accurate per line — better than most benchmarks would predict from PDF-only sources — produces around ten defective lines. Some will be caught at review. Some will be caught at first article. And some will be caught when a distributor confirms that the part went not-recommended-for-new-designs eighteen months ago.

Three properties make this worse than a naive error-rate calculation suggests:

• The errors are not random. They cluster on exactly the parts with the poorest published data, which correlates with older catalog entries, second-tier manufacturers and recently acquired product lines — the same population most likely to have lifecycle risk.
• They are silent. A missing parameter produces a confident answer built on the parameters that were present. Nothing in the output distinguishes a value that was retrieved from one that was inferred.
• They survive review. Engineers verify the parameters they consider critical. Lifecycle status, minimum order quantity and regional certification are frequently not on that list at schematic stage, and are precisely the fields most often absent from machine-readable records.

The Weidmuller data sharpens the timing problem: 43% of engineers finalise component selections during early concept or schematic design, with a further 26% selecting iteratively across multiple stages. Selection is happening early, when the assistant's answer is least likely to be cross-checked against a distributor's live data, and when the cost of a later change is beginning its climb.

Which tools are engineers actually using?

A domain layer has formed above the general assistants, and it is worth knowing because these tools are now the intermediaries between your catalog and your customer's decision.

• Flux raised $37m in February 2026 — a $27m Series B led by 8VC plus a previously undisclosed $10m Series A — for an AI copilot that plans layouts, sources components and tests designs from natural-language prompts. Its 2026 releases added sourcing-aware design with real-time pricing and availability.
• Luminovo, in Munich, launched ElectronicsGPT in January 2026: an agent for electronics procurement that retrieves parts with their technical data, alternates and supply-chain information.
• CELUS, also Munich, generates schematics from requirements against its CUBO component knowledge base, and demonstrated AI-powered component matching at embedded world 2026.
• Cofactr raised a $17m Series A led by Bain Capital Ventures for sourcing constrained by supplier criteria, regulation and internal policy — aerospace, defence, robotics and medtech.
• Parspec, in building products, raised a $20m Series A in July 2025; its models extract product requirements from drawings and specifications and rank compliant products from a catalog of more than six million items, generating quotations, submittals and O&M packages.

The common structure is telling. Every one of these tools is an interpretation layer over manufacturer data it does not own. Their accuracy ceiling is your data quality. When they cannot resolve your part, they resolve a competitor's — and the engineer never learns which parts were silently excluded.

What must a manufacturer publish to be trusted in this loop?

In rough order of payoff:

1. 01Server-rendered catalog pages that a non-browser client can read. When Partsgraph audited 984 distributor and manufacturer domains worldwide in August 2026, 38% returned nothing usable to a standard non-browser client, and the median AI-visibility score was 50 out of 100. This is the cheapest fix and the most common failure.
2. 02Parametrics as structured fields with units, tolerances and test conditions attached — not as a table image, not as prose, not only inside the PDF.
3. 03Lifecycle status as data. Active, NRND, last-time-buy date, obsolescence date. Assistants confidently recommend discontinued parts because nothing told them otherwise.
4. 04Compliance and certification as fields. RoHS, REACH SVHC, UKCA and CE status, AEC-Q or equivalent qualification, country of origin. In building products: fire classification, third-party certification, declared performance.
5. 05Explicit alternate and equivalence relationships. If you do not publish what your parts replace and what replaces them, an assistant will guess, and its guess will be graded on plausibility rather than on qualification.
6. 06A live route for the questions that change. Stock, lead time and price cannot be served from a crawl. Microchip's Model Context Protocol server, published in November 2025, exposes verified product specifications, datasheets, inventory, pricing and lead times directly to AI clients. It is still a very short list of manufacturers doing this — in the same audit of 984 domains, none advertised an MCP endpoint, while third parties have built unofficial MCP wrappers around major distributors' public APIs to fill the gap. Someone is going to expose your catalog to agents. It may as well be you.

The honest limit

None of this makes an assistant a competent engineer. The Weidmuller survey found engineers naming real-time error detection and genuinely autonomous design as the missing capabilities, and roughly 80% ranking reliability and durability as the most important criterion in component selection against only 22% prioritising price. That value structure resists automation, because the cost of being wrong is asymmetric and lands years later.

But that is an argument about who approves, not about who proposes. Proposal has already moved. The engineer who once opened four distributor sites now opens one assistant, and the parts that appear in that answer are the parts under consideration. If your data cannot support a correct answer, you are not being rejected — you are not being enumerated.

The free grader at [/audit](/audit) shows what an assistant currently reads from your catalog, and which of your parts it gets wrong.