You Wouldn't Ship Before First Article. Don't Let an Agent Answer an AOG Call Until It Passes Its Own.

You tried this in 2024 and it embarrassed you. Fair enough.

A lot of aerospace suppliers put a chat widget or a phone answering tool in front of customers two years ago and pulled it back within a quarter. It told a caller a part was in stock when the quantity in the ERP was allocated to a customer-owned consignment bin. It gave a lead time that no planner had agreed to. Worst case, it answered a question about a part number on the United States Munitions List for a caller nobody had screened, and your export compliance officer found out from the transcript.

So the instinct — do not let this thing near customers — was correct in 2024. What changed in 2026 is not that the agents got polite. It is that the tooling for proving one is right before you widen its authority finally grew up. You can now run an agent against your own real cases, watch every step it took, and open its authority one notch at a time based on measured results rather than on how the demonstration felt.

Every shop in this trade already knows this discipline under a different name. You do not ship a production lot because the setup looked good; you ship it because the first article passed and the data is in the package. An agent deserves exactly the same treatment: a defined test set, a recorded result, and a signature before it goes to production.

The calls that actually come in, and which ones are dangerous

Take a spares distributor or a Part 145 repair station with an inside sales desk of three people. A normal day's inbound: do you have this part number in stock; is it traceable with an FAA Form 8130-3 or only a certificate of conformance; what is the lead time on a repair; can you overhaul to the latest service bulletin; is that a new-surplus or overhauled unit; what is your CAGE code; can you take a DO-rated order and what does that do to my delivery.

Sprinkled through those are the calls that can genuinely hurt you. An aircraft-on-ground request at 6:40 p.m. that is worth thousands if you answer it and worth nothing if it goes to voicemail. A caller with an overseas email address asking for the specification detail on a defense part. A request to confirm a part is "certified to Rev K" when your certificate says Rev H. A broker fishing for stock levels. Those are not edge cases. Those are Tuesday.

So the question for the owner is not "can an agent answer the phone." It obviously can — end-to-end voice now answers in roughly two-tenths of a second and can look something up mid-sentence. The question is which of those calls it is allowed to answer, and what evidence you demand before you widen the list.

flowchart TD
  A["Pull 250 recorded stock-check and AOG calls from last year"] --> B["Inside sales manager marks the right answer on each"]
  B --> C["Run the agent against all 250 with no customer on the line"]
  C --> D{"100% correct on the 20 export-screening calls?"}
  D -->|No| E["Narrow what the agent may say, run it again"]
  E --> C
  D -->|Yes| F["Tier 1 live: after-hours answering and message capture"]
  F --> G{"30 days: zero export slips, stock status 95% right?"}
  G -->|No| E
  G -->|Yes| H["Tier 2: quote stock and lead time, commercial part numbers only"]

Building the test set out of your own history

This is the work, and it is not glamorous. Pull 250 real inbound contacts from the last twelve months — recorded calls if you have them, otherwise the sales inbox and the chat log. Your inside sales manager writes the correct outcome next to each one: what the right answer was, what should have been asked, what the caller should have been told and not told. That is roughly sixteen hours of one experienced person's time. It is also the most valuable sixteen hours anyone in your building will spend this quarter, because that file becomes the standard everything is measured against, including your new hires.

Deliberately load it with the nasty ones. Twenty export-screening cases: foreign callers, defense part numbers, requests for technical data, a caller asking for a drawing. Fifteen traceability traps: a part you can only certify with a certificate of conformance where the caller wants an 8130-3, a superseded part number, a lot with a shelf-life expiry. Ten availability traps: consignment stock, allocated stock, a quantity in a quarantine cage awaiting disposition. Ten priority-rating cases where the caller mentions a DO or DX rating, which under the Defense Priorities and Allocations System means you must accept or reject the rated order in writing — fifteen working days for DO, ten for DX — and your agent must not be the thing that accepts it.

Then run the agent against the whole file with nobody on the line. You are not looking for a score out of a hundred; you are looking at the failures one at a time, the same way you would look at a first-article dimensional report. Two failures in the export set is not "97% accurate." It is a stop.

Watching the steps, not just the answer

The second half of the 2026 change is that you can see what the agent did, step by step: which part record it opened, which stock quantity it read, whether it checked the caller against your screening list before it said anything, what it decided to say and what it declined to say. Two years ago you got the final answer and a shrug. Now you get something that reads like a traveler with stamps on each operation.

That matters for a very practical reason: when the agent is wrong, you usually find the cause is your data, not the agent. It quoted stock that the ERP showed as available because nobody had posted the consignment allocation. It offered a lead time from a routing that has not been updated since a machine was retired. In most shops, the first month of this is a data cleanup project wearing a different hat, and that cleanup pays for itself whether or not you ever go live.

The arithmetic: what the gated rollout costs and what the after-hours calls are worth

Illustrative assumptions for a distributor with three inside sales people and a published line that rolls to voicemail after 5:30 p.m.

Note what the test set costs relative to what it protects. Fourteen hundred dollars and a month of restraint, against a single export-control incident that could cost you a voluntary disclosure, outside counsel and a very bad year. Nobody in this trade needs that argument explained twice.

Where the human stays, permanently

Some things should never move to an agent no matter how well it tests. Accepting or rejecting a rated order under the Defense Priorities and Allocations System is a written commitment with legal weight — a person signs that. Confirming airworthiness certification status on a specific serial number is a person's call, because the tag is the product. Anything that touches controlled technical data stays inside your export compliance process, full stop. And quoting a repair after a visual assessment of damage is engineering judgment; an agent can gather the photos and the serial number and open the file, but it does not price the repair.

The other permanent human is the reviewer. Somebody reads transcripts every week — twenty minutes, a sample, and every single case the agent flagged as uncertain. The day that review stops is the day your evidence stops, and evidence is the only reason you were allowed to widen the agent's authority in the first place. Treat it like your internal audit schedule: on the calendar, with a name against it.

Frequently asked questions

How many test cases is enough before I go live after hours?

Two hundred to three hundred real contacts covers most shops, provided you deliberately include the hard categories rather than sampling randomly. A random sample of 500 easy stock checks tells you nothing about the twenty calls that matter.

What is a realistic first tier of authority?

Answer, identify the caller and the company, capture the part number, quantity and need-by date, say plainly that a person will confirm availability and certification in the morning, and book the callback. No stock quantities, no lead times, no certification statements. That tier is genuinely useful and almost impossible to get wrong.

Does an AS9100 auditor care that an agent is answering the phone?

They care if it touches a process in your quality system — customer communication and contract review are in scope. Keep the agent out of contract acceptance and document it as a message-taking step, and it is straightforward. Have it quoting delivery commitments without a defined review, and you have created an unapproved process.

What do I do when it gets one wrong in front of a customer?

Treat it exactly like an escape: contain it — narrow the agent's authority that day — find the cause in the step record, fix it, add that call to the test file permanently, and verify. Your quality manager already owns this method. It works fine on software.

The Monday version of this

Do not buy anything first. Ask your inside sales manager to spend one hour pulling the twenty worst calls of the last year — the ones that made somebody's stomach drop — and write the correct handling next to each. That single page is the beginning of your test file, and it will tell you more about whether you are ready than any vendor conversation.

When you are ready to put something on the line, CallSphere builds AI voice and chat agents that answer business phone lines and web chat around the clock, capture the caller and the request, and book callbacks — with transcripts you can review, which is the part that matters for a shop that has to prove things. Start it in the message-taking tier after 5:30 p.m., measure it against your own twenty worst calls, and widen it only when the evidence says so.