It Quoted 19 Seconds; the Press Ran 26. Test the Agent Against 200 Closed Molding Jobs First.

You tried this. Somewhere in 2024 a vendor put a quoting assistant in front of your estimator, it looked at a print, and it came back with a 19-second cycle on a part that ran 26 on the floor. Nobody could tell you why it said 19. Your estimator went back to his spreadsheet and the whole thing died in six weeks. That was the right call at the time.

What changed in 2026 is not that the agents got magically better at reading prints. It is that the tooling for watching and grading an agent grew up. You can now run an agent against a pile of jobs you already know the answer to, see every step it took to reach its number, and only widen what it is allowed to do once it earns it. The gate is measurement, not faith — and measurement is something a molding shop is unusually well equipped to do, because you already have the answers sitting in your production monitoring history.

The four numbers a molding quote gets wrong, in order

An injection molding quote is not one estimate, it is a chain of them, and the errors compound. Cycle time is first and worst: cooling dominates, wall thickness drives cooling, and a half-millimeter of nominal wall on a PC part is the difference between 24 and 31 seconds. Then cavitation — quoting a 4-cavity tool the customer eventually buys as a 2-cavity doubles your machine hours per thousand. Then the scrap and startup allowance, which everyone sets at 2% because that is what the template says, on a part that historically runs 5.5% because of gate blush on the B-side. Then secondary operations: the pad print that adds 11 seconds of labor per part and never made it onto the router.

Miss cycle by 20% on a 1.2-million-piece annual program and you have given away roughly a fifth of your machine revenue on that part for four years. That is the risk anyone is really talking about when they say they do not trust an AI quote.

What matured in 2026: you can watch the work, step by step

Through the first half of 2026, the practical part of the agent story stopped being the model and started being the instruments around it. You can record every job an agent handles, replay exactly what it looked at and in what order, score its output against a known-correct answer, and hold it at a narrow scope until the score holds. Enterprises doing this at scale are not turning agents loose; they run them shadow-mode against real history first and widen authority in steps.

A test set, in this trade, means a batch of your own closed molding jobs where you already know the true cycle time from the machine monitor, the true scrap rate from the job history, and whether you won or lost the quote — used to grade an agent before it is ever allowed to answer a customer.

flowchart TD
  A["Pull 200 closed jobs from DELMIAworks or Plex"] --> B["Attach the real cycle from the machine monitor, not the router"]
  B --> C["Agent re-quotes each job with the customer name hidden"]
  C --> D{"Cycle within 10% and scrap within 1 point?"}
  D -->|No| E["Estimator replays the agent's steps and fixes the rule it used"]
  E --> C
  D -->|Passes 9 of 10| F["Agent drafts quotes under 250,000 pieces a year only"]
  F --> G["Quality manager re-scores the same 200 jobs every 90 days"]

Building the 200-job set out of your own history

Most shops already have everything they need and have never assembled it in one place. If you run DELMIAworks, Plex, Global Shop, or E2, the closed job history has the estimate. If you run Mattec or any machine monitoring at all, the presses have the truth. The job is to sit them next to each other.

Pull 200 jobs closed in the last three years. For each one, capture: the quoted cycle and the actual average cycle from the monitor, not the standard on the router — those two drift apart the moment a process tech dials in a real setup. Capture the quoted scrap allowance and the actual scrap from the job. Capture the quoted price per piece and whether you won it. Capture the resin, shot weight, cavitation, tonnage, and whether the tool was yours, the customer's, or offshore-built.

Weight the set deliberately. Half should be your bread and butter — the parts you quote in your sleep. A quarter should be the ugly ones: thick sections, hot runner with valve gates, insert molding with a brass insert an operator loads by hand, anything with a texture spec. And a quarter should be jobs you lost, with the price you lost at, because an agent that quotes accurately but always high is a different failure than one that quotes low.

The four grades, and what "passing" means

Score it on four things and do not let anyone add a fifth until those four are green:

The part that is new and worth the effort: when it fails a job, you open that job and read what it did. Usually it is one bad rule — it used nominal wall from the print instead of the thickest section, or it applied your standard 22-second cooling rule to a filled nylon that behaves nothing like unfilled PP. You fix the rule, re-run all 200, and see the score move. That loop used to be guesswork.

What the mistakes cost, with the assumptions on the table

Assume you quote 34 programs a year and win 8. Assume the average awarded program is 600,000 pieces a year at 22 seconds on a 300-ton press, and you cost machine time at $62 per hour. A 20% cycle miss on one awarded program is worth roughly 3,700 hours of machine time you did not price over a four-year life at 600,000 pieces a year — but let us keep it to a single year to stay honest.

One caught cycle error on one program pays for the exercise several times over — and the same 200 jobs also tell you how often your human estimator misses, which is the number nobody has ever measured in most shops.

Where the estimator does not get replaced

Keep a human on anything the test set does not cover. A resin you have never run in production. A tool coming from a mold shop you have not worked with, where the real risk is steel condition and water line layout, not geometry. Anything going into a validated medical process where the quote implies IQ, OQ and PQ runs and a customer-witnessed capability study. Anything where the customer is supplying the tool and you have not seen it — the quote is really a bet on someone else's steel.

Keep a human on the strategic price too. Sometimes you quote a part at a number that makes no sense on the spreadsheet because it holds a family of six other part numbers on your floor. No test set contains that reasoning, and it should not.

And keep a human on the conversation. When the buyer calls to say the price is 9% high, the answer is usually a change to cavitation, a different runner, or a volume commitment — a negotiation, not a recalculation.

Frequently asked questions

How many past jobs is enough? I only have about 90 clean ones.

Ninety with trustworthy monitored cycle data beats 400 where the actuals are the router standard copied forward. Start with what is clean, and add jobs as you close them. The one thing you cannot compromise on is that the actual cycle comes from the press, not from the estimate.

Our machine monitoring only covers twelve of our twenty presses.

Then build the test set from those twelve and be explicit that the agent is only graded on that tonnage range. Scope the authority to what you measured. That is exactly the discipline this approach is for.

What do I do about jobs where the process tech changed the setup and cut three seconds?

Those are the most valuable jobs in the set. Record both the original and the improved cycle. If the agent consistently quotes the improved cycle you have an optimist; if it quotes the original you have a conservative. Either is fine as long as you know which one you have.

Can I let it send quotes directly to customers once it passes?

Not on the first pass. Start with drafting quotes that your estimator signs, on programs under a volume threshold you set. Widen it after two quarters of scores that hold. The whole point is that authority follows evidence.

A note on the calls this creates

Faster quoting only helps if you can also handle the calls that come with it — the buyer chasing a number, the customer asking whether you can hold the price if the volume doubles, the follow-up nobody made because everyone was on the floor. CallSphere builds AI voice and chat agents that answer the phone line and the web chat, capture who called and what program they were asking about, and book the follow-up on your estimator's calendar. It does not price your parts — your estimator and your test set do that.