No MES records the idle time that costs you orders. It begins when an email with four PDF attachments arrives in production planning and ends when someone types the first figure into the costing. In this gap, enquiries are lost that you would technically have won long ago.
A constructed example, like all the cases in this text: at 4:40 pm, an enquiry comes in for 250 laser-cut parts made of 3 mm S235JR, two bends, one H7 fitted bore, no surface specification. A few minutes later, before the end of the working day, it lies sorted with production planning. Material, thickness, quantity and deadline have been recognised, the H7 bore is flagged as relevant for costing, and the missing surface specification as a query. Nobody has done any costing yet. The groundwork, for which someone would otherwise type out the title block and email text, is done.
For a contract manufacturer who wants to automate quotation preparation without touching the ERP, exactly this groundwork is enough as a first step.
The bottleneck is in the inbox
Since 2019, job cuts in the metal and electrical industry have added up to almost 350,000, Gesamtmetall (the German metal and electrical industry employers’ association) reported on 14 October 2026 (gesamtmetall.de). Every job cut takes with it experience that is not in any master data table: the feel for drawings, materials and realistic setup times.
The enquiries keep coming, though, and they come differently than they used to. Platforms such as Laserhub advertise a “24/7 instant quote” and cite more than 13,000 customers (as of October 2026, laserhub.com). Buyers who order there are likely to expect this speed of response from you soon too.
That the time is lost before the shop floor is stated, of all places, on the homepage of a costing software provider. KeMeta, which specialises in automated quotation costing for sheet metal and tube parts, writes there: “Many manufacturing companies don’t lose valuable time in production, but before it: in the quotation phase.” (kemeta-software.de)
What an AI agent reads in a request for quotation: title block, dimensions, tolerances
An RFQ agent reads. It has access to your inbox and a few clearly defined tools, and its first job is triage: enquiry, repeat order, complaint or advertising? Only then does the work on the attachment begin.
In the PDF drawing, it is interested in the title block first. That is where you find the drawing number, revision, material (such as 1.4301, S235JR or AlMg3), sheet thickness or raw dimensions, surface specification and the general tolerance. From the body of the drawing, it takes the outer dimensions, number of bends or machining sides, threads and tight individual tolerances, which later determine the choice of machine. From the email text come quantity, batch-size tiers, requested date, delivery address and special requirements such as inspection certificates to EN 10204.
Whether an agent is useful or dangerous depends on what it cannot read reliably. A good agent returns a confidence value for each field and flags what is uncertain: a scanned hand sketch, a title block covered by a stamp, a tolerance that appears in the dimensioned view but not in the text.
These fields land with production planning as open items, the missing surface specification from the example above as a query. A bad agent would simply enter “none” there. Then the costing is missing a galvanised finish that the customer tacitly expects.
Handover to the ERP: create, yes; decide, no
The handover decides whether the agent helps production planning or makes work for it. In Proalpha or abas, it creates an enquiry or prospect record, attaches the email and drawing, suggests existing items or previous projects with the same drawing number and hands over a structured costing sheet. In PSIpenta and SAP Business One the pattern is the same; only the objects have different names. Writing prices, sending quotes and changing master data are outside its permissions.
Take drawing number 4711. The agent finds a previous item for it and suggests it. What it doesn’t see: the new enquiry is at revision C, the item in the ERP was costed according to revision A, and in between the tolerance on the bore has become tighter.
If the old price went out without approval, nobody would know about the new tolerance. This way, costing sees the suggestion, spots the revision jump, adds setup time and a material surcharge and approves. Only then is the quote created in the ERP.
We recommend restricting the agent’s read and write permissions to exactly the objects it needs and logging every write operation. What a connection with an explicit tool allowlist and approval for every write call looks like technically is covered in our article on MCP servers for ERP and CRM.
In short: the agent is costing’s front-office assistant. It sorts, reads out and opens the file. Someone else signs.
AI agent or AI costing tool: what fits costing at a contract manufacturer
Our recommendation up front: if 80% of your enquiries are clean STEP files for laser-cut parts, buy a costing tool and leave out the agent. If your enquiries arrive as emails with a PDF, a sketch and three follow-up questions, the agent is what gives the tool clean input in the first place. The further your parts range is from standard laser-cut parts, the more likely the second scenario is. The strongest combination is then: agent at the front, tool or human in the middle, ERP at the back.
| Criterion | AI agent for enquiry intake | AI costing tool |
|---|---|---|
| Core task | Understand email and drawing, structure data, create ERP record | Analyse geometry, calculate times and costs, suggest price |
| Typical input | Unstructured email, PDF drawing, scan, follow-up questions | CAD file (STEP, DXF) or machine-readable drawing, stored machine and hourly rates |
| Result | Complete enquiry with open items, ready for costing | Quoted price, often with a PDF quote |
| Strength | Works with the chaos that is actually in the inbox | Reproducible costing for standard processes (laser cutting, bending, milling) |
| Limitation | Does not provide a price | Needs geometry data and maintained costing rules; exotic parts remain manual work |
It gets harder with a welded assembly of nine parts, three of them as hand sketches and one as a photo taken on the workshop bench. A costing tool that calculates on CAD data has nothing to go on for four of the nine parts. The agent sorts what is there, takes over the five clean parts as line items in the enquiry record and words the missing CAD files as a query to the customer. In this case, production planning starts with a sorted file, and the query is already written.
Oroox and KeMeta, for example, explicitly advertise with AI. Oroox calls its price engine “AI-powered”, claims “over 95% accuracy” by its own account, promises quotes “in seconds” and connects ERP and MES via API (oroox.com).
At KeMeta, according to its homepage, an AI recognises geometries, materials and machining steps; DXF, STEP, IGES and PDF are accepted, and the output is a PDF quote with a price breakdown. On its product page, KeMeta also describes how it introduces the Quotation Factory solution into existing system landscapes (KeMeta page on Quotation Factory).
Working in the same field are PROfirst CALC, which costs laser, plasma, oxy-fuel and waterjet parts from DXF, PDF, STEP and other formats (profirst-group.com), and costing24 with the promise “simply calculate models” (costing24.com). These tools calculate. An agent reads, sorts and hands over.
The pilot: 50 real enquiries before you automate quotation preparation
Don’t start with the integration. Start with a folder. Take 50 closed enquiries for which you know the quote and the order outcome, and mix them deliberately: clean STEP cases, pure PDF drawings, scans and at least one assembly. Let the agent process these cases blind and compare field by field. First a test set of real cases, then scale: that is also what our lessons on AI agents as digital employees say.
Plan two to three weeks for this. Compared with an ERP interface, that is a manageable effort, and it tells you whether the interface is worth it at all. The fields the agent is supposed to fill are clear, and enquiries come in every day. That is the kind of process with clear rules and high volume that, according to our article on AI automation in SMEs, you should start with.
Before the first customer drawing is uploaded, clarify in the pilot contract where the model runs and whether your data will be used for training. Customer drawings are someone else’s intellectual property. Also specify when the drawings will be deleted after the pilot ends, on your side and on the provider’s. Because RFQ emails contain names and contact details of contact persons, you will as a rule also need a data processing agreement (DPA) with the provider under Art. 28 GDPR.
As of October 2026, preparing quotes is not one of the high-risk applications under the AI Act, whose obligations for Annex III only apply from 2 December 2027 anyway (digital-strategy.ec.europa.eu). Art. 4, on the other hand, has applied since February 2025: whoever deploys the agent must take measures to support the AI literacy of the people who work with it.
What else the Digital Omnibus has postponed and who is responsible in Germany is covered in our article on the EU AI Act for SMEs. It is no substitute for legal advice.
Measure four things:
- The hit rate per field: material, thickness, quantity, deadline and tolerances separately, because the errors are distributed very differently across them.
- The rate of correctly flagged uncertainties, because an agent that confidently enters wrong values is worse than none.
- The time between the email arriving and a record that is ready for costing.
- The error classes that remain: quantity “2 × 250” read as 2, thickness taken in inches although millimetres were meant, material taken from the email text although the title block says something else.
These error classes show you whether you have a model problem, a scan problem or a process problem.
What remains: more enquiries reach costing in time
An AI agent in enquiry intake doesn’t make your quote better or your price sharper. Both stay with production planning and sales. What the agent changes is the number of enquiries that reach them in time at all.
If you have abas, Proalpha, PSIpenta or SAP B1 in-house and are thinking about automating the quotation process, you need neither a new ERP nor a big platform to get started. Take 50 enquiries, an agent with read permissions and a production planning team that is willing to compare fields for three weeks. Afterwards, you will know whether your business needs a costing tool, an agent or both. Before that, nobody knows, not even an agency.