RFQ automation turns a customer's request for quotation into a quote in your ERP without someone retyping it. For spare parts, the request is usually an email, the parts are described in the customer's own words, and somewhere behind it is a machine waiting for a part, with a customer who would like a price before it has been idle much longer.
Reading that email is the easy bit. Any capable language model can summarize it. The work is in matching the right customer, finding the right material numbers, and writing proper records into SAP, in a way the engineers who own the quotes are willing to trust. This post sets out six habits an AI agent should have before it writes anything to SAP, then looks at how one agent puts them into practice.
Key Takeaways
- Spare parts quoting makes a strong first AI agent: high volume, well-defined steps, and customers waiting on equipment that may be down.
- An agent that writes to SAP should show its plan, score each step, hand off what it shouldn't handle, flag what it inferred, wait for approval, and leave a trail.
- A frontier LLM can read the email, but it has never seen your SAP records or parts documentation. The agent needs a platform that supplies that context.
- Before choosing an RFQ automation tool, ask who approves a quote before it reaches your ERP, and whether every quote traces back to its email.
Why Is Spare Parts Quoting a Good First Job for an AI Agent?
Most aftersales teams know the routine. A request arrives, written in whatever language the customer prefers, sometimes with a scanned attachment. Someone works out who sent it, finds the business partner in SAP, translates "the usual seal kits" into material numbers, checks the catalog, types in the line items, creates the service notification and the quotation, and files the email. Then the next one arrives.
None of these steps is difficult, and that is the point. The work is repetitive, the volume is steady, and the steps barely change from one request to the next. It also falls to the engineers who know the equipment best, the same people customers want on the phone for technical questions. Nobody went into engineering to retype part numbers.
Those are the conditions we look for in a first agent: friction that is easy to see, a process with clear edges, and a result that people will notice within weeks. We wrote more about choosing that starting point in how the first AI agent sets up the next ones.
What Should an AI Agent Do Before It Writes to SAP?
An agent that only drafts text can afford to be approximately right. One that creates records in your ERP can't, because a quotation is a commitment to a customer. In our experience, engineers hand work over to an agent once they can see how it reached each result. Six habits make that possible.
Show the plan before it starts. The engineer should know which steps the agent will take, and in which order, before anything runs. A plan on the table is easier to trust than a result that appears out of nowhere.
Score each step. A single confidence figure for the whole quote says little. A score for each step, such as matching the customer or identifying a part, tells the engineer where to look, which beats rechecking everything.
Know when to hand off. Some requests shouldn't be automated at all: duplicates, requests a colleague is already handling, or anything the agent can't read reliably. These cases need defined exits to a person, with the reason attached. Those exits should be decided before the run, so nobody discovers them halfway through.
Say what it inferred. Customers don't always write tidy lists. When an agent picks up something the customer implied without stating it, it should say so, so the engineer can confirm it or take it out.
Wait for approval before writing. Drafting can be automatic. Creating records in a system of record should wait for a person who is accountable for the quote.
Leave a trail. The quotation, the approval, and the original request belong together. When a customer calls about a quote months later, nobody should have to reconstruct what happened from memory.
The same pattern applies wherever an agent touches a system of record. Our RFP and proposal agent applies it to bid answers, our guide to human-in-the-loop AI agent approval covers the principle in general, and agents that run on a schedule, without anyone watching need even clearer rules about what they may do alone.
How Does InstaQuote Put These Habits Into Practice?
Principles are easier to write than to build, so it helps to look at an agent that follows them. InstaQuote is an AI agent for spare parts quoting in the Squirro Agent Catalog, a collection of ready-to-deploy agents built on the Squirro platform. It reads incoming requests from the shared mailbox where they arrive, matches the customer and the parts against SAP and the company's technical documentation, and prepares a quotation for an engineer to review. Our interactive demo shows it working on sample data, and the example below comes from there.
A customer emails a request for two seal kits and six fuel filters for an MTU 396 engine. Before doing anything else, InstaQuote lays out its plan: 11 steps from the email to the quotation in SAP. The early steps check whether the request belongs in the automated path at all. If it's a duplicate, if a colleague has already picked it up, if it's in a language the agent hasn't been set up for, or if the customer match comes back uncertain, the request goes to a manual queue with the reason attached.
This request passes those checks, and the agent works through it step by step. Reading the request, identifying the language, matching the business partner, and identifying the materials each get their own confidence score, so the engineer can see at a glance which parts of the quote are solid and which deserve a second look.
Further down the email, the customer mentions an O-ring set for the same engine series without adding it to the list. It's the kind of detail that is easy to miss on a busy morning and awkward to discover after the quote has gone out. InstaQuote adds it as a third line item and marks it as taken from the text.
Every quote is approved before it reaches SAP. InstaQuote stops after preparing the line items and presents them to an engineer, with confidence scores and the inferred item flagged. No service notification, quotation, or other record is created in SAP until the engineer approves. Only then does the agent run the steps that write to SAP.
After approval, the agent creates the service notification and quotation as ordinary SAP records and archives the original email with them. All six habits show up in a single request. For the engineer, the result is a prepared quote with the uncertain parts already marked.
Where Does a Frontier LLM Fall Short?
A frontier LLM will read a spare parts request in seconds and summarize it nicely. Then it gets stuck, because it has never seen your SAP records or your parts documentation. It can't tell you which business partner sent the email, and it has no way of knowing whether "seal kits for the MTU 396 engine" corresponds to a real material number in your catalog.
InstaQuote runs on the Squirro platform, which supplies that missing context:
- LLM extraction pulls customer details, material numbers, and quantities out of the email body and its attachments.
- Connectors to SAP, through custom API and OData integrations, validate the business partner and each material against your live records.
- Technical knowledge base ingestion brings in product documentation, manuals, and past quotes, including scanned documents that need OCR before they can be read.
- Cognitive search combines keyword and semantic matching, so an exact part number and a loose description can both find the right entry.
- Entity extraction and aggregation turn those documents into data you can filter by customer, equipment, or part.
Put together, they give the agent a working digital spare parts catalog to search, built from documents you already have. The language model handles the reading. Knowing what the email actually refers to comes from the platform underneath.
What Changes for the Engineer?
Before an agent is involved, an engineer handles every request from start to finish: reading, matching, looking up, typing, and filing.
With InstaQuote, the engineer opens a quotation that is already prepared and spends their time on the parts that need judgment:
- Checking items the agent flagged as inferred.
- Looking at any step with a lower confidence score.
- Approving the quote, or correcting it first.
Requests the agent shouldn't handle arrive in the manual queue with the reason attached. Engineers get time back for technical questions, and customers get their price sooner, which is usually the first thing they ask about.
What Should You Ask Before Choosing an RFQ Automation Tool?
There's more than one sound way to build RFQ automation, and the right fit depends on your requests, your ERP, and how your team works. Whatever you evaluate, these questions tend to show how much the tool can be trusted with:
- How can your team see how a quote was put together, and where the tool was unsure?
- What happens to requests the tool can't handle reliably, and who picks them up?
- How does it confirm that a customer and a part exist in your ERP?
- Which of your documents can it learn from, and what does it take to keep them current?
- Who approves a quote before it becomes a record, and can you set that rule yourself?
- Could you trace a quote back to the original request months from now?
Good answers will sound different from vendor to vendor. If they're vague, though, the quotes will be too.
We walked through InstaQuote end to end, alongside an agent for RFP responses, in our on-demand webinar, Proven AI Agents Driving Transformation in Manufacturing. You can also try the interactive InstaQuote demo, which uses sample data.
To talk through how InstaQuote could fit your quoting process, book a demo.