AI Quoting Agent, Self-Hosted: Faster Quotes, Data Onshore
Your quotes sit for days while the customer asks someone else
A self-hosted AI quoting agent drafts each quote from your own price book within the workflow, a person approves it, and your pricing data stays on infrastructure you control in Australia. You can build it without writing code, and whether your IT team agrees to it depends mostly on where the AI runs.
If you run sales, operations or estimating at a business that quotes a lot, you know the queue. Requests for quote land in a shared inbox. Someone opens the PDF, finds the items in the price book, applies the customer's terms and keys it all into a quote. Meanwhile the customer has emailed two other suppliers.
What is a self-hosted AI quoting agent? It is a workflow that reads a quote request, looks up prices in your own systems, drafts a priced quote using your margin rules and sends it to a person for approval. "Self-hosted" means the workflow runs on your server, and the AI model runs either in an Australian cloud region under your own account or on hardware you own.
This guide covers how the agent works as a business flow, the two ways to host the AI behind it, how to choose between them, and when self-hosting is the wrong call.
Working out whether this fits your business? Book a free 30-minute review
Why does quote speed matter so much?
Research published in Harvard Business Review in 2011 ("The Short Life of Online Sales Leads", Oldroyd, McElheran and Elkington) found that firms contacting a lead within an hour were nearly seven times as likely to qualify it as firms that waited longer.
A quote request is a lead where the customer has already told you what they want and is waiting for a number.
Most of the delay sits in the reading and the looking up, which a machine can draft. The judgement call, whether this customer gets the sharper price, stays with your people.
| What slows a quote down | Who it waits on | Can an agent draft it? |
|---|---|---|
| Reading the RFQ, PDF or spec | Whoever opens the inbox | Yes |
| Finding each item in the price book | The person who knows the codes | Yes, if item data is clean |
| Applying customer terms and margin rules | Estimator or sales lead | Yes, as fixed rules |
| Deciding on a special price | Sales manager | No, a person decides |
| Checking and sending | Estimator | A person approves; the send is automatic |
Why won't IT let you use the obvious AI tool?
This is a common starting point. Someone tries a public chatbot, pastes in a customer request and a page of the price list, and gets a decent draft. Then IT or the finance director asks, fairly, where that data went.
Your price book, margins and customer list are the business. Under Australian Privacy Principle 8, an organisation that discloses personal information overseas must take reasonable steps to ensure the recipient does not breach the APPs. Under section 16C of the Privacy Act it stays accountable for what the overseas recipient does, according to the OAIC's APP 8 guidelines.
Customer contacts on a quote are personal information. Your pricing is commercially sensitive whether or not the Privacy Act applies to it.
Self-hosting moves the AI to somewhere your IT team already accepts, with a person approving every quote. For the wider picture of where AI data travels, see our guide to AI data sovereignty and onshore deployment.
How the quoting agent turns an RFQ into a draft quote
The agent is a chain of ordinary steps. Only one or two of them involve AI. The rest are lookups and rules your team already follows.
Here is the flow for a typical wholesale or trade supply quote:
How a quote request becomes an approved quote
The AI reads the request and matches messy descriptions ("20 x 90mm PVC elbows, DWV") to your item codes. The prices and margins come from your system and your rules.
Then a person approves the draft before anything reaches a customer. In n8n this is built in: the human-in-the-loop documentation shows how to require approval before an AI agent takes an action, and the Slack node can send and wait for an approve or decline.
For quotes built from photos and site walks rather than written requests, the flow changes. Our posts on AI field quoting for service operators and roofing estimation automation cover those cases.
Where does "no code" fit in?
No code means you build the workflow by dragging steps onto a canvas in a tool such as n8n. Your operations team can read it, and a change to a margin rule is a change to one box.
n8n suits this job because you can host it yourself. Its Community edition is free to self-host and includes almost the full feature set. It leaves out enterprise features such as single sign-on, environments, shared credentials and Git version control.
Two licence points matter before you commit:
- n8n's Sustainable Use License lets you use the software "only for your own internal business purposes or for non-commercial or personal use". Running your own quoting fits that. Reselling n8n as a hosted service to others does not.
- On a self-hosted instance, n8n states in its privacy documentation that workflow data "stays on your infrastructure and we never receive it". Telemetry is on by default, and the administrator turns it off with environment variables.
n8n also ships ready-made model connectors for Microsoft Foundry (Azure OpenAI), AWS Bedrock, Google Vertex and Ollama. So switching the AI behind the agent later is a change of settings inside the workflow.
If you are still choosing a workflow tool, our Zapier vs Make vs n8n comparison covers pricing and fit.
Option 1: AI in an Australian cloud region, under your own account
The first way to host the AI brain is a managed model service from Microsoft, Amazon or Google, set to an Australian region in your own cloud account. Nobody on your team looks after servers or GPUs. You pay for what you use.
The catch is that each provider offers a different set of models in Australia, and the deployment type decides where your data is processed. Here is what the official documentation says as at 25 October 2026.
Azure OpenAI (Australia East)
For Standard (regional) deployments, Microsoft states that it "processes prompts and responses in the region associated with your deployment". Its regional availability page lists gpt-4.1-mini and gpt-4o (2024-11-20) for Standard deployment in australiaeast, plus embedding models. Newer models appear under Global or Data Zone types. Microsoft defines the Asia Pacific data zone as processing "within any Asia Pacific nation", which is wider than Australia.
Amazon Bedrock (Sydney)
The Bedrock regional availability page lists Amazon Nova Pro, Lite and Micro in-region in Sydney. Claude models are reached from Sydney through cross-region inference. AWS documents an au profile "which routes only among Australian Regions", alongside an apac profile that can route anywhere in Asia Pacific and a global profile.
Google Vertex AI (Sydney, australia-southeast1)
Google's locations page lists Gemini and partner models, including Claude, at the Sydney endpoint. It also warns that "endpoints don't guarantee data residency or in-region ML processing". Check the data residency page for the specific model before you rely on Australian processing.
What this means for a quoting agent
Across all three, the newest models usually arrive first under global routing, and in-region options lag. This is the trap a regulated buyer hits when the model they wanted is not offered for regional deployment in Australia East. For a quoting agent it rarely matters, because reading an RFQ and matching item codes does not need the newest model.
For a deeper side-by-side of the three clouds, see AWS vs Azure vs Google Cloud for AI in Australia.
Option 2: Your own model on your own server
The second way is to run an open-weight model yourself, using a tool such as Ollama or vLLM, on a server in your building or on your own Australian virtual machine. Nothing leaves your network.
Ollama's FAQ is direct about it: "Ollama runs locally. We don't see your prompts or data when you run locally." Ollama also offers cloud-hosted models, and its FAQ explains how to switch cloud features off for local-only mode. Whoever sets it up should do that.
Running your own model means accepting a few limits:
- The models you can run on a single server are smaller than the cloud flagships. They handle structured reading and matching well. They are weaker on long, messy specs.
- You own the upkeep: hardware, patching, model updates and monitoring. If the server stops, quoting falls back to manual.
- The cost is mostly upfront hardware rather than a monthly usage bill.
Our OpenAI vs Claude vs Ollama comparison goes further into model quality.
Which hosting option suits your business?
Both options keep your price book out of overseas consumer AI tools. The difference is who looks after it and how much the model can do.
Australian cloud region vs your own server
| Metric | Australian cloud region (Azure, AWS, Google) | Your own model (Ollama or vLLM) |
|---|---|---|
| Where the data sits | Your cloud account, Australian region, if the deployment type is regional | Your server, your building or your own Australian VM |
| Cost shape | Pay per use, no hardware | Hardware upfront, low running cost |
| Upkeep | Provider runs the model | Your IT team or a partner runs it |
| Model quality | Larger models, though newest ones may be global only | Smaller open-weight models |
| Fits your IT team if | You already run Azure, AWS or Google | You run on-prem servers and want no external calls |
| Main risk | Choosing a global or data zone deployment by mistake | Nobody owning the server after go-live |
Use this to narrow it down:
Pick where your quoting agent's AI runs
When should you not self-host a quoting agent?
If you quote a few jobs a week, a good template and a clean price list will save more time than any agent.
A self-hosted workflow needs an owner for updates, backups and broken connections. With nobody in that seat, expect it to break and stay broken.
The agent can only look up prices that exist in a system. If your best estimator prices from memory, write that down first.
Bespoke design and fabrication jobs need an estimator's judgement on every line. AI can still read the drawing pack, but the draft saves less time there.
In our lead's prior work on operational data platforms for BHP and Rio Tinto, gained while working with previous consulting employers (these companies are not Solve8 clients), the same lesson came up repeatedly: the quality of the master data decided the result far more than the tool on top. Quoting works the same way, and clean item codes do more for you than a bigger model.
How long does it take to get the first draft quotes?
A sensible build starts with one quote type, measures today's turnaround, and runs the agent beside your team before any customer sees its work.
From first workshop to approved quotes going out
Running side by side (we call it shadow mode) means the agent drafts quotes your team never sends. You compare its drafts to the real ones each week. You set the accuracy level it must reach before it moves to approval mode, and nothing reaches a customer until it does.
Each phase ends with a decision to proceed, pause or stop, and you keep everything built so far.
What could a quoting agent free up?
Consider a hypothetical trade wholesaler. The figures below are assumptions for illustration, not results, so swap in your own.
Worked example: estimator hours on quote preparation (assumptions)
Measure your own minutes per quote before anything is built. That baseline is how you will know whether the agent earned its place.
Here is what changes in the day-to-day for the estimator:
An estimator's quote, before and after
| Metric | Manual quoting | With a self-hosted quoting agent |
|---|---|---|
| First look at the request | When someone gets to the inbox | A draft is waiting for review |
| Item lookup | Searched by hand in the price book | Matched to item codes, flagged where unsure |
| Margin rules | Applied from memory, varies by person | Applied the same way on every draft |
| Final decision | Estimator | Estimator, unchanged |
| Where pricing data goes | Stays in your systems | Stays in your systems and your Australian region or server |
Your next step
Start with one number: how long a typical quote takes from inbox to sent, today. Then list where your prices actually live, whether that is an ERP, Xero or MYOB items, or a spreadsheet.
Those two facts shape both the hosting decision and the build. Our process automation and private AI infrastructure teams scope exactly this kind of work, and our own SupportAgent product ships as a self-hosted Docker install for the same data reasons.
Book a scoping call and leave with a straight answer on whether a quoting agent is worth piloting for your quote volume, and which hosting option your IT team will accept.
Common Questions
Can I build an AI quoting agent without writing code?
Yes. A workflow tool such as n8n lets you build the steps on a visual canvas: read the email, call the AI model, look up prices, apply rules and wait for approval. Connecting to an ERP or price book sometimes needs technical setup, but the quoting logic itself stays readable to your operations team.
Is self-hosted n8n free for business use?
n8n's Community edition is free to self-host and includes almost the full feature set. Its Sustainable Use License allows use for your own internal business purposes, which covers running your own quoting. Enterprise features such as single sign-on and Git version control need a paid licence.
Which AI models can I run in an Australian region?
It depends on the provider and deployment type. As at 25 October 2026, Microsoft lists gpt-4.1-mini and gpt-4o for Standard deployment in Australia East, AWS lists Amazon Nova models in Sydney with Claude through an Australia-only routing profile, and Google lists Gemini models at its Sydney endpoint. Check each provider's page before you commit.
Is running my own model with Ollama better than an Australian cloud region?
Neither is better in general. Your own model keeps everything inside your network but is smaller and needs someone to run the server. An Australian cloud region gives you larger models with no hardware, as long as you choose a regional deployment rather than a global one.
Will the agent send quotes to customers on its own?
It should not. The safe design has a person approve every draft before it is sent, and n8n supports approval steps for exactly this. The agent drafts and looks up; your estimator still decides the price a customer sees.
How much quote volume justifies a quoting agent?
If you quote only a few jobs a week, a clean price list and good templates will usually do more. The case gets stronger when quotes queue behind one or two people, when pricing varies by who prepares it, or when requests arrive faster than your team can read them.
Related Reading:
- AI Field Quoting for Australian Service Operators - How quoting from site walks and photos differs from quoting written requests.
- AI Data Sovereignty: Where Data Goes - Where AI data actually travels, and how to keep it onshore.
- AWS vs Azure vs Google Cloud for AI in Australia - A wider comparison of the three clouds for Australian AI workloads.
- AI Agent Governance: Privacy and Human Override - How to set data access and human override rules for any agent.
Sources: Microsoft Learn, Region availability for Foundry Models sold by Azure (updated 19 October 2026); AWS, Amazon Bedrock regional availability and cross-Region inference documentation (accessed 25 October 2026); Google Cloud, Vertex AI generative AI locations and data residency pages (accessed 25 October 2026); n8n documentation: Community edition features, Choose how to use n8n, Privacy, human-in-the-loop and model node pages, and the Sustainable Use License (accessed 25 October 2026); Ollama FAQ (updated 22 October 2026); OAIC, APP Guidelines Chapter 8; Oldroyd, McElheran and Elkington, "The Short Life of Online Sales Leads", Harvard Business Review, March 2011. Our data sovereignty guide is at /guides/data-sovereignty-australia/.