Most small businesses in Ireland lose deals not because their work is bad but because their quote arrived three days after the buyer asked for it, half-templated, with the wrong VAT line and a scope paragraph copied from the last job that wasn't quite the same shape. I've watched this happen for twenty years inside Tesco, Dunnes and Oracle, and then again from the outside while building IMPT. The quote-and-proposal layer is where an intelligence brain earns its keep for an SME, because it's the part of the business that touches cash directly and is almost always done by the owner at half nine at night.
What "proposal AI" actually means inside a small business
When I say proposal AI I don't mean a chatbot that writes marketing copy. I mean a system that has read every quote you've ever sent, every job you've delivered, every email thread that turned into a purchase order, and every one that didn't. It knows your pricing logic, your standard exclusions, your payment terms, and the way you phrase a scope when the client is a school versus a factory. When a new enquiry lands — by email, by web form, by a phone call your VOIP system transcribed — the brain drafts the quote in your voice, with your numbers, against your templates, and hands it to you for review.
The important word there is drafts. The owner still signs. But the cognitive load drops from "compose a quote from a blank page at 21:30" to "read this draft, fix two lines, send". That is the entire game in SME sales automation. You're not replacing judgement, you're removing the keystrokes between judgement and a sent PDF.
The data layer: what the brain actually reads
An intelligence brain that drafts quotes properly needs a coherent view of four things, and most small businesses already have all four — they're just scattered.
- Historical quotes and proposals. Usually PDFs in a Drive folder, sometimes Word documents, sometimes a folder per year on a shared drive. These are the ground truth for tone, structure and pricing logic.
- Delivered job records. What was actually built, supplied or done, what it cost you, and what you charged. Often in an accounts package — Sage, Surf, Xero, Big Red Book — and partly in a project folder.
- Inbound enquiry traffic. Email inboxes, web forms, WhatsApp Business, sometimes a CRM if you got that far.
- Reference documents. Your terms and conditions, your insurance certs, your tax clearance, your H&S statement, the line about retention you always forget to include.
The brain needs to ingest all four into a single embedding store with proper metadata — date, client, job type, value, outcome (won/lost/withdrawn). I keep this on-premise for our customers because the historical quote archive is one of the most commercially sensitive data assets a small firm owns. It contains your margins. It does not belong in someone else's training set. The architecture I build for this on the small business intelligence brain deliberately keeps the vector store and the inference loop inside a box you can point at.
Retrieval is the part everyone gets wrong
The naive version of quote automation AI is to stuff a model with your last fifty quotes and hope. It produces text that sounds like you and gets the numbers wrong, which is the worst possible outcome — confidently incorrect prices going out under your name.
The version that works splits the problem into three retrieval stages:
- Job classification. When an enquiry comes in, the brain classifies it against your historical taxonomy. Not "construction" — specifically "single-storey rear extension, blockwork, under 40 sqm" or "annual statutory audit, micro company, single trading entity". The classification is built from your own delivered jobs, not a generic ontology.
- Comparable retrieval. Pull the five-to-ten closest historical jobs by both content embedding and structured filters (size, location, client type, year). This is where the margin logic lives — not in a rules engine, in the actual record of what you charged the last ten people who asked for something similar.
- Template assembly. Pull the matching proposal template, the right T&Cs version, the current insurance cert, the right VAT treatment. This is deterministic. The model doesn't get to be creative about whether you're charging 13.5% or 23%.
Only then does the language model do its job, which is to write the narrative paragraphs — scope, assumptions, exclusions, timeline — in your voice, grounded in the retrieved comparables. Numbers come from structured retrieval. Words come from the model. Mixing those two is where bad systems hallucinate prices.
The human review loop, and why it matters more than the draft
I've never shipped a quote-drafting system that sends without human approval, and I won't. Not because the technology can't, but because the failure mode is asymmetric. A draft sitting in your inbox waiting two hours for review costs nothing. A wrong quote sent to a client costs the deal and possibly the relationship.
The review interface matters more than the model. What works in practice:
- The draft renders as a near-final PDF, not as raw text. The owner reads it the way the client will read it.
- Every number is traceable — hover or click and you see which historical jobs and which pricing rules produced it. No black boxes.
- Edits feed back. If you change "supply and fit" to "supply only" three times in a row for a particular client type, the brain learns that pattern. This is where the system actually gets better, not from retraining a model but from updating retrieval signals.
- A reject-with-reason flow. "Too expensive for this client" is a different signal from "wrong scope" and the system needs to distinguish them.
This is the part Irish sales automation tools imported from US SaaS playbooks tend to skip. Owner-led businesses don't want a pipeline dashboard. They want the next quote drafted before they've finished their tea.
Pricing logic that survives contact with reality
The hardest engineering problem in proposal AI for SMEs is not language generation. It's encoding pricing logic that's currently in someone's head. Every small business owner I've worked with has rules like:
- "I add 15% if it's beyond Cahir because of the travel."
- "I don't quote fixed-price for renovations on anything pre-1950."
- "If they're a returning client and the job is under five grand I don't bother with a deposit."
- "Public sector work needs the longer T&Cs and the tax clearance reference up front."
None of this is written down. All of it matters. The brain has to extract these rules from the historical archive — not by asking the owner to list them, because they won't remember them all — but by analysing the variance between similar jobs. Why did this extension get quoted at one rate and that one at another? The structured comparison surfaces the rule, the owner confirms or corrects it, and it goes into the rules layer as an explicit, auditable condition.
This is also where the on-premise architecture pays back a second time. The pricing rules of a small business are arguably more sensitive than the quotes themselves. You don't want them sitting in a multi-tenant cloud where a prompt injection could exfiltrate your margin structure.
Where the whole thing connects to the rest of the business
A quote-drafting brain in isolation is useful. Connected to the rest of the operation, it compounds. Won quotes flow to the job-management side. Lost quotes flow to a follow-up sequence the brain also drafts. Delivered job actuals flow back into the comparables store, so next year's quotes are calibrated against this year's reality, not last year's optimism. Cash collection chases the right invoice at the right interval, in your voice.
This is why I describe the whole thing as an intelligence brain rather than a quoting tool. The quoting module is the most visible piece because it's the one tied directly to revenue, but the data it produces — patterns of what wins, what loses, what gets discounted, what gets walked away from — is the operating intelligence of the business. Most SMEs have never had access to that view of themselves. The big firms I worked at had teams of analysts producing it. A well-built brain gives a five-person company the same visibility, on a server in their own office.
Where to start this week
If you run a small business in Ireland and you want to know whether this applies to you, do one thing this week: gather your last twenty-five sent quotes into a single folder, alongside a note for each one of whether you won or lost it and roughly why. That folder is the seed of your intelligence brain. Everything else — the retrieval, the templates, the review loop, the pricing rules — is engineering on top of that data. Without it there's nothing to learn from. With it, the rest is a few weeks of work. If you want to talk through what that looks like in your specific business, get in touch and I'll walk you through what we'd actually build.