How We Cut Quote Turnaround From Days to Minutes With AI Triage

Two years ago we had a problem that was quietly costing us jobs. A customer would fill in our repair quote form or ring the shop, we would promise to get back to them, and then the real work began: read the request, work out what was actually wrong, check whether we could help, look up parts and pricing, build a quote, get it approved, send it back. On a good day that took a few hours. On a busy day it stretched into the next morning, and by then the customer had often already spoken to a competitor.

We were not losing jobs because our prices were wrong or our service was poor. We were losing them because we were slow. This is what we built instead, what it changed, and how any service business with the same bottleneck could copy the approach. We are a Manchester technology business that started at the repair bench, so this is our own workflow, not a client’s.

What does a slow quote actually cost?

A slow quote costs you the jobs you never hear about, because the customer has already booked someone else by the time you reply.

The customer is not waiting for you

When someone requests a repair quote they are not sitting by the inbox. They are ringing the next shop on Google, asking a friend, or deciding whether it is even worth fixing. Every hour you take is an hour where their urgency cools and somebody else frames the conversation.

The research says the same thing

This is not just our impression. Forrester’s State of Business Buying 2024, reported by Cincom, found that 86% of B2B purchases stall during the buying process and 81% of buyers end up dissatisfied with the provider they chose, often citing responsiveness. The same Cincom piece notes that firms following up within an hour of an enquiry are nearly seven times more likely to qualify that lead.

Repair customers are not B2B buyers, but the behaviour rhymes. Whoever answers first gets to set the terms.

The pattern we kept seeing

A request would arrive late afternoon. We would reach it the next morning. The quote went out by lunchtime. Half the time, the customer had already made other plans. We knew exactly what the problem was. We just could not answer every request instantly without hiring, and the quoting itself was too manual to speed up by trying harder.

What did we actually build?

We built a triage and quoting workflow that sits inside our CRM and carries a request from “customer asks” to “quote in their inbox, ready to accept”.

Where requests come from

A repair request arrives one of three ways: through the quote form on our website, over the phone where staff log details using a guided wizard, or by email. We are office based and collect and return rather than taking anything over a counter. Whichever route it takes, it lands in the same repair inbox with a unique reference and a status showing where it sits in the pipeline.

At that point it is raw customer information: a postcode, a device, a description of the fault, sometimes a photo. Nothing has been triaged or priced.

The step that used to eat the morning

Everything slow about the old process lived in the gap between that raw request and a person having enough context to make a decision. Reading, interpreting, researching, assembling. That gap is what we automated, and it is the part worth automating in almost any business.

How does the AI triage work?

Two things run automatically the moment a request lands: a summariser and a classifier.

Reading and summarising

The first skill reads the customer’s description, which can be anything from “won’t turn on” to three paragraphs about what happened beforehand, and writes a clean structured summary: what is broken, what they have already tried, what they need, and any red flags.

Classifying against our own rules

The second skill sorts the request into one of four categories:

• Standard repair, which we can handle confidently.
• Needs more information, where there is not enough detail to quote.
• High risk, meaning complex, uncertain or low success rate.
• Do not take on, such as data recovery from a physically damaged drive, which needs a clean room we do not have.

These are not guesses. They come from a triage policy we wrote with our engineers, reflecting years of knowing which jobs are worth taking. If someone describes a laptop that will not charge and has already tried another charger, the classifier marks it standard and flags the likely causes.

The whole pass takes seconds. By the time a human opens the request, the first read is done.

How does the quote get built?

For standard repairs the AI proposes a quote, assembled from pre-approved parts rather than invented from scratch.

A library, not improvisation

Proposals are built from a library of quote templates and line items covering the jobs we see most: screen replacements, charging port repairs, battery swaps, software reinstalls, virus removals. Each carries parts, labour, VAT and a realistic turnaround. The AI matches the request to the closest template and adjusts line items where the description warrants it.

This is the detail that makes it fast and safe at the same time. The AI is not pricing anything. It is assembling components a human already approved. It is also why the running cost is trivial: matching a description to a template is exactly the routine work that small, cheap models handle well, for the reasons set out in why AI got cheap.

The approval queue is the point

No AI-drafted quote ever reaches a customer unreviewed. Every one lands in an approval queue where a person checks it, adjusts it, and sends it. That human-in-the-loop step is what lets us keep control of pricing and tone while still getting the speed.

Approving a draft takes about 30 seconds. Building one from scratch took 10 to 15 minutes. That single ratio is the whole business case.

What does the customer see?

The customer gets an email and an SMS linking to their own quote page. There is no login: the link itself is the authentication, signed with a token that expires and is tied to that request.

Three things they can do

The page shows every line item, the total including VAT, an estimated turnaround and a plain-English description of the work. From there they can accept and book a collection slot, decline and tell us why, or ask a question that routes back into the CRM.

If they accept, they pick a date and time window. The system checks real availability against that day’s collection route, the same-day cut-off and the distance from our depot, then sends a confirmation and generates a QR code for the driver to scan on arrival.

From quote to booked in under two minutes

The whole path from quote sent to collection booked can happen while the customer is still on their lunch break. When the driver scans the code, the request converts into a job and moves into the engineer workspace for diagnosis, repair, testing and return.

What changed when it went live?

We rolled it out gradually, testing on a handful of requests before opening it to everything. The difference was immediate.

The honest version of the numbers

Our conversion rate improved, but we cannot give you a clean before-and-after percentage because the old process was too manual to measure properly. What we can say is that the “sorry, we have already arranged something else” replies dropped off and the “wow, that was fast” messages increased. We would rather tell you that than invent a figure.

What the freed time was actually for

The gain was not really the hour saved. It was that the team stopped spending mornings on information assembly and started spending them on the jobs that genuinely needed a person: the complex cases, the edge cases, the customers who needed a conversation rather than a form.

Why does human approval not slow it down?

We worried it would. It does not, and it turns out to be the reason the system is trustworthy rather than merely fast.

Thirty seconds against fifteen minutes

Reviewing a draft is a different cognitive job from building one. You are checking a proposal, not constructing it. That is why the approval step costs seconds while the original task cost a quarter of an hour.

It keeps the team in the loop

The team are not rubber-stamping. They spot the edge cases, notice when a template is wrong, and feed that back. The approval queue is not friction; it is quality control and a continuous improvement loop in the same step. It is the same principle we describe in our guide to AI agents: the boundary matters more than the capability.

How could another service business build this?

You do not need a custom CRM. You do need to do the unglamorous parts in the right order.

Start with the triage rules, not the technology

The most important part of our system is not the AI, it is the written policy behind it. We built it by sitting down with our engineers and asking which repairs we do constantly and what they cost, which jobs look simple and turn out not to be, which requests we should refuse, and what we always need to know before quoting.

If the rules are wrong the AI will amplify the wrong decision faster. Write the logic down before you think about models or integrations. That document is worth more than any tool.

Build the quote library second

Every time you send a quote, save it as a template. Every time you price a common job, add it to the list. Within a few weeks you will have most of your quoting covered by reusable components, and that library is what makes the AI both quick and safe.

Use AI to read and match, never to invent

Ours does three jobs: read the description, summarise it, match it to a category and a template. It does not invent pricing, conjure parts we do not stock, or send anything unreviewed.

That split matters. OpenAI’s State of Enterprise AI 2025 report found that ChatGPT Enterprise users save roughly 40 to 60 minutes per working day, with the gains concentrated in exactly this kind of context assembly rather than in end-to-end automation. Our experience matches: the engineer’s expertise was never the bottleneck. The ten minutes of reading and research before they could apply it was.

Make accepting frictionless

Remove every barrier between “customer receives quote” and “customer says yes”. If you are not building a custom page, you can get most of the way with a clear PDF and a scheduling link, or a simple form where they confirm and pick a slot. Every extra click is a chance to drift away.

Then measure four things

Time from request to quote sent, approval time, conversion rate, and error rate. We built a small dashboard showing these by week. If the error rate climbs, the triage rules need work. If conversion drops, the problem is pricing or wording, not the AI.

What did we get wrong?

Three things, and they are the useful part of this story.

The classifier was too cautious

The first version flagged almost everything as “needs more information” because we had not shown it enough examples of a complete request. We fixed it by taking 50 real requests, labelling them ourselves, and using those to refine the rules. The AI did not get smarter. We got better at teaching it.

We nearly skipped the boring foundations

It was tempting to start with the clever part. Had we done so, we would have built a fast system that made confident, wrong decisions. The templates and the triage policy were where the actual value sat.

We had to handle customer data properly from day one

The workflow touches names, addresses and device details, so the usual duties apply regardless of how clever the automation is. The government’s data protection overview is the plainest summary of the basics, the ICO’s advice for small organisations and its guidance on artificial intelligence cover what that means for an AI system, and the NCSC’s Small Business Guide covers keeping it secure. Deciding this at the start cost us nothing. Retrofitting it would have cost a great deal. If certification matters to your customers, that is what our Cyber Essentials service covers.

The bigger point: new workflows, not faster old ones

The goal was never to do the old process quicker. It was to build a process that was not previously possible.

The old shape and the new one

The old workflow was: customer asks, staff member reads, staff member researches and builds a quote, staff member sends, customer replies. It worked. It did not scale.

The new one is: customer asks, AI reads and classifies and proposes, human approves in 30 seconds, customer accepts and books online. The difference is not only speed. It is that the second version handles 10 requests a day or 100 without more staff.

Where to look in your own business

Ask how long it takes you to turn a request into a sent quote, how many of those requests follow a predictable pattern, and where you genuinely need judgement rather than information assembly. The answers point straight at where automation pays. Our guide to what AI automation can actually do covers the general version, and missed calls covers the same leak one step earlier in the funnel.

You may not need a triage classifier at all. You might need a template library and a form that fills it in. That would still be most of the win.

What we are building next

The system is now core to how we operate, and it is not finished. We are working on classification that learns from completed jobs, pricing that accounts for parts availability and current workload, follow-up that nudges customers who viewed a quote without accepting, and voice integration so someone can ring, describe the fault, and get a quote by SMS before the call ends.

Every one follows the same principle: remove friction, speed up the decision, and leave the judgement with a person.

If quoting, triage or repetitive admin is the bottleneck in your business, this is exactly the work we do for clients through our workflow automation and CRM service. If you would rather find out where the biggest gain sits before committing to anything, our AI Opportunity Assessment maps your current process and gives you a written plan for £99.99, and you keep the report either way. If you just want a straight answer about whether this would work for you, ask us.