AI sales automation: what to automate from enquiry to quote

In the proposals that reach sales departments, “sales automation” usually means one thing: bots that write to strangers by themselves. Inside the company, revenue leaks somewhere else entirely — between the enquiry that arrived and the quote that went out. This article answers what AI sales automation can cover across that stretch, who approves what, and how you will know it worked. We read AI in sales narrowly here: as the work around selling, not as a replacement for a salesperson.

We are writing for B2B companies that quote — distributors, manufacturers with a quotation desk, service firms with a handful or a dozen or so salespeople. Below: how to measure this stretch, five stages in one template, a worked example with its assumptions stated, the data and GDPR decisions, and the situations where this is the wrong project. Model and platform prices were read from the vendors’ own pages on 6 September 2026 and checked again on 9 October 2026 (unchanged; Claude Sonnet 5 and Haiku 4.5 are now legacy models), because a figure without a date is useless here.

What AI sales automation is — and what it is not

An operational definition: the system reads, assembles and prepares; the salesperson decides and sends. Everything that goes out to a customer or changes data in a system starts in a mode that requires human approval — the only way to know, from day one, how often the system gets it wrong.

Two distinctions save the most money. First: the classic automation your CRM already includes in its subscription — reminders, opportunity assignment, templates. Switch that on first. AI only enters where rules run out: reading an unstructured enquiry, or matching price list lines to a description like “same as last time, but the stainless version”.

Second: B2B sales automation “from the acquisition side” — automated email series to people who never contacted you, and automatic rating of prospects (lead scoring) — is a different category of risk; we come back to it at the end. Here we deal only with enquiries that have already arrived: AI in sales, read this way, does not bring new customers, it recovers revenue lost on the way to the quote. Which processes are suitable at all — the five conditions for a good candidate — is set out on our page about AI automation in a company.

Measure the stretch from enquiry to quote before you automate anything

One number describes this stretch best: hours from the enquiry arriving to the quote going out. Use the median (the middle value when you line the times up from shortest to longest), not the mean — one enquiry that sat in an inbox for two weeks will inflate the mean into uselessness.

Measuring it needs no tooling: for two weeks, record in a spreadsheet when each enquiry arrived and when the quote went out. Add two companion numbers: the share of enquiries with no reply within 24 hours, and the share of quotes that came back for revision — wrong price, missing line.

Then write down a sentence you can check unambiguously after the rollout: “Today the median time from enquiry to quote is X hours. After implementation it should be Y.” We set this step out at greater length in the guide on how to implement AI in a company. If nobody can extract those dates, that is your first result: the project starts with measurement, not with a model.

Five stages between the enquiry and the quote

The same stages exist in every company that quotes; what differs is where the median grows most. We describe each with the same template — input, output, who approves, typical mistake, metric — because a description like that can be handed to your own IT team or to any contractor. Four of the five are the sales versions of processes 1, 5, 7 and 9 from our piece on examples of AI automation; here we go a level lower.

1. Qualifying and logging the enquiry

Qualifying enquiries is the first stage, because without it there are no others — and today the enquiry waits for someone to retype it.

• Input. The body of the message or form, attachments, sender, channel.
• Output. An opportunity in the CRM with a full set of fields — company, contact, product or service, quantity, date, source — and a suggested priority.
• Who approves. The salesperson on duty, with one click; they correct the fields if needed.
• Typical mistake. A duplicate customer or the wrong salesperson assigned — visible at once and reversible.
• Metric. Minutes from the enquiry arriving to the CRM entry.

Priority follows written rules — value, deadline, existing customer — not a model trained on history: rules are cheaper, predictable and can be explained.

2. Clarifying questions

A large share of enquiries are incomplete: no quantity, no version, no date. The salesperson notices when they sit down to quote — two days later, by which time the customer has asked your competitors.

• Input. The enquiry from stage 1 and the list of fields your company needs in order to quote.
• Output. A list of what is missing and a draft of a short email with two to four questions.
• Who approves. The salesperson: reads, corrects, and sends from their own inbox under their own name.
• Typical mistake. Asking about something that was in the attachment — visible on reading the draft.
• Metric. The share of enquiries that come back with further gaps after the first reply.

3. Quote automation: the quote as a draft for approval

In most companies that quote, this is where the median grows most — and where the temptation to hand the pricing decision to a machine is strongest. Quote automation does not mean quotes that generate and send themselves; it means that preparing a quote starts from a draft rather than from an empty template.

• Input. The price list (a file or the ERP), the customer commercial terms, three to five comparable historical orders, the quote template.
• Output. A draft quote with lines taken from the price list and the uncertain points marked: “no equivalent found for line X”.
• Who approves. The salesperson decides the price and the discount, and sends. The pricing decision is never automatic — every ruling has a person who answers for it.
• Typical mistake. The wrong price list line or an out-of-date price — catchable while reviewing the draft, before the quote goes out.
• Metric. Hours from enquiry to quote — the same ones you measured before starting — and the share of quotes coming back for revision, which must not rise after implementation.

Precondition: the price list and the discount rules must be written down. If they live in the salespeople’s heads, the project starts by writing them down. The same mechanism handles the repeatable questions in requests for quotation and RFPs — with a list of the questions that need a human decision.

4. AI in the CRM: the entry after the call, instead of evening notes

The CRM loses to the phone: entries get made in the evening, from memory, or never. AI in the CRM in this form predicts nothing — it prepares the entry.

• Input. The email thread, or a recording or summary of the call.
• Output. A proposed entry: the contact, what was agreed, the amounts, the next step and its date.
• Who approves. The salesperson, with one click, right after the call.
• Typical mistake. A misread next step or date — visible on review; if it slips through, stage 5 exposes it.
• Metric. The share of open opportunities with a current, dated next step.

Two preconditions: a CRM that can take data automatically (through an API or a file import), and call recording and correspondence processing settled under GDPR — before you build, not after; more on that below.

5. A follow-up that does not get lost

Follow-ups do not go out because twenty new enquiries arrived, not because the salesperson does not want to send them. This stage looks after the conversations already under way.

• Input. Opportunities with no movement for an agreed number of days, and what was agreed in the last CRM entry.
• Output. A reminder for the salesperson and a draft follow-up that refers to specific agreed points, not to “just checking in”.
• Who approves. The salesperson, always — and it goes out from their own address.
• Typical mistake. A follow-up after the customer replied through another channel — which is why stages 4 and 5 only work together.
• Metric. The share of opportunities with a follow-up sent on time.

A boundary worth writing into the requirements: this is not an automated email series. Nothing goes out without a person, and nothing goes to anyone who did not write to you first.

A worked example: a distributor with four salespeople

The example is illustrative — it is not a client implementation, and every figure is an assumption you can swap for your own. We do not quote “industry benchmarks” for savings; the ones circulating online have no stated method.

Assumptions. Four salespeople, 60 enquiries a week, so roughly 260 a month (60 × 52 / 12). Baseline measurement: median from arrival to quote 31 hours, 22% of enquiries with no reply within 24 hours, 15% of quotes coming back for revision. Repetitive work per enquiry: 10 minutes reading and logging in the CRM, 25 minutes gathering prices, terms and history, 10 minutes formatting, 5 minutes updating the CRM — 50 minutes in total. Per month: 260 × 50 / 60 ≈ 217 hours.

After a pilot in “the system prepares, the salesperson approves” mode, what remains is review, correction and sending — call it 15 minutes, so about 65 hours a month. Recovered: about 150 hours a month. We present that as hours and as a target for the median — “from 31 hours to 6” can be checked after a quarter — not as headcount.

Running model cost. A token is the unit a provider counts text and charges in — roughly part of a word. Assume about 30,000 input tokens and 5,000 output tokens per enquiry across the three steps combined — an assumption to verify in the pilot. At Claude Sonnet 5 pricing — 2 USD per million input tokens and 10 USD per million output, read on 6 September 2026 — that is about 0.11 USD per enquiry and about 29 USD a month; on Claude Haiku 4.5 (1 USD and 5 USD per million), about 14 USD. The model cost is a rounding error; it is not what the project price turns on.

Platform. n8n Cloud Starter: €20 a month on annual billing for 2,500 executions (an execution is one run of the automation); assuming three executions per enquiry, 260 enquiries comes to about 780, with headroom. The alternative is the Community edition on your own server: the licence is free, but the server and its upkeep are yours.

Build and integration. No price range here: the cost is set by how many systems have to be joined up (CRM, email, price list or ERP, template), the state of the data and the level of certainty required — we broke the components down in our piece on how much an AI implementation costs. Plan maintenance as a standing line item: a monthly review of a sample of quotes, and tests after each model version change.

The formula for you: (hours recovered × your cost per salesperson hour) − (subscriptions + maintenance), set against the build cost, gives the payback in months. The counter-example by the same method: a company with 25 enquiries a month recovers about 15 hours. That rarely covers integration and maintenance — switch on your CRM reminders and come back to this when enquiries run into the hundreds.

Customer data, price lists and GDPR — what to settle before integrating

Three decisions before the first systems are connected.

Which data leaves the company at all. Enquiries and order history contain personal data belonging to your customers’ employees, and the model needs neither names nor phone numbers to produce a quote. Narrow the scope in the data preparation step; data that was never sent needs no guarantees at all. Replacing names with identifiers is sound data minimisation, but pseudonymisation does not take a project outside GDPR — pseudonymised data is still personal data.

To whom, and where. The model vendor, the countries or regions where it processes the data, whether it uses your content to train models, and the data processing agreement. The business tiers of the large vendors’ services usually exclude training on customer content, the consumer tiers often do not — always check the terms of the specific service. The model vendor’s role is established separately for each operation; who is the controller and who the processor, and what to ask each vendor, is set out on our page about security and GDPR.

Price lists and commercial terms. A trade secret rather than personal data, but the test is the same: does the whole price list have to leave the company, or are the lines matching the enquiry enough?

Two separate matters. Recording calls in stage 4 is a decision to take with a lawyer, and carries a duty to inform the other party. And the EU AI Act: as long as the salesperson reads, corrects and sends, the message is theirs. If, after the pilot, some category of reply were to go out without a person, the disclosure duty under Article 50 of the AI Act applies — in force from 2 August 2026 (as at 6 September 2026) — as does our own rule that an automated system always identifies itself; we keep the current dates on our page about the AI Act technical compliance sprint. This describes the technical and organisational side, not legal advice — the legal assessment belongs to your lawyer or data protection officer.

Where we draw the line in sales automation

Most offers labelled “sales automation” promise mass outreach to cold contacts, which is exactly what is missing here. We build systems that protect your domain, your brand and your customer relationships, so you will not find in them:

• Cold-email and bulk mailing tools. Legal risk — consent, the electronic communications rules, liability sitting with the sender — and reputational risk: a company email domain whose mail then lands in spam and a “spammer” reputation outlast the campaign. This rule applies to every project, with no exceptions.
• Lead scoring with no data history. The model has to learn from past opportunities described consistently, with an outcome. If the CRM does not have that, a sensible project starts by tidying the data and says plainly that scoring comes later — it does not sell data collection under the name of prediction.
• Bots impersonating a salesperson. Anything that talks to a customer by itself identifies itself as an automated system and has a route to a human. Nothing in this article talks by itself.
• Decisions on price, discount and negotiation. The system prepares, a person signs; “automatic pricing” is a ruling nobody is answerable for.
• “AI will close the sale for you”. AI removes the repetitive work around the sales conversation, not the conversation.

The principles behind every implementation are on our standards page.

When AI sales automation is the wrong choice

1. Too few enquiries. A few dozen a month recover a dozen or so hours, and the cost of integration is largely fixed.
2. The price list and discount rules are not written down. That is a project to do first, not alongside — an organisational one, not an IT one.
3. Every quote is bespoke engineering. AI can assemble the inputs, but it will not shorten the estimator’s work; the stretch to shorten is somewhere else.
4. A CRM nobody uses — automation will only entrench the emptiness — or a CRM with no API and no file import.
5. The salesperson’s work is not the bottleneck. If the quote is waiting on the boss’s discount decision or on a date from production, you will speed up the part that was never the problem.
6. A ready-made tool exists. If the reminders, assignment and templates in your CRM subscription are enough, use them.

How to work this out for yourself: a checklist

1. Export the last 100 enquiries with the dates they arrived and the dates quotes went out; if you cannot, that is your first result.
2. Calculate the median hours from enquiry to quote and the share of enquiries with no reply within 24 hours.
3. For a week, note the minutes of repetitive work per enquiry.
4. Check whether the price list and the discount rules exist in a file that can be opened without a salesperson.
5. List the systems — where enquiries arrive, where the quote is produced, where the CRM is — and whether they have an interface for connecting other software (an API), or a file import.
6. Settle which mistake is acceptable, and how you will see it before the customer does.
7. Write down “today X hours, after implementation Y” and the name of the person answerable for that number.

What next

Start with the stage where the median grows most — usually stage 3, but the first project should be chosen by measurement, not intuition. Which stage we would start with in your case, and what an AI project for a sales team costs to build, we settle in the first conversation; the service from the supplier’s side is described on our AI in the sales team page, and which processes are suitable at all on our page about AI automation in a company.

Tell us what the road from enquiry to quote looks like in your company today. We will tell you which stage we would start with and what can be measured in that first stage.

Sources, as at 9 October 2026

• Claude pricing (Anthropic) — Claude Sonnet 5 ($2 / $10) and Claude Haiku 4.5 ($1 / $5) per million input and output tokens; both are now listed as legacy models, and the current Sonnet 5.5 costs the same as Sonnet 5; read on 9 October 2026.
• n8n pricing — Starter at €20 a month billed annually for 2,500 executions; read on 9 October 2026.
• Regulation (EU) 2024/1689 (AI Act) — Article 50; read on 9 October 2026.