How to build a sales pipeline from scratch (a step-by-step guide)

Build a sales pipeline backwards from a meeting target, and build it in an order that respects three dependencies you cannot reorder: a sending domain needs weeks of calendar time before it needs work, stages have to exist before leads can be placed in them, and the ICP has to be written in fields before a list can be filtered on it. Most first pipelines fail not because a step was done badly but because the steps were done in the wrong order, and by the time that shows up the evidence needed to diagnose it was never recorded.

Start from the meeting target and work upwards

Do the arithmetic before you do anything else, because it tells you how big the first list needs to be and whether the whole plan is viable at all.

Pick the number of qualified meetings a month the business actually needs. Then walk the funnel upwards with placeholder rates, marking each as an assumption you will replace with your own measurement. Illustrative arithmetic, not a benchmark: if you want 10 meetings a month, and roughly a third of positive replies become meetings, you need about 30 positive replies. If a fifth of all replies are positive, that is 150 replies. If replies run at 3% of delivered messages, that is 5,000 delivered messages a month.

Now look at that number and decide whether you believe it. Five thousand delivered messages a month is roughly 250 a working day, which is roughly ten mailboxes at a conservative per-mailbox rate. If that is more infrastructure than you intend to run, the honest conclusion is not to send harder — it is that the reply rate assumption has to be better, which means the list has to be tighter, which is a targeting decision, not a volume one.

Write the assumptions down with the date. In six weeks you will have real numbers for two of the four, and the difference between what you assumed and what you measured is the most useful thing you will learn in the first quarter.

Three dependencies that fix the order

The domain clock runs whether you are working or not. A newly registered sending domain has no history, and history is what mailbox providers use. Google's Email sender guidelines require senders of more than 5,000 messages a day to Gmail accounts to authenticate with SPF, DKIM and DMARC and keep the spam rate reported in Postmaster Tools below 0.30%, effective 1 February 2024; Microsoft began routing high-volume non-compliant domains to the Junk folder from 5 May 2025, rejecting with 550 5.7.515 Access denied, sending domain does not meet the required authentication level. Publishing records takes an afternoon. Building the sending history takes weeks. Start on day one.

Stages must exist before the first lead is placed. Not because the software demands it, but because a lead placed before the stages are agreed gets placed somewhere arbitrary, and every report you run afterwards is computed over that arbitrary placement.

The ICP must be expressed in fields before the list is built. An ICP written as a paragraph cannot filter anything. It has to become industry values, country codes, headcount bounds and title strings, because those are what a query can act on.

Everything else in the build can move around. These three cannot.

Days 1 to 7: the parts that need calendar time

Register the sending domain — a separate one from your primary company domain, so a deliverability problem in outbound cannot damage the mail your finance team sends. Publish SPF, DKIM and DMARC. Start DMARC at p=none so you receive reports without risking legitimate mail, and read the reports before tightening. Connect the mailboxes and start warmup at genuinely low volume.

While that clock runs, do the things that need no infrastructure. Write down what you sell in one paragraph, who it is for, and what it does not do. That last item is worth more than it looks: the exclusion list is what stops the first list filling with plausible non-buyers.

If you want the deliverability groundwork in detail rather than in summary, setting up a cold email domain covers the records and the sequence properly, and the SPF, DKIM and DMARC generator will produce the records themselves.

Days 3 to 10: define the stages before any lead exists

A stage is a gate with an entry test, not a label on a column. Six is enough to start. Autocloz seeds a default pipeline named "Sales Pipeline" on first use of the deals module with exactly six stages and a default probability on each: Lead at 10, Qualified at 25, Proposal at 50, Negotiation at 75, Won at 100 and Lost at 0. Won and Lost carry is_won and is_lost flags rather than being recognised by name, which is what lets reporting distinguish closed from open without string-matching a stage somebody later renamed.

Two decisions matter more than the names.

Write the entry test for each stage as a sentence a second person could apply. "Qualified" means nothing on its own. "The person we spoke to confirmed the problem, has budget authority or named who does, and agreed a next step with a date" is a test. If two reps would place the same deal in different stages, the stage definition is the thing that is broken.

Decide the close-reason vocabulary now. Autocloz stores a machine key on the deal — price, champion, timing, competitor, no_decision, budget — alongside the rep's free text. Agreeing that list in week two costs ten minutes. Agreeing it in month six means every deal closed before then is uncategorised, and the analysis you wanted to run is impossible. Sales pipeline stages explained goes deeper into what each gate has to prove.

Days 8 to 14: the ICP in fields, then a deliberately small list

Turn the paragraph into fields. Concretely: which industry values, which country codes, what headcount range, which job titles. In Autocloz an autopilot ICP is expressed as exactly five keys — titles, industries, locations, headcount_min and headcount_max — and the filter each one produces is worth knowing before you rely on it. Industry is matched exactly and case-sensitively, so "SaaS" and "saas" are two different filters. Locations are upper-cased and matched against country codes on either the person or the company. Titles go through a token-subset match with an alias table, so "CHRO" matches "Chief Human Resources Officer" and "HR Director" does not match "Threading Systems Director".

The headcount filter has a behaviour worth stating plainly, because it changes what you get: a company whose size is unknown passes the bound rather than failing it. The clause reads as "size is null, or size is within range". That is the right default for a list you would rather over-include than starve, and it is the wrong assumption if you believed the filter was strict.

Then build 100 to 300 rows. Not more. A small list is not a compromise; it is the only size at which you can read every reply yourself, which is where the actual learning happens in month one. Building a targeted lead list covers the sourcing and hygiene end of this.

Days 12 to 21: one sequence, two channels, small

Build a single sequence. Two channels, not five — coverage beats breadth at this stage, and a step that has no handle to send to is a gap rather than a touch. Four email steps and one LinkedIn step across two weeks is a complete first sequence.

Three things to get right and nothing else:

• One ask per message. Stacked asks reduce the reply rate for a mechanical reason: they require a decision about which one to answer.
• A stop condition on reply. A follow-up sent after someone answered is the single most damaging automation failure, because it is the one the recipient definitely notices.
• A close-out step. The last message says you are stopping. It reliably produces replies from people who meant to answer earlier, and it makes the silence at the end deliberate rather than accidental.

Autocloz's free plan covers 5 users and 10 mailboxes with warmup and DMARC monitoring built in, which is enough to run this entire build — start free and set the stages up before the first import.

Days 21 to 30: the review that turns a queue into a pipeline

A pipeline is a thing you look at on a schedule. Without the schedule, it is a list of deals that ages.

Book thirty minutes weekly and answer four questions in this order. How many new leads entered? Where did leads leave, and from which stage? Which deals have not moved in fourteen days? What is the worst-converting transition?

That last one is the whole point of the review. Every pipeline has one transition that leaks worse than the others, and improving it is worth more than improving the other five combined, because the funnel multiplies. Doubling a 5% reply rate and doubling a 40% meeting-to-opportunity rate produce very different amounts of pipeline for the same effort.

Set a rotting threshold at the same time. A deal that has sat in one stage for longer than your median cycle in that stage is either dead or neglected, and either answer is actionable.

A worked example of the first month

Illustrative arithmetic, to show the shape rather than to predict anything.

Two hundred rows enter the sequence. Suppose 6% bounce, leaving 188 delivered. Suppose 4% reply across the whole sequence — 8 replies. Of those, suppose three are positive, three are "not now" and two are opt-outs.

Three positive replies against a target of ten meetings a month looks like a failure. It is not, and reading it as one is the most common early mistake. What the run bought you is four measurements you did not have: your real bounce rate, your real reply rate, the ratio of positive to negative, and a list of the objections people actually raised. Those four now replace four of the assumptions in the arithmetic at the top, and the second run is sized from measured numbers rather than guesses.

The decision rule at this fork: if bounce is above about 5%, fix list hygiene before anything else, because every other number is being computed over a broken denominator. If bounce is fine and reply rate is near zero, the problem is targeting or offer, not copy. If reply rate is reasonable but positives are near zero, the message is reaching the right inbox and describing the wrong problem.

The five ways a first pipeline dies

• Volume before deliverability. Ten thousand sends from a three-week-old domain produces a reputation problem that takes months to unwind.
• Stages invented after the leads arrived. Every report is then computed over placements nobody agreed.
• No stop-on-reply. The automation contacts people who already answered, and those are the people most likely to tell others.
• No review cadence. Deals age quietly; nobody notices until the quarter ends.
• Reading a close rate at thirty days. With eight replies, the confidence interval on any rate you compute is wider than the decision you want to make with it.

What a new pipeline cannot tell you, and what Autocloz does not do

Three limits are worth knowing before you rely on the reporting.

Average days-to-close is computed over won deals only. In Autocloz it is the average of closed_at minus created_at across deals in a stage flagged is_won. Deals still open — including the slow ones that will eventually close — are not in the denominator. That makes the figure right-censored, and it will read optimistically low in any month where the long deals have not landed yet. It is a useful number as long as you know it answers "how long did the deals that closed take", not "how long does a deal take".

Stage history is not recoverable after the fact. The stage_entered_at timestamp is re-stamped on every transition, so a deal tells you when it entered its current stage and nothing about the stages before it. Time-in-stage across a whole cycle is therefore not reconstructible from the deal row later. If that analysis matters to you, capture the transitions as you go.

One heat metric on the deal record is inert. engagement_score is documented as a 0-to-100 number computed nightly from recent activity, and the only code that touches it reads it — nothing writes it. It sits at its default of 0 and stays there. Do not build a prioritisation rule on that field, and do not read a Kanban card's heat as a signal until something populates it.

Autocloz also does not discover leads for you as part of this build, and it does not decide who to target. The free CRM with deals and stages will hold the pipeline, run the sequence and record the events; the judgement about who belongs in it stays yours. Teams weighing this against a per-seat incumbent will find the pricing arithmetic laid out on the Autocloz and HubSpot comparison, which is a more honest place to make that decision than a feature list.