Three Sales Ops Automations, With What They Actually Cost

TL;DR: $6,500 to build all three, $270/month to run, roughly 50 hours a month of manual work replaced. Payback around two months. The numbers are the easy part - the section on what went wrong in each build is the part worth reading before you commission one.

I am Ivan, I build these for a living. Rather than another piece on what AI could theoretically do for sales operations, here are three systems with real numbers, including the parts that did not go smoothly.

1. Lead enrichment and scoring, B2B SaaS

The client was taking 200+ inbound leads a month through a website form, and the sales team was spending two hours a day researching each one before deciding whether it was worth a reply. Two hours a day, to sort a list.

What we built: a form submission fires a workflow. A data provider fills in company details. A model reads the company description and whatever the person typed in the form, and assigns an ICP score from one to five with a written reason. High scores route to a rep with a notification. Low scores go to nurture. Nothing is discarded.

Numbers:

• Manual research: from 2+ hours a day to about 20 minutes

• Rep attention: spent on the top fifth by score rather than in arrival order

• Build: 6 days, $3,500

• Running: $180/month across the data provider, the model and the platform

What went wrong

The first version scored confidently on leads it knew almost nothing about, because an empty company description still produces a number. Reps lost trust in the score within a week, which is very hard to win back. We rebuilt it so the model must state its confidence separately, and anything below the threshold is marked unscored and goes to a human rather than getting a plausible-looking two. The lesson generalises: a scoring system that cannot say I do not know will be ignored.

2. Pipeline reporting, logistics

The operations manager spent three to four hours every Friday building a pipeline report in a spreadsheet for the leadership meeting. Copy, paste, recalculate conversion by hand, write the summary.

What we built: a scheduled job at seven on Friday pulls the pipeline, computes stage conversion, average deal size, performance by rep and movement against last week, and a model writes a short plain-English summary of what changed. Data and narrative land in the team channel before eight.

Numbers:

• Manual reporting: three to four hours to zero

• Consistency: computed the same way every week, so week-to-week comparison finally means something

• Build: 3 days, $1,800

• Running: $60/month

What went wrong

Two weeks in, the report went out with a large drop that was not real: an API token had expired and one source returned nothing, which the arithmetic dutifully treated as zero. Leadership spent a morning on a crisis that did not exist. Now a missing source makes the report say it could not be built, loudly, instead of publishing a confident zero. Any reporting automation without that guard is a false-alarm generator.

3. Stalled deal alerts

Reps were forgetting to follow up on deals that went quiet. Deals sat in negotiation for thirty days with no activity and revenue leaked without anyone deciding to let it go.

What we built: a daily check across every open deal. Seven days with no email, call or note and a task appears for the owner with the context attached: last interaction, deal size, days in stage. Fourteen days and the manager sees it too.

Numbers:

• Deals sitting 30+ days with no activity: from about 40% of pipeline to 8%

• Build: 2 days, $1,200

• Running: $30/month

What went wrong, and the one I would not build again

The manager alert at fourteen days was a mistake in its first form. It read as surveillance, reps started logging token activity to reset the clock, and the data got worse rather than better. We changed it to a weekly summary to the manager of how many deals were ageing, with no names, and the behaviour stopped immediately.

If you build this, make the alert serve the rep and not the audit trail. Anything that is experienced as monitoring will be defeated by the people it monitors, and they will be creative about it.

The arithmetic across all three

$6,500 to build, $270/month to run, roughly 50 hours a month of manual work removed across the teams. At a fully loaded $60/hour that is about $3,000/month of labour, so payback lands around two months and everything after is recovered capacity.

One honest caveat on that figure. Fifty hours removed is not fifty hours of new output; nobody converts recovered time at 100%. What actually happened is that the reps made more calls and the operations manager stopped losing Friday. That is worth the money, but it is not the same as hiring 0.3 of a person.

The pattern in all three

None of these is clever. Each takes a task that is consistent, high frequency and low judgment, and moves it off a person, leaving the judgment where it was. The scoring system does not decide who to call, it decides what order to look in. The report does not analyse, it assembles. The alert does not chase, it reminds.

Every one of the three failures above came from the system acting more confident than its inputs justified. That is the failure mode to design against, and it is worth asking any vendor how they handle it before you ask what it costs.

Want the equivalent numbers for your own sales operation? The audit at 2pizza.team/audit takes two minutes, no call, and gives you a breakdown for your specific stack.

Originally published at 2pizza.team. We build AI and automation systems for small teams - fixed price, two to six weeks. See the work.