You know which work eats your week, and you have a list of ten things you would like a machine to do. The first process you automate decides whether the second one gets a budget, so it is worth choosing on numbers rather than on which task annoys you most.
This post is the scoring method I use on a first call. You can run it yourself with a spreadsheet and an afternoon.
Why the first pick goes wrong
Most people pick the process that hurts the most. That is usually the one with the most exceptions, the most people involved, and the least written down. It is painful because it is messy, and messy processes are the slowest to automate and the hardest to measure.
The second common mistake is picking the process that is easiest to automate. A report that takes someone twenty minutes on a Friday is easy, and automating it saves twenty minutes. Nobody will fund the next project on that result.
The process you want sits between those two: enough hours to matter, rules clear enough to write down, and a result you can count.
Score each process on five questions
Take every process on your list and answer these. Write numbers, not impressions.
Hours a month: how many person-hours does it take today? Count everyone who touches it, including the person who checks the work and the person who fixes the mistakes. If you do not know, have the people who do it log a week.
Volume and repetition: how many times does it happen? A process that runs 2,000 times a month with the same steps each time is a better candidate than one that runs four times a month with a different shape each time.
Rules you can write: could you hand the process to a new hire on two pages? If the honest answer is "it depends on who is asking," the rules are not written yet, and writing them is the first part of the project.
What the input looks like: is it already a record in a system, or does someone read a document, a handwritten slip, an email, or a voice on the phone? This does not rule a process out. It tells you what kind of build it needs.
What a mistake costs: a wrong number in a weekly report costs little. A wrong amount posted to a ledger or a missed payroll deadline costs a lot. High-cost processes are still good candidates, but they need a review step, and that changes the design.
Turn the scores into a ranking
Multiply hours a month by the loaded hourly cost of the people doing it. That is the monthly cost of the process as it runs today. Then sort.
At the top of the list, look for the process that has high cost, high repetition and rules you can write. That is your first project. Leave the process with the highest cost but no written rules for second place; you will write its rules while the first one runs.
A quick check before you commit: can you describe, in one sentence, the number that will tell you it worked? "Slips entered per hour." "Calls answered in the first ring." "Bills posted without a person touching them." If you cannot name the number, pick a different process.
Match the process to the kind of build
The input question from the scoring above tells you what you are buying. In our work it sorts into four kinds.
Someone reads and types: a data-entry or bookkeeping agent. It reads the document, applies your rules, posts what passes to your CRM, ERP, spreadsheet or accounting system, and puts what fails in a review queue with the reason it stopped.
Someone talks on the phone: a voice agent. Ours answers inbound calls as a receptionist and places outbound calls as a cold caller. Same system, described by the job. It holds a real two-way conversation, handles interruptions, responds in under a second, transfers a live call to a person, books on your calendar, and writes the note.
Two systems that cannot see each other: custom software that reads and writes both, with field mapping, validation and a log of every record it touched.
A record moves between two connected systems on a fixed rule: a workflow tool or a short script. This is the cheapest kind, and if your top process fits here, buy that and save the build fee.
What a good first project looked like
A delivery company in Dallas-Fort Worth had about 2,000 handwritten delivery slips, each with about twenty fields, and 72 hours before drivers' pay was due. It scored high on every question above: a large, fixed volume, rules the office could state, a clear deadline, and a number to measure against.
We built a data-entry agent that watched the folder the company's own scanner fed. It read each page, four slips at a time, and set aside the pages it could not read instead of guessing. About 30 minutes later there were more than 1,900 rows in a spreadsheet, checked against the paper, and payroll went out on time.
The point of the example is the scoring, not the speed. The process had a count at the start, a count at the end, and a deadline. Anyone could see whether it worked.
On the phone side, our proven deployment is outbound qualification and meeting setting in oil and gas. The voice agent has placed more than 30,000 outbound calls since December 2025. Other industries are a fit rather than a proof, and on the call I will tell you which one yours is.
Plan the first month after launch
An automated process is not finished on the day it goes live. Plan for this.
Week one: the rules are conservative on purpose, so more records go to the review queue than you expect. A person reads the queue every day.
Weeks two to four: most of what lands in the queue turns out to follow a pattern. Each pattern becomes a rule, and the queue gets shorter.
End of month one: compare the number you named before the project against the number now. Hours spent, records processed, calls answered, errors found in a sample. That comparison is the case for your second process.
Every record our agents post carries where it came from, which rules it matched, and when it posted, so the person checking the work can sample any record and see why it went through.
What it costs
Prices are starting points. Bookkeeping and data-entry agents are from $5,000, then from $500 a month. The AI receptionist and answering service is a build fee from $2,500, then from $500 a month, and for the receptionist a busy month costs the same as a quiet one. The cold caller and appointment setter is from $5,000, then from $2,500 a month. Custom software and one-off jobs, like the 2,000 slips, are priced per project, from $10,000.
The monthly fee covers AI model usage, monitoring, fixes, and the changes you ask for, plus the phone line and call minutes on the voice agents. More systems to connect, more volume, more complicated rules, and more agents or phone lines raise a quote. The pricing section of our FAQ covers the rest.
Set the price against the monthly cost you worked out in the ranking step. That is the comparison that matters.
Bring the list
If you have scored your processes, bring the spreadsheet. If you have not, bring the list and we will score it together. Either way, we will tell you which process to automate first, what kind of build it needs, and the number you should measure it by.
Book a call and bring the processes you would take off someone's desk first.