Hi Friends,

Last week I shared the sales narrative I wrote by hand for Allenix. (Read that one here.)

This week is the building part.

We are in the trenches right now with companies putting AI into real 8 and 9 figure operations. Every one showed up with a list of 30 things they wanted automated.

The valuable work is not the building. It is walking those processes and finding out which items deserve to be on the list.

Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027. Escalating cost, unclear value, weak controls. Not one of those is a technology problem.

So here are the 5 questions we ask before anyone builds. About 10 minutes each. They take a list of 30 maybes and leave you with the 2 that actually move money.

Every process is a pile of small steps masked as one. Write the process steps down.

A client wanted to automate their sales proposals. That is 6 steps: pricing, scope of work, matching case studies, discount approval, assemble, review and send. Build it as one thing and you own a 6 month project where nothing works until everything works. Break it apart and the scope of work ships in 2 weeks, because that is where the saved hours are.

Now the part most people skip.

With the steps written down, ask which ones should exist at all. Steps survive for reasons that expired. Someone left. Two systems could not talk. A customer complained once in 2019. In that proposal process, the discount approval was a rule that belonged in the CRM, not an AI project at all.

Then ask which of the survivors collapse into each other. Two approvals become one. Three handoffs become a shared view. Five steps become three before anyone writes a prompt.

Automating a step you should have deleted is the most expensive thing you can do with AI. You pay to build it, then pay monthly to keep it alive, and now it lives in software where nobody questions it again.

Delete first. Reimagine next. Then AI what survives.

Name every input the work needs. Where it lives, who owns it, and how fresh it has to be.

Access. The software needs to reach the files, the records, and the emails on its own, with its own login. The version we see most: a vendor emails a price list every month and somebody retypes it so the software can read it. That is not automation. There is still a person inside doing the work, except now it sits somewhere nobody counts.

People. If a human has to hand something over for this to run, that human is part of the automation. Their sick day is your outage.

Timing. Every input has a shelf life and almost nobody writes it down. Your quoting tool prices off a vendor sheet that expired in June. Nothing breaks, no alert fires, and the quote looks like every quote you have ever sent. It goes to a customer.

This is the question that kills the most builds, and Gartner expects 60% of AI projects without AI-ready data to be abandoned through 2026. If the answer is "we would need to export it," or "it lives in her inbox," you are funding an integration project masked as AI.

Something has to start it. A file landing in a folder, a deal changing stage, a call ending, a clock hitting 6am. Name it before you build.

If the only thing that starts it is a person remembering, it runs for 2 weeks and quietly stops. That is the failure nobody catches, because it produces no error. Nothing ran, so nothing complained. Weeks later somebody asks how the new tool is going and learns it died in March.

So name a check and an owner. The check is a weekly count of how many items should have been handled against how many were. The owner is whoever reads that count in month 7, when everyone who built it has moved on.

An automation nobody owns is a liability with a subscription.

Write down what a great output looks like before you build. If you cannot describe it, you cannot check it, and you cannot tell the model what you want either.

Then sort it, because there are two kinds and they behave nothing alike.

Some outputs have one right answer. A total, a mapped field, a date calculated from a contract. You can test these. Write the rule, run real cases, count the misses. Once it passes, it keeps passing.

Some have no single right answer. A summary, a scope of work, a recommended next step. You cannot test these, only judge them, and judging means a person reads them. They will be right most of the time and confidently wrong the rest, and the wrong ones look exactly like the right ones.

Most teams price the second kind like the first. That is where the payback disappears.

So decide the “good” upfront. Which fields you would eyeball, how many you would sample, who does it. If it reaches a customer, one named person approves every one.

Then ask the question that kills more of these than anything else. If a person needs an hour to trust an output that took an hour to produce, you did not automate anything. You moved it and added a subscription.

Everything above decides whether this is buildable. This decides whether it is worth it, and it goes last because the math needs those answers: how many steps survived, whether you need an integration project, how much review the output takes.

Start with which number moves. We bucket everything into three.

Then the math, and it should take an afternoon, not a quarter. Ask the team for directionally accurate numbers. Time per run, times how often it happens, times $150 an hour. Do not let anyone spend a week refining that number.

It’s also important to see where the saved hours actually go. Example, if you save the company 200 hours, but spread across 12 people and everybody is slightly less busy.

It’s important to differentiate absorbed hours vs true savings.

Take the things on your company AI initiatives and put every one through all 5 questions.

Two things happen when you do. The few initiatives worth real money separate themselves from the noise. And your team stops arguing about what to build first, because the answer is sitting on the page in front of them.

'Til next time,

--Ali

P.S. Want the Excel version we use? Hit reply and I will send it over. It is the same file we work through with clients, all 5 questions with the payback math and a worked example.

About Me: I am Ali Mamujee. I run Allenix, where we build AI revenue engines for B2B companies between $5 million and $25 million with complex sales cycles. Kids back in school. Counting down the days for Houston Texans and Texas Longhorns football.