Omar Trejo

AI engineering for ambitious companies. Custom AI designed, built and run for you, at higher quality, faster and at lower cost. Founder, ML LABS.

[Omar Trejo @otrenav]

Your system speeds up the invoice packet. The contractual billing cycle, disputed charges and collections all sit after it. Count the part it changes directly; keep the rest as a hypothesis until your own evidence supports it.

Which of your business case numbers does it change?

[Omar Trejo @otrenav]

One quote prices AI seats per person, the other prices a build, and each is pitched to you as the whole decision. Price both over three years, then decide which layer you need to own: https://t.co/AKqMaDX98S

[Omar Trejo @otrenav]

Deployment metrics can look healthy while risk grows in what you ship. In a 2022 Scientific Reports study, 91% of 128 model-dataset pairs degraded over time, and no DORA metric would have moved. Here is the check to run beside the dashboard. https://t.co/GTmkAsZiFx

[Omar Trejo @otrenav]

@karrisaarinen Telling whether the work is good gets easier when a team keeps a handful of past examples it agrees were good and checks new output against them, whatever tool made it. The tool can change; the bar stays put.

[Omar Trejo @otrenav]

@KennethCassel Moving from sub-assemblies to the full robot multiplies the part numbers you track. Having each supplier's ship date feed straight into the build schedule, rather than living in email, is often what keeps the first full line on time.

[Omar Trejo @otrenav]

Your CEO asks why you cannot just use a chatbot. Your 10-second answer: alone, a chatbot helps one person write faster, but it does not read your systems, follow your rules or log what it did. Use it for drafts. Build for the work that runs every day.

[Omar Trejo @otrenav]

@pedrofragoso Me, personally. Every miss becomes a dated rule, and anything that sends, spends or publishes waits for my OK. Those three verbs cover a new agent from day one, before it has made the mistake that would add its first line.

[Omar Trejo @otrenav]

82% of 418 IT and security professionals had found AI agents they didn't know about in the past year (Cloud Security Alliance, Autonomous but Not Controlled). The chat in this video is illustrative.

Does every AI agent in your company have a named owner and an off switch?

[Omar Trejo @otrenav]

If a manual step still gates one of your workflows, it is likely capping how much work your team can take on. Count both costs each month: the staff time it burns, and the decisions that wait on its output. Price that before you call the step temporary.

[Omar Trejo @otrenav]

How to put a number on an AI saving that your board can check.

Pick one step, say matching supplier invoices to purchase orders. Your team handles 4,000 invoices a month, and 600 of them fail the match and wait for a person to clear them.

Price the step as it runs today. If each exception takes 12 minutes of an analyst's time, those 600 cost 120 hours a month; at $60 an hour that is $7,200 a month, before the late fees and lost early-payment discounts those delays cost you.

Then name the change. The software clears the exceptions that have a known cause, such as a price inside tolerance, a split delivery or a PO number it can find, and sends the rest to a person on your team with the evidence attached.

Measure the step the same way afterwards. If 400 of the 600 now clear on their own and the other 200 take 8 minutes each, the step costs you about 27 hours instead of 120.

That chain is your saving: which step changes, how often it runs, what each run costs now, and what it costs after, counting the exceptions still left to an analyst to clear by hand. Multiply the runs by the difference. Here the analyst cost falls by about $5,600 a month; that figure less the software's monthly running cost is the saving you report. A savings slide without that chain is only a guess that neither you nor your board can check.

[Omar Trejo @otrenav]

Your delivery dashboard can stay green quarter after quarter while your team loses its grip on the code going out.

A 2022 study tracked 128 deployed models. 91% degraded over time. Not one would have moved a DORA metric.

AI writes the change, the dashboard looks healthy, and comprehension is the thing nobody is measuring.

[Omar Trejo @otrenav]

@datachad Yes! And a smaller job for the model usually means it's simpler to run and easier to trust, and that makes the scale-up case stronger too.

[Omar Trejo @otrenav]

BearingPoint surveyed 1,050 senior executives: nearly 3 in 4 of those using AI report impact they can measure, yet only 13% of organizations scaled it as their business case planned (Scaling AI for measurable impact, 2026-10-01).

Has your pilot defined what scaled means?

[Omar Trejo @otrenav]

@Deepa_54236 Yes! Thanks for passing this along; the conditional formatting is usually the only documentation, and the color legend lives in one person's head.

[Omar Trejo @otrenav]

Will the output of your next AI project land in a tool your people already open, or in a new one they have to remember? An accurate result that never reaches the person who acts on it changes nothing in how your work runs.

[Omar Trejo @otrenav]

Before anyone quotes you a build, list every item from one recent period that left the main path, and what happened to it. If far fewer left than you assumed, the build is simpler. If far more did, the exceptions are the real product and the main path is the easy part.

[Omar Trejo @otrenav]

Your pilot answered one question honestly: can the software do this task? Whether your company relies on it is a separate question.

Ask what would stop if you switched it off tomorrow. What is your answer?

[Omar Trejo @otrenav]

@Cafflynuz Ha, and it's usually the one nobody draws until they need it. The parts that tend to matter most are the boring ones: who owns the data, what format it leaves in, and whether the docs are good enough for someone else to pick it up.

[Omar Trejo @otrenav]

Whatever you build now will need changing within a year or two. In seven minutes: how to buy software you can afford to replace, what to ask for by name before anything is built, and how to put a price on a change before you sign the contract.

[Omar Trejo @otrenav]

Your team still gives defensible answers from last week's numbers while margin leaks from routes nobody had time to re-price. Routing becomes an AI problem when your decision space outgrows your review cycle. Three signals say you have crossed that line. https://t.co/CS5TRHBOZU

[Omar Trejo @otrenav]

@datachad Fixing the process first also gives you a clean baseline. When it's time to show the pilot scaled as planned, you can tell what the process change did and what the model added.