Kaprin
All insightsFDE Field Notes

Start AI Projects With the Number You Want to Move

By Kaprin Engineering 7 min read

This is our thesis, written for CTOs, operations leaders and CFOs who've been asked to "do something with AI" and want to know what to point it at. Kaprin is an embedded AI engineering firm that practices forward deployed engineering (FDE), where senior engineers work inside a client's team and systems. In our work one pattern holds. Conversations that start with the technology rarely end well, and conversations that start with a number usually do.

What do buyers actually ask for?

When operators describe what they want, models rarely come up. These are themes we hear in different words, and none of them is a quote from a client. Each one is a measurable change in how the business runs.

  • Orders entered in minutes, where today they take days. Documents arrive by email and someone retypes them.
  • No duplicate card charges when the network drops. A payment gets retried and the customer is billed twice.
  • Reporting that doesn't eat a morning, because someone assembles the same numbers by hand every day.
  • A software bill that stops growing with headcount. Every new hire adds a seat, and every seat adds a cost.

Nobody describes these in terms of AI. Some need a model. Others need a database lock or a replaced spreadsheet. The buyer wants the friction gone and a number to prove it.

What is the four-step method?

Our method has four steps: name the operational friction, put one number on it and agree the baseline, build the leanest system that moves it, and ship it to production and measure. Each step has a written output, so anyone can check progress.

  1. Name the operational friction in one sentence, in operator language. "Staff retype purchase orders from PDFs into the ERP."
  2. Put one number on it and agree the baseline: minutes per order, hours per report, duplicates per week. The decision-maker signs it off.
  3. Build the leanest system that moves it, meaning the smallest design that reaches the target. A plain deterministic script beats a model call when it does the job.
  4. Ship it to production and measure. Real traffic, real data, the same metric, compared with the baseline.

The method maps onto our delivery rhythm: Scope within 24 hours, Embed on day 1, Harden in weeks 2-6, Handoff by day 45. The first two steps happen in scoping and week one, which we describe day by day in what an embedded engineer does in week one.

Why put a number on it first?

A number gives the project a finish line. Without one, pilots drift: demos look impressive, nobody can say whether the business improved, and the budget gets questioned at the first review. With one, every demo answers the same question, which is whether the number moved.

It also keeps you from building the wrong thing. If the baseline shows the delay sits in an approval step, automating data entry won't help, and you'd rather learn that on day 2 than in month three.

The number changes the budget conversation too. A CFO can't put a value on "an AI initiative". A reduction from four hours to ten minutes a day, or from days to minutes per order, converts into hours, salaries or capacity. That's the language in which projects get renewed or canceled.

A good metric already exists or can be measured roughly from current data, would be accepted by finance or operations as evidence, and responds directly to what the engineer builds, without depending on factors outside the project.

Does the answer always involve a model?

Not always. Many operational problems are solved by deterministic code: rules, validation, locks, queues and scheduled jobs. We reach for a model only when the input is unstructured or the decision needs judgment, and then we wrap it in validation, spend caps and audit logs. The leaner the system, the easier it is to run.

This is a stance as much as a cost decision. A plain script behaves the same way every time, costs almost nothing to run and can be fully tested. A model call can vary, costs money per request and needs guardrails. If the script does the job, a model would add risk for no benefit.

Where models are the right tool, we apply what we call deterministic scaffolding: idempotency keys, atomic locks, schema-validated outputs, hard spend caps, circuit breakers and immutable audit logs, with evals before agents. See anti-vibe-coding and deterministic scaffolding for more on that approach.

What does this look like in practice?

Three of the four requests above map to a Kaprin engagement we can show, and the fourth is covered below the table. These are our own published case studies, with the figures as stated on each case-study page. They're results for those clients, and we don't promise them for yours.

FrictionNumber that movedCase study
Purchase orders retyped from PDFs into an ERP and freight system 90% faster order entry; $150k+ annual impact; runs 24/7 Slender Morris
Daily client reporting consuming analyst time Reporting cut from 4 hours to 10 minutes; clients grew from 30 to 60 without new analysts; over $100k in avoided salary cost Hunt Digital
Stadium payments and ticketing under crowds of 50,000+ fans Contactless tap-to-pay in under 800ms; row-level locks give 100% idempotent ticket allocation; average donations tripled FansChance UK

Two further examples, Dingo Cement and Thomas Logistics, each reported about a 95% reduction in manual order processing. You can find them on our success stories page.

The duplicate charge problem

The FansChance work is the clearest example of a requirement that sounds simple and is hard in practice. A tap-to-pay terminal on a crowded concourse will lose connectivity, and a system that retries without protection can charge a fan twice. The fix is plain engineering: idempotency keys and locks. We explain the mechanics in preventing duplicate payments when a POS terminal process dies.

The software bill that follows headcount

The fourth request is about cost structure. Per-seat SaaS pricing penalizes headcount growth, because every new employee adds a license fee. Custom software deployed in the client's own cloud has no per-seat fee. Whether replacing a given product pays off depends on your numbers. The SaaS seat tax and build versus buy sets out how to check.

What are the common mistakes?

The most common mistake is choosing the technology before the number. Others follow: measuring effort when you should measure results, starting with the hardest workflow, and declaring success from a demo. Agreeing the metric and checking it on live traffic catches each of them.

  • Starting from the technology. "We need an agent" is a solution looking for a friction.
  • Vanity metrics say nothing about the business result. Counting prompts handled or documents processed tells you how busy the system was.
  • A first target that's too big. A narrow workflow in production teaches more than a broad one in review.
  • If nobody on your side owns the metric, nobody will notice when it slips.

What should you do with this?

Start with one friction and one number. Write the sentence, find the baseline, and ask any provider, including us, how they would move it and by when. If the answer starts with a technology name, push back. If it starts with your number, you're talking to the right people.

  1. List the operational frictions people complain about every week.
  2. Pick the one with the clearest, most measurable cost.
  3. Measure the current number for a week.
  4. Ask what the leanest system is that would move it.
  5. Agree a date for production and a date for the first measurement.

We also plan the ending: ownership, runbooks and credentials in your account. That's covered in the first 90 days of an FDE engagement. If you have a friction in mind, tell us about it and we'll reply with how we'd measure it and what we'd build first.

Questions readers ask

What do companies actually want from AI projects?

Usually a specific operational improvement: orders entered in minutes when they used to take days, no duplicate card charges when the network drops, reporting that doesn't consume a morning, or a software bill that stops growing with headcount. The technology is a means to that number.

How do you choose a baseline metric for an AI project?

Pick one number that finance or operations already recognizes, such as minutes per order, hours per report or failed transactions per week. Measure it as it is today, even roughly, and get the decision-maker to agree to it in writing before any build starts.

Do you always use AI models?

Not always. If a deterministic script solves the problem, we use the script. It's cheaper, faster, easier to test and has no variance. Models earn their place where the input is unstructured or judgment is needed, and even then they sit inside validated, capped and logged scaffolding.

How long before you can show the number moving?

Kaprin typically deploys to production within 30-45 days of starting, with weekly live demos along the way. The first thin slice runs on real data in week one, so you see movement toward the baseline metric well before the final cutover.

What if the number does not move?

We say so, find out why and decide whether to change the approach or stop. A single agreed metric keeps that conversation short. Learning in week three that a workflow isn't worth automating beats learning it in month six.

Bring us the broken prototype, the SaaS bill, or the manual process

Tell us what is stuck. We reply within 24 hours with a scoped plan, a fixed timeline and the one number we will move.

More insights

FDE = forward deployed engineering. Strict NDA protection. Senior architect review.