Skip to content
Metzger Creative
Services

AI integration

AI that answers questions from your own data and takes real work off your team, built into your product.

Book 20 minutesSee the other services

Who this is for

  • You have a corpus your customers keep asking questions about
  • Your team is copying and pasting between a chat window and your admin panel
  • A vendor quoted you for something you suspect is a thin wrapper
  • You need to know whether this is worth doing before you fund it

What gets built

A question hits the index and three chunks come back. Everything the model can call on the other side is typed and validated first.

How it works

  1. A two week sprint against your real data, ending in a working prototype and a written recommendation

  2. An evaluation set built early, so "is it good enough" has a number behind it

  3. If it is worth building: every tool the model can call is typed and validated, so it cannot take an action nobody defined

  4. A prompt-injection and red-team pass before anything goes near a customer

What it costs

The bands that apply to this work, with what each one includes and what pushes it up or down. Nothing here is a number you have to book a call to hear.

AI integration sprint

Two weeks, one fixed fee, and an honest answer at the end.

$6K

What it includes

  • A working prototype against your real data, not a demo dataset
  • An evaluation set, so "is it good enough" has a number attached
  • A written recommendation, including "do not build this" when that is the answer
  • Everything built stays yours whichever way the decision goes

What moves the number

  • Nothing. The fee is fixed and the two weeks are the whole scope.

Sometimes the answer is that the full build is not worth doing. Finding that out here costs a fraction of finding it out at the end of a build.

AI build

The assistant or agent from the sprint, made production grade.

$10K–$40K

What it includes

  • Retrieval built and tuned against your actual corpus
  • Typed tools with validation, so the model cannot take an action you did not sign off on
  • Prompt-injection defense and a red-team pass before launch
  • Shipped behind your auth, inside your product, not in a separate widget

What moves the number

  • Retrieval quality work: how messy the source corpus is
  • How many tools the agent needs, and whether they write as well as read
  • Evaluation and red-teaming depth
  • Whether it ships behind your auth and into your existing UI

Questions people actually ask

Everyone is selling AI right now. Why you?
Because I will tell you not to build it when that is the right answer, and the sprint is priced so that outcome is still worth what you paid. Most of the value in this category is knowing which problems are actually retrieval problems.
Where does my data actually go?
That decision is yours. I will lay out what leaves your infrastructure, what does not, and what the options cost, including running smaller models yourself. Nothing gets sent anywhere before you have agreed to it.
Can it take actions, not just answer questions?
Yes, and that is where the engineering is. Every tool is typed and validated, write actions are scoped and logged, and anything irreversible needs a human confirmation step.

Sound like your situation?

Twenty minutes, no deck, and a straight answer at the end about whether this is the right piece of work for you. If it is not, I would rather say so now than at proposal stage.

Book 20 minutesSee every band