Custom AI Development

You tried the off-the-shelf AI, and the output came back generic, or worse, invented.

Off-the-shelf works when the answer is already in the model. Custom starts when the answer depends on data only you have.

Tell us what you’re trying to do. Custom work is the plumbing that puts your own data in front of the model at the right moment, built around a real definition of good, with a human in the loop.

Or text 562-726-2287. No form marathon, no discovery call.

Describe the problem

See what custom covers

$3,000

Where AI projects start

Your data

In the prompt, not the open web

Human

In the loop on anything that matters

Sound familiar

You’ve tried the generic tools, and you still can’t trust what they hand back.

You bought the subscription, and every draft comes back sounding like the internet’s average take on the topic, never yours.

It states things with total confidence, including customer wins that never happened, so nobody on the team trusts what it produces.

The tool doesn’t know your product, your data, or your pricing rules, so the output stays generic no matter how carefully you prompt it.

Your people copy data out of one system into a chat window, then paste the result back into another, several times a day.

You’ve already tried a few tools, and not one of them changed the actual quality of the work.

The definition of what good looks like lives in one person’s head, so nothing can be checked, repeated, or handed to someone else.

None of this means the tools are useless. It means an off-the-shelf tool was never going to know your business. Here’s the version that does.

The version where this works

The model answers from your data, your catalogue, your rules, and a human stays in the loop on anything that matters.

Not the average of everything ever written on the internet. The build isn’t the intelligence, it’s the plumbing that puts your own data in front of the model at the right moment. It’s scoped around a real definition of what good looks like, so you can actually tell whether it’s working, and the copy and paste that eats your afternoons stops being somebody’s job.

What custom actually is

There’s one line, and almost everything sits on the cheap side of it. Off-the-shelf works when the answer is already in the model. Drafting, summarising, rewriting, translating, classifying against ordinary categories. A subscription does all of that, and you should buy the subscription.

Custom starts when the answer depends on data only you have. Your CRM records, your product catalogue, your pricing rules, your account history. No general tool has that, and no amount of prompting invents it. The build isn’t the intelligence, it’s the plumbing that puts your data in front of the model at the right moment.

The signal to watch for is copy and paste. If your people are pasting information out of one system into a chat window, then pasting the result back into another, you’ve found the gap. That pasting is the integration, done by hand, several times a day.

The other signal is a rule nobody’s written down. Custom work often turns out to be half integration and half getting a decision out of somebody’s head and into something you can test. If you’re doing neither, buy the subscription. We’ll tell you that for free.

Which side of the line

Off-the-shelf, or custom.

If the answer is already in the model, buy the subscription and we’ll tell you so for free. If it depends on data only you have, that’s the custom part, and no amount of prompting gets you there.

What you’re asking for	Off-the-shelf	Custom
Drafting, summarising, rewriting, translating	Buy the subscription	Overkill, and we’ll say so
Answering from your CRM, catalogue, or pricing rules	Can’t, it doesn’t hold your data	Yes, this is the point
Copy and paste between systems, all day	That paste is you, integrating by hand	Automated, no person in the middle
A rule that lives in one person’s head	No way to apply it consistently	Written down, and testable

What custom actually covers

What custom covers.

Integrations. Getting your systems to hand data to a model and take the result back, reliably, without a person in the middle copying things.
Retrieval over your own content. Answering from your documentation, your policies, your catalogue, rather than from the open internet.
Structured output you can act on. JSON a database can sort, not prose somebody has to read and retype.
Evaluation. A way to tell whether it’s working, because a system nobody measures drifts, and nobody notices.

What it looked like

What this looked like at Workhuman.

A SaaS company whose sales team couldn’t personalise outreach at the volume their quotas demanded. Experienced reps resisted changing what worked for them. Newer reps didn’t know what good personalisation was. Several tools had been tried, and none had changed the actual quality of the emails.

The diagnosis was that it was a people problem, not a tool problem, which is why the tools hadn’t helped.

What got built was a workflow using GPT-3.5 with Zapier and Airtable, drafting a sequence of personalised emails from CRM data, company size, focus, and the rep’s own notes, with context-aware follow-ups and event details pre-processed into the prompts. The rep then edited the draft rather than starting from a blank page.

That’s the custom part. Not the writing, which any tool can do, but the CRM data arriving in the prompt automatically. It beat the testing control, response quality improved, and sequences that had taken real time were ready in seconds.

Read the Workhuman case study for the full build.

Before you hire us

What we’ll talk you out of.

A chatbot on your website

Unless you can say specifically what it’s for. Most of them answer questions nobody asked, and annoy the people who did have one.

Replacing a person’s judgment

Augmenting someone’s throughput is a good project. Replacing their judgment is usually a bad one, and it’s how organisations discover the judgment was doing more than they thought.

Anything you can’t evaluate

If there’s no way to check whether the output is right, there’s no way to improve it, and you’ll be paying to maintain something you can’t trust.

Fine-tuning a model, in nearly every case

Almost every problem people bring to fine-tuning is actually a retrieval problem or a prompt problem, and both are far cheaper.

What it costs

What custom AI costs.

AI projects start at $3,000.

Custom work is scoped, not tiered, because the cost is in the integration surface, not the model, and the projects described here usually sit above the floor. Full detail is on the pricing page. If the work is continuous rather than a project, it belongs in a retainer.

If it’s a different problem.

AI consulting

The whole picture. This is one spoke of it.

AI workflow automation

When the task is repeated rather than bespoke.

AI content operations

When the bottleneck is publishing.

Questions we get

What people usually ask.

The subscription answers from what the model already knows. Custom work puts your own data, your CRM, catalogue, and rules in front of the model at the right moment. If the answer’s already in the model, keep the subscription. If it depends on data only you have, that’s custom.

Almost never. Nearly every problem people bring to fine-tuning turns out to be a retrieval problem or a prompt problem, and both are far cheaper. We’ll say so before you spend the money on the wrong fix.

Probably not, unless you can say exactly what it’s for. Most website chatbots answer questions nobody asked and annoy the people who did have one. If you can name the job, we’ll build it. If you can’t, we’ll talk you out of it.

Then we don’t build it that way. Evaluation comes first, because a system nobody measures drifts and nobody notices. If there’s no way to check the output, there’s no way to improve it, and you’d be paying to maintain something you can’t trust.

AI projects start at $3,000, and the ones described here usually sit above that, because the cost is in the integration, not the model. Detail is on the pricing page. If it’s continuous rather than a project, it belongs in a retainer.

Close

You’ll know which side of the line you’re on.

Off-the-shelf when the answer’s already in the model. Custom when it depends on data only you have. Tell us what you’re trying to do, and we’ll tell you which, for free.