73%
of AI pilots never reach a second workflow
The easy task gets automated. The next one is half as valuable and twice as hard, so the program stalls there.
The problem
Everyone bought AI. Almost nobody got intelligence.
Three years into the boom, the pattern is consistent enough to be boring. A company buys a dozen seats, runs a pilot, sees a good first month, and then watches the curve flatten. Spend is recurring. The savings were not.
What went wrong is rarely the model. It's that the work was scoped as a tool purchase instead of a system design. Nobody owned the workflow the tool sat inside, nobody captured what happened after it ran, and nobody was accountable to a number 90 days later. So the organization got faster at the same work, and no smarter at any of it.
Meanwhile the two ways to get help both failed in predictable directions. Consultancies diagnose beautifully and leave before anything runs. Dev shops build exactly what you spec but can't tell you what to spec — and will never build the part that learns, because it wasn't in the ticket.
14
overlapping AI subscriptions at a 200-person company
Bought by five different teams solving four versions of the same problem. Nobody owns the total.
0
outcome data captured by the average deployment
Without a record of what the output caused, there is nothing to learn from — and no moat a competitor can't buy.
Figures reflect patterns across our intake diagnostics and published industry surveys; directional, not audited.
Why we built it
CRFT started as the thing we kept having to do before the real work could start.
AWSM LABS is an applied AI lab. We build systems that run in production, for companies whose margins depend on them working. And on nearly every engagement, the first six weeks went the same way: before we could build anything worth owning, we had to untangle what the client had already bought, find the workflow that actually moved money, and prove which opportunity deserved to go first.
That pre-work was the highest-leverage thing we did, and it was invisible — folded into a build fee, undocumented, unrepeatable. So we made it a product. CRFT is that diagnosis formalized, priced in the open, and welded to the build and support that follow it, so the analysis never gets divorced from the thing it was supposed to inform.
What we believe
Five positions the whole practice rests on.
01Systems, not products.
A model is an ingredient. The system is the workflow, the humans in it, the data it produces and the loop that feeds back. We scope all four or we don't take the work.
02Compounding beats automating.
Automation is a cost you cut once. A learning loop is an asset that appreciates. We'll happily build the first, but we price and design for the second.
03Price in the open.
Fixed scope, published price, no hourly drift. If we can't quote it before we start, we don't understand it well enough to build it.
04Adoption is the deliverable.
Deployment is not adoption. We sit with the people who run the workflow before we design it, and we ship a scorecard that measures use, not uptime.
05You own everything.
The code, the prompts, the data and the playbook, outright at handover. Ongoing capacity exists because the loop keeps improving — not because leaving breaks it.
Who it's for
Companies with a workflow worth defending.
Typically venture-backed founders and ops leaders at 30–500 people, past product-market fit, with real operational volume and a board asking what the AI line item bought.
You have volume, not just ambition
Enough repetitions of a workflow that a loop has something to learn from.
Someone owns the outcome
A named person whose number moves if this works. Committee-owned projects don't ship.
You want to own what gets built
Not rent a black box from a vendor whose roadmap isn't yours.
You need a demo for a board meeting
We're the wrong shop. Plenty of good ones build those faster than we do.
The practice
Five disciplines on every build, not one.
A system is a living thing operating inside a human team. That's why data science alone has never been enough — and why the handoffs, the trust and the process redesign are first-class build work here.
Design
How the system talks to a person — and whether they trust what it says.
Change management
What happens to a role when a loop takes over one of its steps.
Product
One outcome, ruthless scope, and a version two that's already sketched.
Data science
What the data can honestly support, and the readiness level you're actually at.
Engineering
Production systems your team can run, extend and audit without us.
See how the four stages run.
How CRFT worksGet started
Build the intelligence your company compounds on.
Five inputs, a published price, a report in 48 hours. Decide with numbers — then own what gets built.