Everyone has a pilot.
Nobody has a system.
The demo worked because someone was holding it up. We build the version that runs Monday morning with nobody in the room.
Voice & world models
Workflow automation
GTM systems
Agent infrastructure
It does what you said, not what you meant. That gap is the entire job.
Two altitudes, one studio.
The frontier work keeps us honest about what this generation of models can carry, which changes every few months. The operations work keeps us honest about everything else, because a system that is elegant on Friday still has to survive a Monday morning with somebody pasting in a PDF that does not match the template.
Operations & go-to-market
We start with the people who run the process day to day, because they know which step is slow. It is generally something dull, like a quote that has been sitting in one approval queue since the morning. What we build learns your catalog and your exceptions, reads the documents nobody ever standardized, and hands back anything it is unsure about.
- Lead routing that reads the whole thread
- Quote and order work, edge cases included
- Reporting that assembles itself
- Agents with the authority to stop
Frontier experiences
A model release means very little to a team until they can put hands on it. Give us a week or two and they will have something working, often voice or generative video, along with a straight read on where the current generation still falls over. Some of this is sold direct and some runs through studios and agencies, with Golem Works as the technical partner inside their client engagements.
- Voice-native and real-time interfaces
- World models, avatars, generative video
- Vendor and model evaluation under load
- White-label builds for studios and agencies
Three builds, in detail.
Client names are withheld under NDA. Everything else here is as it happened.
Voice-native interactive storytelling
Product-led an interactive storytelling experience for a voice assistant, built on real-time generative world models, with full voice navigation, generated scenes and a live control panel the client could drive in front of stakeholders. Most of my time went to the experience design with their product managers, arguing about how a chapter should begin and what day three ought to feel like when day one has already set expectations.
I also ran the evaluation on the video, avatar and voice vendors sitting underneath it, which is where the latency numbers decide whether a concept can be demoed at all.
One vendor’s latency made the whole concept undemoable. We said so before anything was built on top of it.
AI audit, then a phased build
Two weeks inside the operation first, watching how the work moves rather than reading the process document. That produced a ranked build list with hours-per-month and a build cost against each line. We then shipped the highest-payback systems in order, and each one was carrying live traffic before the next was scoped.
Roughly half the ranked list was work we recommended against. Some of it was too low volume to matter, and one line was already handled by a $40 tool they had forgotten they were paying for.
Their returns problem turned out to be a badly designed form. We fixed the form.
The AI capability inside someone else’s studio
Embedded as the technical partner on frontier client work, covering model and vendor evaluation, prototype engineering, and product direction on builds the studio sells but does not staff internally. White-label throughout, with everything shipping under their name.
This is how most studios should buy AI capability. Hiring for it means committing headcount to a field that reshuffles every quarter, and renting it means somebody else carries that problem.
We argued for killing a concept in week one, because the model underneath it could not sustain the demo they wanted to give.
Different company, same mess.
Most engagements begin in one of these four situations. Whichever one sounds like your week is generally where we start.
The inbox that runs the company
Orders, quotes and complaints all land in one shared mailbox, and somebody sorts it by hand every morning before anything else can happen. That sorting is generally the first thing we automate. Anything unusual still goes to a person.
The process that lives in people’s heads
Nothing is written down, and no single person can describe the whole flow start to finish. So we map it together with the people who run it. That mapping is part of the audit, and the build follows the map.
The report that eats a Friday
Every week somebody senior copies numbers out of four systems into a document the rest of the company is waiting on. The assembled version can arrive on schedule without them, checked once rather than built by hand.
The pilot that stalled after the demo
There is a proof of concept from last year that impressed everyone in the room and then quietly went nowhere. In most of these the model was fine. What was missing was the plumbing around it.
If your version isn’t on this list, it is probably a cousin of something that is.
The name on the door
is on your calls.
I’m Alan Goldenberg, and I lead every engagement here. If we work together, I’m on the calls, in your systems, and answering the email when something misfires at eleven at night.
On larger builds I staff from a senior bench I have shipped with for years, and you will know who is working on yours. This year that has run from intake plumbing for a national distributor through to a voice-native experience for a Fortune 100 device platform, which is a wider range than I expected to be covering when the year started.
Email gets a reply within one business day. A scoped answer, a price, or a no, within three.
Start with an audit. Stay on retainer.
Audit
A few weeks inside your operation, on calls with the people doing the work, in your own data and your own systems. You leave with every workflow we think is worth automating, ranked by payback, with hours saved and a build cost against each line. The list also includes what we recommended against, and why.
A written map, a working proof against your own records, and a recorded walkthrough. All of it is yours to keep, including to hand to a builder who is not us.
Build
Scoped from the audit and built in phases, highest payback first. Something is working by the end of week two. Each phase goes live in production and has to earn the one after it.
Every system ships with a runbook, credentials in your name, and a written failure mode covering what breaks it and how you will know.
Retainer
How most clients end up working with us, because priorities never hold still for a year. We stay integrated, closer to an employee than a vendor, running what we have built and picking up the next piece as it surfaces. When the quarter changes your priorities we re-sort the list with you.
Or take the handover instead. Documentation, a walkthrough, and a clean exit, which is cheap to do because everything already lives in your accounts.
Every phase is quoted in writing before it starts, and each one has to earn the next, so you are approving a known number rather than watching a meter run.
Larger builds run with a small team I’ve worked with for years.
The terms, published.
You talk to the people building it.
There is no account-manager layer here. You are in a channel with whoever is writing the code, and you know their name.
Something working in two weeks.
The first thing we hand over is a system you can open and use.
We write down what it won’t do.
Every build ships with the list of cases it hands back to a human.
We fix the process before we automate it.
Several of our better results have been a rewritten form, one rule, and a deleted step.
We measure the before.
Whatever number we are trying to move, we agree on where it stands today before anything gets built.
You own everything from the first commit.
Your repo, your cloud, your keys. It all keeps running whether or not we are still involved.
The ones you’d ask on the third call.
The oldest story we tell about a made thing that acts on its own, and the whole myth turns on language. Inscribe the word and it moves. Remove one letter and it stops. Every version is also a warning: it does exactly what it’s told, never what was meant.
Tell me what your team keeps redoing.
Bring one process that eats more hours than it should. Half an hour is normally enough to work out whether it is worth automating and roughly what that would cost. I do not bring a deck to these.
Clay until someonewrites the word.