AI Assessment
Before you build the AI, has anyone priced it?
Not the pilot. The thing running every day, on your real volume, against whatever you're doing today instead.
A situation you may recognize
Someone decides the company needs to do something with AI. A board member asks. A competitor announces one. It's not an unreasonable instinct.
A process gets chosen, usually the one where people are visibly retyping things. A pilot gets built. It works on twenty documents and everyone is encouraged.
Then it meets four thousand documents a month, and the edge cases, and the ones that were scanned crooked in 2019. Someone finally asks what this costs to run.
Nobody has the number. And by now the project has a name, a slide, and a person whose reputation is attached to it. The question gets answered softly.
Four questions, asked early, cost almost nothing
What does it cost per unit of work, at your real volume?
Not the pilot's volume. Per document, per ticket, per record, every month, forever. Most AI cost surprises are volume surprises.
What is it replacing, and what does that cost today?
Usually a person's hours, or code you already own. If the comparison has never been written down, there's no way to know whether this is a win. A surprising share of work that looks like AI is cheaper as ordinary code.
What number would make you stop?
Decide it before you measure, and write it down. A threshold agreed in advance is a decision. One argued afterward is a negotiation, and the project usually wins.
What has to be true besides the price?
Who is allowed to see what data, and for what purpose. This is the thing that actually stalls projects, and it's almost never on the slide. Price is what gets debated. Governance is what blocks it.
One engagement measured the AI at 150 times the cost of the code it would have replaced.
A large enterprise wanted to use generative AI to enrich customer profiles across millions of records. The architect wrote the kill criterion before measuring anything, then found the AI ran at up to 150x the cost of the existing SQL for under 1x return. Eight in ten of the candidate jobs were cheaper as plain code.
He recommended against shipping his own design. Then he built the governance that was the real blocker, kept re-measuring against each new model release, and the pipeline shipped later at cost parity. It runs in production today.
That engagement was Paul Laudeman's own consulting work, not a Red Barn project. He is a Principal Data & Agentic AI Architect, and he advises us on enterprise AI and data work. The four questions above come from how that decision was made.
Why we'll tell you not to
Most firms selling AI don't get paid when the answer is no. We build software, and we'd rather build the thing that pays for itself than the thing that was easiest to sell you.
Sometimes the honest answer is a script, a better form, or fixing the process that creates the paperwork. We'll say so.
Bring us one process
Thirty minutes, no charge. Pick one workflow you're considering AI for: documents being read by hand, data retyped between systems, the same customer questions over and over.
You'll get:
- Whether AI can realistically do it, or whether something simpler wins
- A cost range at your volume, not a pilot's
- Where it would break, and what would have to be true first
A one-page written summary either way, yours to keep and circulate internally. No follow-up unless you ask for one.
Goes straight to [email protected]. Tell us what the process is and roughly how often it runs.