Proof

Three processes. Three industries.
One method.

The claim worth testing is not that I did one impressive thing. It is that this works on processes generally, and that the industry does not decide the answer. So here are three unrelated ones, each running in production.

Nobody should take a consultant's word for throughput claims. Every figure below traces to a log or a command I will run in front of you on a call. Client names are stripped, the systems are live and commercial, and the counts were taken in September 2026. All of them have gone up since.

Case 01 · Logistics

Freight forwarding operations, run by agents

An AI operations platform for freight forwarders. Built from zero in five months, running as a fleet of independent single-tenant stacks with live commercial customers on them.

Freight operations is the seam in its purest form. A rate request arrives by email in whatever format the customer felt like sending. Someone reads it, checks it against carrier rates, quotes a price and a transit time, and then operations has to actually deliver against that promise. Every step is a document, and every handover is where the margin goes.

The loop reads the request, structures it, prices it against the rate data, drafts the quote, and stops. Price and promise are gated: a human commits those, always. Everything upstream of that decision runs without anyone retyping a PDF into a system.

Row-level multi-tenancy was rejected in favour of one full stack per customer, with a separate database, mailbox and volumes. The data is commercially sensitive freight rates, and one missed WHERE clause is not a bug, it is a cross-customer breach discovered by the customer rather than by you.

5 months
from first commit to a production fleet
9
single-tenant customer stacks operated
45
automated checks gating every change
1
customer affected by a failed gate, never the fleet

When agents do a large share of the execution, choose architectures where the dangerous mistake is structurally impossible rather than merely unlikely. A reviewer who catches 95% of missing WHERE clauses stops being good enough once the volume goes up tenfold. The operational discipline is the same as it always was. What changes is who types.

Case 02 · Enterprise sales

Quote to proposal to contract, computed rather than guessed

A commercial system built in six weeks that turns discovery into the two documents a large enterprise deal actually needs, with the price derived rather than typed.

The proposal process ran on hand-written templates whose entire pricing logic was one typed number and a division by twelve. Separately, the platform metered nine usage dimensions accurately and priced none of them. Same missing artifact on both sides: a deterministic function from scope to price.

A large account needs two documents, and both go to the same person. The price proposal stops at the sponsor and the CFO. The plan gets forwarded, to the steering committee, IT, infosec, the integration vendor, the subsidiaries. You never see that list and you are not on it.

So the plan is not a document with a wider audience in any sense an access-control system can express. It is a document that leaves your custody with an unbounded blast radius, and it must therefore never carry a number with a currency symbol on it.

6 weeks
empty repository to shipped
2
document types from one scope model
0
prices computed by a language model

The model writes prose. Code computes price. A model that hallucinates a discount is not an embarrassment, it is a commercial liability, and a price that cannot be explained line by line will not survive a negotiation anyway. This is the rule I would apply to any pricing or promising step in any process.

Case 03 · The stress test

The same method, pointed at software delivery

The process I understand best is my own. Talent-Ray, the AI hiring platform I co-founded and run the technology for: backed by NVIDIA and Google, live with more than 20 companies within three months of launch. Its entire build-test-ship loop is agent-run, and one person operates it.

This one is here as a load test rather than a sales pitch, because software delivery is probably not your process. It is the case where the loop was pushed hardest, so it is the honest answer to whether the method holds up at volume.

A tester reports a problem in whatever form is easiest, and in whatever language they think in. It is transcribed, structured and ticketed. An orchestrator groups related tickets and hands each group to an agent working in isolation. Checks run. On green it merges under a lock and moves to testing. Anything too vague to act on goes back to the reporter as a question rather than becoming a confident wrong answer.

Nothing reaches production without an explicit human instruction. My own testers speak English perfectly well and still report in Turkish, because nobody describes a problem as precisely in their second language. The loop does not care which they use. That generalises further than it sounds: asking people to translate before they report is a tax paid in lost detail, and the detail you lose is exactly the specifics that made the report worth filing.

3,400+
autonomous agent runs, in six months
3,451
changes shipped in a single quarter
38
background workers in production
5
production environments, one operator
14,067
changes authored since Feb 2025
1
person on the team

The orchestrator decides which model account to run on every cycle, against per-account usage ceilings, and fails over when its own is exhausted. Agent capacity is a scheduling problem, and treating it as one is the difference between a demo and something that runs unattended for six months.

Before any of this

Twenty years of doing it by hand

Ten years at PSI Metals building production planning and due-date quoting systems for steel producers in Germany, Turkey, Poland, Brazil and Indonesia. Due-date quoting is this entire business in one problem: what sales is allowed to promise, given what the plant can actually do. Then five years as a McKinsey Engagement Manager on operations, advanced planning and digital transformation.

The agents are new. The problem is not, and that is the part most AI vendors are missing.

The demo

Fifteen minutes, no slides

None of this beats watching it run. Describe a request out loud, on the call, in your own language. Then watch it get structured, routed, worked, and stopped at the gate where a human belongs. If that does not convince you, nothing in a deck was going to.

Ask for the demo