ECONOMICS()
Cost per verified outcome
I count inference, tool calls, retries, review time and failed work. Token price alone does not tell me what the business paid.
AI CTO / KRAKÓW, POLAND
Michał Piszczek. AI CTO.
I am CTO at Archdesk. I work on AI systems used inside enterprise workflows, where a wrong answer can delay a project, trigger rework or create liability. My focus is the layer around the model: cost, data, permissions, verification and recovery.
CURRENT ROLE
CTO, Archdesk
ORIGINAL WORK
PAA | Joule Wars
BUILDER RECORD
4 ventures | 1 acquisition
OPERATING MODE
Strategy -> production
01 / OPERATING THESIS
What an AI CTO is accountable for
A model can generate an answer. The company still needs to know what data it used, what it changed, who can approve the result and how to recover when it is wrong. I treat that complete path as the system.
Generation is easy to demonstrate. The harder job is deciding which machine-produced work the organization can accept without repeating the work by hand. NULLIUS IN VERBA // TAKE NOBODY'S WORD FOR IT
02 / WHAT AN AI CTO OWNS
What I own in production
I use four ledgers when reviewing an AI deployment. Each one needs a named owner and an answer that can survive an incident review.
ARCHITECTURE()
Models, data and the harness
I keep business context, permissions, evaluations, observability and recovery outside the model API, so the model can be replaced.
PROOF()
Independent validation
The model cannot be the only judge of its own output. I use tests, replay, a different model family or a human at the irreversible boundary.
LIABILITY()
Ownership when AI is wrong
Before production, every consequential action needs a named owner, an evidence record and a recovery path. If nobody owns the failure, the feature is not ready.
03 / ORIGINAL FRAMEWORKS
The concepts I use
I named these measures because model benchmarks did not answer the questions that kept appearing in architecture and budget reviews.
PAA / PROOF
Proof-Adjusted Autonomy
The share of completed work that is autonomous, independently checked and delivered before the decision deadline.
->
JOULE / ECONOMICS
Joule Wars
A way to compare AI systems by useful work per unit of energy, instead of treating raw capability as the only score.
->
HARNESS / SYSTEMS
The Agent Harness
The control plane around the model: context, tools, permissions, memory, evidence, evaluation and recovery.
->
REVOKE / SECURITY
Revocation Exposure
The irreversible work an AI system can still accept while a revocation decision propagates through gateways, workers and regions.
->
J/VT / EFFICIENCY
Joules per Verified Task
Energy per accepted result, including failed attempts and retries - the production denominator behind the Joule Wars.
->
Definitions, formulas, DOI records and attribution guidance live in the canonical concept registry.
04 / OPERATOR EVIDENCE
Work that shaped this view
I did not arrive at this view from model benchmarks. It came from building systems that move money, monitor structures, interpret regulations and run enterprise workflows.
ENTERPRISE OPERATIONS
CTO of Archdesk
I lead technology at Archdesk, where software sits inside construction, manufacturing and engineering workflows. Access control, auditability and recovery are product requirements.
FOUNDER SYSTEMS
Four ventures
I founded Lextron.ai, Inclify, Rejsomat.pl and Robotero. The products covered regulatory research, structural monitoring, travel marketplaces and algorithmic trading.
ADVERSARIAL ROOTS
White-hat hacker
Security research from 2004-2012, including responsible disclosure of critical flaws in Sun Microsystems' sun.com. Read the record.
05 / FIELD NOTES
Recent technical notes
Experiments, operating models and arguments that I am willing to publish with the numbers attached.
AI AGENTS / VERIFICATION
Proof-Adjusted Autonomy: The 90% Agent Is a 61.6% Agent
READ FIELD NOTE ->AI ECONOMICS / LABOR
The Unit of Work Is the Agent-Hour
READ FIELD NOTE ->AI ENGINEERING / PROOF
Verification Cost Is the New Bottleneck
READ FIELD NOTE ->AI INFRASTRUCTURE / ECONOMICS
OpenAI's Economics: GPU-Constrained, Not Demand-Constrained
READ FIELD NOTE ->06 / OPERATING PRINCIPLES
Rules I use in reviews
Find the irreversible action
A draft can be regenerated. A payment, production deploy or message to a customer needs a different release gate. I design backward from that boundary.
Use a separate verifier
If the same model writes and grades the answer, both errors are correlated. The check needs another mechanism.
Count the accepted result
I include retries, review time and human exceptions. A cheap token can still produce expensive work.
Keep the harness replaceable
Models rotate. Context, permissions, evaluations and recovery logic should remain under the operator's control.
Report autonomy after verification
If a human has to reconstruct the work before accepting it, the work was not autonomous in any useful operating sense.
CITE_THIS_PROFILE()
Piszczek, M. AI CTO - Enterprise AI Systems, Proof and Economics. piszczek.pl. https://piszczek.pl/ai-cto
07 / DIRECT ANSWERS
Direct answers
What does an AI CTO do?
An AI CTO turns model capability into a reliable operating system for the business. The role owns the economics of verified outcomes, architecture and data, independent validation, security, liability boundaries and the organizational change required to put AI into production.
Who is Michał Piszczek?
Michał Piszczek is the CTO of Archdesk and a technology founder based in Kraków, Poland. He founded Lextron.ai, Inclify, Rejsomat.pl and Robotero, previously worked as a white-hat hacker, and is the author of Joule Wars, Proof-Adjusted Autonomy, Revocation Exposure and Joules per Verified Task.
What is Proof-Adjusted Autonomy?
Proof-Adjusted Autonomy measures the share of completed work that an AI system executes autonomously and supports with complete, independent, reliable and timely evidence. It discounts completed work that lacks proof.
What are the Joule Wars?
Joule Wars is the thesis that the durable AI competition moves from raw model capability toward useful intelligence per joule. As capability commoditizes, energy efficiency becomes an economic, infrastructure and geopolitical constraint.
Where is Michał Piszczek based?
Kraków, Poland. Work and speaking are available in Polish and English. For press, podcasts, panels and keynotes, use the media and speaking page.
OPEN_CHANNEL()
Contact
AI deployment, systems architecture, C-level advisory, podcasts and keynotes.