From Demo to Production: The Compliance Layer
Why the demos prove less than they seem
Three demos ran across this course, and all three worked. Here's why that proves less than it seems: those demos ran on clean data, staged folders, and a rehearsed path. An ordinary Tuesday doesn't look like that.
The research on this is blunt. Gartner projects that over 40% of agentic AI projects will be cancelled by 2027, and the reason named isn't that the model was too weak. Most of them die in the gap between demo conditions and production conditions.
Five checks before trusting an agent with real work
Is the input structured and clean, or buried in PDFs and tribal knowledge? Budget extra weeks if it's messy. This is the single biggest killer of these projects, and it's rarely the one anyone budgets for.
Can every output be checked, and is every action logged? This is practice 03, institutionalized rather than run by hand each time.
How many systems does the agent have to touch? Each one multiplies the effort. Every demo in this course was deliberately folder-based (files in, files out, no IT project required), and that's not a limitation, it's the correct place to start.
Who owns the output? Who signs? Which plan or deployment is in use, and is the data excluded from model training?
Handling time, error rate, or cost per case, measured against today's numbers before the project starts, and judged at 90 days. Projects with a baseline survive their first mistake; projects running on enthusiasm alone don't.
The EU reality: data and AI literacy
Two duties attach the moment real, non-sanitized data enters the picture. Neither is difficult, but both are a conversation to have before scaling, not after.
GDPR: the data question. Personal data needs the right plan, not a personal account. Consumer plans carry no data-processing agreement: fine for learning on sanitized data, unsuitable for customer data. For the real thing, a Team or Enterprise plan, or an EU cloud deployment such as AWS Frankfurt, is the appropriate path: a conversation with IT beforehand, rather than after.
EU AI Act, Article 4: the literacy question. Since February 2025, organizations have carried a legal duty to ensure staff working with AI are demonstrably AI-literate, with national enforcement beginning in 2026 as of this writing. Most everyday productivity use sits in the low-risk tiers of the Act, but the literacy duty applies regardless of risk tier, worth documenting, not assuming.
The works council
Co-determination in Germany applies to any system that could technically monitor performance, and an agent that logs its own work can, whether or not monitoring was ever the intent. It's the system's capability that triggers the right, not what anyone meant to do with it.
Deploying quietly and letting the works council find out later breaches co-determination before day one. The likely result is an injunction, a forced rollback, and a workforce that distrusts the whole program afterward. Involving the works council early, and framing the deployment honestly as assistive tooling with human sign-off, is simply the truth if practice five has actually been followed. That path is slower to start and faster to stick.
Say a finance team quietly rolls out the monthly-report skill from Lesson 08 across the whole department, without a word to the works council, reasoning that "it's just a reporting tool, not a monitoring system." Three weeks in, someone notices that the skill's own Flags tab (the exceptions list practice 03 asked for) incidentally shows who submitted late, incomplete, or repeatedly-corrected reports, department by department. Nobody built that as a performance metric. But it is one, technically, whether anyone meant it to be or not, and that's precisely the trigger this lesson opened with: capability, not intent. The honest, unresolved part of this case: there is no clean fix once the tool is already live and the works council finds out from a concerned employee rather than from the rollout plan. Rolling it back looks like an admission of bad faith even if none was intended; leaving it running compounds the original breach. The only real fix is the one this section already named: have the conversation before deployment, not after. That's exactly why it's a "before scaling" line, not a "before some employee complains" line.
None of this is a substitute for legal advice specific to a given company and jurisdiction. It's the plain-language version worth knowing before that conversation happens, not instead of it.
The same five practices, industrialized
There's a question worth asking once the five practices feel solid: what does this look like at serious scale, running a company's real workflows rather than one person's desk? Teams building the most advanced orchestration systems around agents like Claude Code aren't doing anything different from what's in this course. They're running the same five practices, turned from personal habits into machinery that enforces them.
Brief before you delegate becomes an ambiguity score. Instead of a person judging whether a brief is clear enough, the system measures how vague the request is and physically blocks the agent from starting until it clears a bar. That's the plan-mode correction from practice 01, made mechanical.
One bounded job, not a mission becomes a hard-coded "split, never simplify" rule. The agent is forbidden from quietly shipping a watered-down version of an oversized task. If the job is too big, it must halt and split it: not a suggestion, a constraint it cannot route around.
Demand evidence, not assurances becomes an adversarial second agent. A completely separate reviewer, instructed to assume the work failed until the evidence proves otherwise, checks the output. The checker is never the doer. An agent grading its own work is a sycophant, whether the grader is a person skipping a step or a model marking its own exam.
Write the handbook once becomes all state in files. Every decision lives on disk, not in a conversation, so a session can crash and the next one resumes exactly where it stopped by reading the files. The phrase worth carrying forward: the conversation is volatile; the files are permanent.
Nothing ships without your signature stays a hard-coded human checkpoint. Even maximally autonomous systems reserve visual checks, architectural decisions, and approvals for a person, by design. Autonomy at the frontier still never means unsupervised.
There is no sixth practice. Getting better at this isn't a matter of learning more rules. It's going deeper on the same five.
The last lesson compresses this whole course into three concrete moves for Monday morning.
Reply here and it goes straight to Rod. Same as replying to one of his emails.
