← All articles AI Insights · Speed vs Quality

10x Faster with AI? You're Only as Fast as You Can Verify

There is a growing expectation that with AI, everything should happen 10x faster.

On the AI side, that's true. You can run several agents in parallel. One drafts the report, one writes the code, one researches the market. Each finishes in minutes.

But the work isn't done when AI finishes. It's done when someone you trust says it's right.

The human sets the speed

I believe in a Human in the Loop workflow: AI drafts, humans decide. People review, correct and approve AI output before it goes live.

It's not only good practice. For high-risk AI systems, the EU AI Act makes human oversight a legal requirement (Article 14)¹.

The consequence is simple. If ten agents produce work and one human validates it, the process runs at the speed of that human. Parallel agents speed up production. They don't speed up judgment.

What verification really means

Verification is not a quick skim before you hit send.

For content and analysis:

• Read what was produced in detail, as if your name is on it. Because it is.
• Check every fact, number and quote has a real source. Then open the source.
• Ask: would I say this, in this tone, to this person?

For code:

• Confirm it covers the test cases, including edge cases and failure paths.
• Confirm it follows good engineering patterns: readable, consistent with the codebase, no duplicated logic, no hard-coded secrets, no silent error handling.
• Passing tests is the minimum. Code you can't maintain is technical debt with a deadline.

Speed vs quality: two conflicting forces

So we face two challenges at once:

1. Speed up the work, because AI makes that possible.
2. Keep quality high, because every AI output needs validation.

These forces pull in opposite directions. And our sense of speed can fool us. In a 2025 METR study, experienced open-source developers expected AI to make them 24% faster. They were measured 19% slower, yet still believed afterwards they had been 20% faster². Tools have improved since, but the lesson holds: feeling fast and being fast are different things.

The two extremes

Too fast, poor quality. Examples:

• A consulting report delivered to a government with references to papers that don't exist and a fabricated quote from a court judgment. Deloitte agreed to a partial refund³.
• Lawyers who filed a brief with case citations invented by ChatGPT and were fined by a US federal judge (Mata v. Avianca, 2023)⁴.
• A sales email that confidently quotes the wrong price, or a feature your product doesn't have.
• An AI-generated pull request merged because the tests passed, while it duplicated existing logic and quietly swallowed errors.

Too detailed, too slow. Examples:

• Rewriting every sentence of an internal status note so it "sounds like you", when the reader only needs the facts to be right.
• Line-by-line review of variable names in a throwaway prototype that will be deleted next week.
• Triple-sourcing common knowledge, like the fact that the EU has 27 member states.
• Three rounds of formatting tweaks on a draft only your team will see.

Neither extreme works. Fast and wrong destroys trust. Slow and perfect destroys the reason you used AI in the first place.

Balance is part of the answer, not all of it

The first step is matching the depth of the check to the risk. Keep a human firmly in the loop for:

1. Anything customer-facing.
2. Decisions about people: hiring, credit, health.
3. Money, contracts and legal commitments.
4. Anything you can't easily undo.

Lighter checks are fine for internal drafts, brainstorming and prototypes.

But balance alone keeps you stuck. You stay at the same speed forever, checking the same things the same way. The real gain comes from changing how trust is built over time.

The real solution: validate early, then earn trust

1. Front-load validation. When you start using AI for a new task, check everything. Read every line. Verify every source. Review every diff. This is expensive, and it should be.

2. Give feedback on where it misses. Don't just fix the output yourself. Tell AI what was wrong and why. "You used a statistic without a source." "This function duplicates the helper in utils." "This tone is too salesy for this client."

3. Make AI remember: write it down as skills. Feedback in a chat disappears when the chat ends. Write it into skills: written instructions that capture your standards, your examples and the checks you expect. Every correction becomes a rule. Every rule makes the next output better.

4. Reduce checking as trust grows. Once a skill consistently produces good results, move from checking everything to spot checks. Then to sampling. Your validation effort drops, and your speed goes up for real.

5. Never go to zero. Models change. Context changes. Your business changes. Keep regular but less frequent checks, and run a deep review from time to time to keep the bar high.

Think of it as a trust curve: heavy validation at the start, lighter over time, never zero.

Can AI check AI?

Yes, and it helps. A second agent can review the first one's work, catch obvious errors and run the tests. Use it as a first filter.

But I'd still keep a human in the loop. It's the only solid way to build real trust in a given AI process. An AI reviewer can tell you the output looks consistent. Only you can tell if it's right for your business, your client and your standards.

Treat AI like a senior expert on day one

The best mental model I've found: AI is an extremely capable senior expert who just joined your company.

Brilliant. Fast. Knows a lot.

But with no context. No business domain knowledge. No idea which path you want it to follow.

What would you do with a senior hire like that? You'd invest in onboarding. Explain how things work here. Review their first deliverables closely. Give clear feedback. Write down the standards.

Over time, they become independent. You stop reviewing every email. But you never stop challenging them, and you still check in from time to time.

Do the same with AI.

A simple checklist

1. Did I read the full output, not just skim it?
2. Does every fact have a source I've opened?
3. For code: do tests cover edge cases, and does it follow our patterns?
4. Is the depth of my check matched to the risk?
5. Did I turn today's corrections into a written skill or instruction?
6. When did I last do a deep review of this AI process?

The bottom line

AI makes production 10x faster. Trust is what makes the whole process faster.

Invest in validation early. Capture feedback as skills. Earn trust step by step. And keep a human in the loop, always.

How are you balancing speed and quality with AI in your team? I'd love to hear in the comments.

Google Tom Dulz for more info on what I do and how we can work together!

Sources

1. Official Journal of the EU, Regulation (EU) 2024/1689 (AI Act), Article 14 "Human oversight", July 12, 2024.
2. METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity", July 10, 2025. Reported by The Decoder, July 11, 2025.
3. Associated Press, "Deloitte to partially refund Australian government for report with apparent AI-generated errors", October 7, 2025.
4. Mata v. Avianca, Inc., US District Court for the Southern District of New York, sanctions order, June 22, 2023.