How We Review AI Work

For all client work: websites, copy, content, ads, and SEO

Our Promise

Nothing goes to a client until a human on our team has reviewed it and stands behind it.

AI can turn out code, copy, ads, and articles faster than we can read them. Our bottleneck is now how long it takes us to review. We should plan our work around what we actually have time to check.

Ground rules

• You ship it, you own it. If your name is on the pull request, the article, or the ad set, you're responsible for every line in it, regardless of how it got written. We never tell a client, "The AI did that."
• Keep batches small enough to actually read. Don't dump a massive AI-generated change or a folder of fifty blog posts onto a teammate. Aim for a max of 400 changed lines of code per PR, and one article, one landing page, or one ad group per review. If it's bigger than that, break it up. Skimming is not reviewing and no one can review weeks of content, code, ad work in an hour.
• Never let AI invent facts. Prices, offers, dates, product specs, stats, testimonials, case-study results, and legal copy must come directly from the client or a verified source. AI can help rewrite them, but we have to check every single fact against the original.
• Keep client data out of prompts. Never paste passwords, API keys, customer lists, or other personal data into an AI tool. Stick to the tools approved in the client's contract, and keep agents sandboxed away from live credentials.
• Don't publish in bulk just because we can. Search engines penalize pages mass-produced to chase rankings, and ad platforms reject claims we can't support. Every page, post, and ad has to be worth a reader's time, and one of us has to have read it.
• Be upfront with clients. We tell clients in our contracts that we use AI, and we don't call something "reviewed" until it's actually gone through the right check below.

What we promise clients

Clients trust us because we can back up what we hand over. That only holds if we never promise more than we can review.

• Estimate on review time, not ai generation time. A draft that took the AI ten minutes can still take a person half a day to review. Quotes and deadlines are based on the review time.
• Say it at the start. In the proposal and at kickoff, explain that AI speeds up drafting, prototyping, first designs, etc... But a person reviews everything we deliver and that the review sets the pace.
• Don't agree to volume we can't check. If a client wants more than we can review by their date, tell them how much we can properly review by then and when the rest will follow.
• Be able to show the review. If a client asks "how do you know this is right?", we point to who checked it, when, and against what. The review note runwell task is that record.

How deeply to review

Client facts & claims

• What it covers: Any copy, content, or ad making real claims: pricing, offers, product specs, stats, testimonials, case-study results, competitor comparisons, or legal, medical, and financial text.
• What you need to do: Fact-check every claim against client-provided source material. If you can't verify a claim, take it out or ask the client. Check ads against the platform's ad policies.
• Who signs off: You, plus written approval from the client.

High-risk code

• What it covers: Auth, permissions, payments, personal data, tracking and consent, form submissioins, webhooks, database migrations, deployment, and infrastructure.
• What you need to do: Read and understand every line. Make sure automated tests cover the change, and test it yourself in staging.

Published content & ads

• What it covers: Blog posts, landing pages, emails, social posts, ad copy, and SEO titles and descriptions that don't make hard claims.
• What you need to do: Read every word of every piece, including bulk items like ad variants and meta descriptions. Check brand voice, originality, links, and that it's useful to a real reader.
• Who signs off: You, plus the client's usual approval before it goes live.

Everyday web work

• What it covers: Standard features, UI, styling, refactors, and simple UI copy (like headings and button labels).
• What you need to do: Read the whole change. Check it in a browser on both desktop and mobile sizes. All automated checks pass.
• Who signs off: You.

Throwaway & drafts

• What it covers: Prototypes, experiments, internal scripts, outlines, keyword research, brainstorms, and rough first drafts.
• What you need to do: Quick spot-check. If a client is going to see it, clearly label it as an unreviewed draft.
• Who signs off: No one (until it turns into real work—then review it properly).

If a change mixes a few of these together, review the whole thing at the strictest level that applies.

Day-to-day workflow

1. Know what you're building first. Decide what the change is and how strictly it needs to be reviewed before you start prompting.
2. Work in small chunks. Feed the AI the client's actual source material so it doesn't guess or fill in blanks on its own.
3. Run the automated checks. Fix or regenerate until they pass: build, tests, and linters for code; spelling, links, and plagiarism for copy.
4. Do the human review. Check the work yourself. If the AI wrote code you don't fully understand, or copy you can't vouch for, don't try to patch it—throw it away and ask for something simpler.
5. Leave a quick review note. On the PR for code, or in the doc or ticket for copy and ads. Mention what the AI generated, what you checked, and who signed off.
6. Ship it. Merge or publish once it's signed off (and once the client approves in writing, if it includes factual claims).

When you're pressed for time

Generate less. Never skip the review. If a deadline is tight and reviews are piling up:

• Break the work into smaller, easier-to-review pieces.
• Talk to the client about cutting scope.
• Push the deadline and be honest with the client about why.

Skimming high-risk code, spot-checking factual claims, publishing a batch nobody read, or merging something because it "looks fine at a glance" isn't how we work.

The batch limits are practical starting points. We'll revisit them every few months based on what's working for the team.