The bottleneck is review, not writing.
Everyone who's tried content automation knows: generation is fine. It's the partner review that kills the rhythm. The partner blocks two hours for review, those hours go elsewhere, drafts pile up, rhythm breaks, the experiment dies by week six. The fix isn't better AI — it's better review criteria.
Generation is no longer the constraint
GPT-4-class models write competent first drafts for most legal topics. Quality is sufficient; the edge cases are not worth solving at generation time.
Partner review is the constraint
Partners have 15 minutes, not 45. Review that takes 20+ minutes per piece becomes impossible to sustain, even with great drafts.
Pre-defined criteria compress review
If the partner knows exactly what to check for — tone, don't-say list, source requirements — review drops to <5 min. Without criteria, review takes forever because the partner has to invent them each time.
The criteria that compress review.
Before any draft is generated, define the review checklist. Five dimensions, each a yes/no question. If all five are yes, sign off. If not, reject with a specific reason — the next draft gets it right.
The goal: a partner scanning these five in under 60 seconds, then deciding. If the criteria are right, 85–90 % of drafts pass on first review.
- Does it match our tone (direct, concrete, no hype, no jargon)?
- Does it stay within the practice area (no cross-area claims)?
- Does it follow BRAO/UWG (no advice, no guarantees, no comparison)?
- Does it have ≥1 concrete example or data point?
- Does it end with a clear reader takeaway?
Shape, tone, structure — per channel.
LinkedIn, blog, and newsletter each want different shapes. Generating one piece and cross-posting doesn't work — but generating three channel-tuned versions from one topic does. Here's what each wants:
150–300 words. Hook in first line, one insight, one reflection question at the end. No hashtag spam, 3 at most. Single image optional.
Attention mediumBlog
800–1,500 words. Problem → analysis → actionable steps. Subheadings every 200 words. One embedded example or screenshot.
Reference mediumNewsletter
600–900 words. Direct address, first-person voice. One main idea, two supporting points, one clear call-to-action.
Relationship mediumA month of content, in 90 minutes.
Here's the monthly cadence we see working. Total partner time: ~90 min. Total drafts: 8. Total published pieces: 8 across three channels (with adapted versions).
- Week 0: 30 min planning session — partner picks 4 topics for the month from a topic bank.
- Week 1–4: AI generates drafts (2 per week). Each draft arrives in Teams on Tuesday morning.
- Tuesday + Thursday: partner reviews drafts (~5 min/piece).
- Approved drafts auto-publish to LinkedIn (same day), blog (Wednesday), newsletter (first of month).
- End of month: 10-minute review of what performed — inform next month's topic selection.
Guardrails for legal content.
Content from legal professionals is held to professional rules — BRAO in Germany, bar rules elsewhere. The generation prompt and review checklist both enforce these. It's not optional; it's a must-have.
No advice
The model is prompted to never give specific legal advice. Phrases like "in your situation" are banned. General education only.
No direct client mention
No naming of clients, cases, or matters. Anonymisation enforced at generation time and verified at review.
No guarantees
No absolute claims about outcomes. Hedging language required where probabilistic claims are made.
No comparative claims
No statements that compare against other firms or claim superiority. Factual differentiation only.
What success looks like.
Content automation is useful only if it's measurably better than manual. Track these three KPIs from day one:
Pieces/month published
Baseline: 1–2. Target: 6–10. Anything below 6/month means the rhythm isn't holding.
Draft-to-approval time
Target: median under 1 day. Longer means review criteria aren't tight enough.
Engagement per piece
Channel-specific. LinkedIn: comments and shares. Blog: time-on-page. Newsletter: opens and clicks.