AI newsletter agent or freelancers: a sub-€9M scoring method

The founder of a €4.2M Dutch DTC pet-nutrition brand walks into our Tilburg office on a Tuesday with three printed quotes on her clipboard. One is for a Claude-driven newsletter agent. One is for a six-person freelance editorial pool out of Utrecht. One is for a hybrid pipeline: AI drafts, human reviews, AVG checks the citations. She has 14,200 subscribers, a 21% open rate she can recite from memory, and twelve months of marketing budget to allocate by Friday.

She asks the question every founder in this revenue band asks: which one am I supposed to pick? The honest answer is that none of the three quotes told her how to decide. They told her what they would build. This post is the scoring method we use when the choice lands on our desk.

The three options, stated plainly

Every quote in this category collapses to one of three shapes. We rewrite the founder's three PDFs onto a single page so the comparison is real.

Option A — agent-only. A Claude-driven pipeline pulls product data, reads the last twelve months of newsletters, drafts the next one, calls a tool to attach images, and triggers the ESP send. No human in the loop after week two. Roughly €18k setup, then API and infra.

Option B — freelance pool. Six freelancers on a rotating Monday-morning brief. A managing editor (usually the founder's marketing lead) signs off. Between €350 and €700 per issue, plus internal time.

Option C — hybrid. An agent drafts, scores its own sources for AVG defensibility, and hands the draft to a single in-house reviewer. That person edits the brand-voice slop out and presses send. €11k setup, plus API and a slice of one salary.

Per-mail cost across 14,200 subscribers

The unit is per-mail-sent, not per-issue. That distinction matters because the agent cost is almost entirely fixed per issue, the ESP cost is linear in subscriber count, and the freelance cost is fixed per issue. Three different curves.

We model a 52-issue year and roll it up:

subscribers: 14200
issues_per_year: 52
esp_cost_per_1k: 0.40   # postmark-tier sender

option_A_agent_only:
  setup_amortized: 6000   # 18k over 3 years
  api_per_issue: 4.20     # ~80k input + 6k output, Sonnet-tier
  esp_per_issue: 5.68     # 14200 * 0.40 / 1000
  human_per_issue: 0
  per_issue_eur: 15.6
  per_mail_eur: 0.00110

option_B_freelance_pool:
  setup_amortized: 0
  api_per_issue: 0
  esp_per_issue: 5.68
  human_per_issue: 480    # avg weekly rate across rotating six
  per_issue_eur: 485.7
  per_mail_eur: 0.03421

option_C_hybrid:
  setup_amortized: 3700   # 11k over 3 years
  api_per_issue: 3.10     # smaller drafts, more retrieval calls
  esp_per_issue: 5.68
  human_per_issue: 95     # 1.5h editor at 63 eur/hr loaded
  per_issue_eur: 108.0
  per_mail_eur: 0.00761

The cost spread between Option A and Option B is roughly 31×. That sounds like it ends the conversation. It doesn't. Cost is one of four dimensions. The other three are where the decision actually gets made.

AVG-defensible source attribution

If you send marketing claims to 14,200 EU residents, every claim has to be traceable to a source you can show the Autoriteit Persoonsgegevens if they ask. That includes “studies show,” “scientifically formulated,” and the small print under a product photo. The AVG, the Dutch implementation of the GDPR, does not care that the agent wrote it.

This is where Option A degrades fastest. A pure agent will produce a fluent paragraph and quietly attach a citation to a paper that does not exist. We tested four agent stacks across the last quarter and every one of them did this at least once per hundred drafts, even with retrieval grounding and a sternly worded “do not invent sources” system prompt. The failure mode is not loud. It is one sentence in the third paragraph that the founder reads, nods at, and ships.

Option B fails differently. Six freelancers means six citation styles, six interpretations of “primary source,” and six different opinions on what counts as evidence. The AVG audit trail exists but it is messy.

Option C is the only one where the source list is generated by the agent, stamped, and signed off by one human who is liable for the send. Provenance is queryable per claim. That matters not because audits are common, but because audits happen exactly once and you need the trail ready before you know they are coming.

Brand voice over a quarter

Voice is the dimension founders underestimate. Over thirteen issues — one quarter — every option drifts. The question is which direction.

The agent drifts toward the mean of its training distribution. Fluent, slightly American, indistinguishable from any other DTC newsletter in the inbox. You will not notice in week one. By week eight the unsubscribe rate ticks up half a point and you will not know why. We have seen this happen in three of three agent-only deployments where the founder did not own the style guide and refresh it monthly.

The freelance pool drifts in six directions simultaneously. Each writer has their own rhythm and the managing editor either flattens it (costing voice) or doesn't (costing consistency).

The hybrid drifts the slowest because there is exactly one human whose job it is to hold the line. That human is the founder, the marketing lead, or a senior copy editor on retainer. If you cannot name that person on the day you sign the contract, do not pick Option C.

The compute-cost wildcard

One dimension that doesn't show up in any of the three quotes: where model pricing goes next. Published per-token rates move, quotas tighten, and new tiers get introduced — usually faster than a DTC marketing contract gets renegotiated. If your newsletter pipeline is a single Sonnet call, your per-mail cost in eighteen months is whatever the model vendor and its cloud counterpart decide it is on the day.

That doesn't change the scoring today. It changes the weight you put on resilience. Option B is immune. Option A is fully exposed. Option C lets you swap the agent for a cheaper model — or, if quotas bite, extend the human-edit window — without rebuilding the workflow.

The rubric we hand the founder

Score each option 1 to 5 on four axes. Multiply by your weights. Pick the highest total. The point is not the number. The point is forcing the founder to name her weights before the vendor pitches her.

Founders running on tight gross margin weight cost at 0.4 and the other three at 0.2 each. The math picks Option A. Founders who have ever received an AP letter weight AVG at 0.4 and the math picks C. There is no universally right weighting. There is only the one you can defend to your board the day something goes wrong.

The smallest thing to do today

Before you pick any of the three: open last quarter's twelve newsletters in one PDF and mark every claim that would need a source if the AP asked. If the count is under ten, your AVG exposure is low and you can optimise for cost. If the count is over forty, your decision is already made — you need a human in the loop, which means Option C or B.

When we built the hybrid pipeline for a Dutch supplements brand last quarter, the thing we ran into was exactly this: the agent's draft was clean, but its citation tool kept attaching plausible-looking DOIs that resolved to papers in adjacent fields. We solved it by wiring a second small model whose only job was to fetch the abstract of every cited paper and compare its topic vector to the claim. Cheap, boring, works. If you are evaluating a similar setup, our notes on shipping AI agents into regulated content workflows go deeper on the architecture.