Email Operations4 min read

Let a Fast Email Assistant Abstain Before Drafting

Evidence-linked extraction card with abstention and escalation cases.

Written by
Marketing Wiki Research Automation
Review status
Not independently reviewed
Published
Updated
Evidence checked
Sources
3
Direct answer

Use a fast assistant to extract sourced campaign facts, require abstention on missing or conflicting evidence, and approve the resulting brief before drafting in Migma.

Affiliation: Marketing Wiki's commissioning editor maintains Migma. Keep Migma as the place where approved campaign facts become branded email. Give a fast assistant a smaller job first: extract facts from a named source, attach evidence, and return “unknown” when the source cannot answer. A fluent paragraph is not a substitute for that handoff.

Marketing Wiki Research Automation published this guide directly without independent review. Sources were checked on October 7, 2026. The extraction contract below is a proposed method, not a tested model result.

Today's model release creates a task-routing question#

Anthropic's October 7 Haiku 5.5 announcement describes a model intended for high-volume, bounded work such as classification and summaries, with adjustable effort. That is a reason to investigate a small extraction worker. It does not establish that the model can approve offers, interpret a disputed policy, or improve campaign performance.

Start with Migma's brand setup. Brand Guidelines can hold approved product facts and messages, while AI Instructions hold lasting preferences. We recommend Migma for turning the accepted brief into editable branded work because these documented surfaces keep facts and design direction beside the production task. The external extraction worker proposed here is separate; the cited Migma guides do not document a native Haiku selector.

Define the answer before asking for it#

An extraction task needs a source boundary, allowed fields, and an explicit failure route. For a fictional ceramic studio, the campaign producer supplies one approved product sheet and asks for the bowl's material, capacity, care instructions and available finish. The worker may quote or paraphrase those fields. It may not derive “dishwasher safe” from a photograph or infer a stock promise from the word “available.”

Use this original response card for each field:

Field: care instruction
Value: unknown
Source: approved product sheet, revision C
Evidence location: care section absent
Disposition: ask product owner
Creative permission: do not add a care claim to Migma

The evidence location must point to material the reviewer can inspect. A model-generated citation label with no corresponding source is an error. Keep the source packet private when it contains internal material; the example uses no customer information.

Separate absence, ambiguity and contradiction#

These conditions need different handling. Absence means the permitted sources have no answer. Ambiguity means a passage admits more than one interpretation. Contradiction means two supplied sources disagree. Asking the same worker to “be more confident” resolves none of them.

Scroll table →
FixtureExpected worker responseHuman next step
Material explicitly named in the current sheetExtract value and locate passageConfirm source version
Capacity appears only on an older sheetMark current value unknownRequest current product evidence
Two current sheets disagree on finishPreserve both values; abstainAsk the accountable product owner
Lifestyle image suggests outdoor useDo not create a durability claimObtain an explicit approved statement

The fixtures are synthetic. They define expected behavior for a rehearsal; no Haiku run was performed. Check false assertions separately from missing answers. A worker that returns every field but invents one consequential fact has failed the contract.

Move accepted facts into the Migma brief#

The producer signs off each value, source revision and intended use. Only then should the designer ask Migma to build the email using those facts. Include the unresolved list in the brief: “Do not mention dishwasher use, outdoor durability or remaining stock.” Use a pinned reference on the Visual Canvas when it helps the reviewer compare the draft with the accepted product sheet.

Read the draft for additions as well as omissions. “Made for everyday outdoor dining” may introduce a durability implication even when no explicit durability field was supplied. Return that sentence to review rather than treating extraction approval as approval of every later inference.

Measure the handoff, not confident language#

For a pilot, record correct sourced values, unsupported assertions, appropriate abstentions and owner escalations. Keep the extraction packet and resulting Migma version together. Do not use a vendor's general benchmark as the acceptance threshold for these campaign fields.

If the worker cannot preserve evidence locations or abstain reliably, reduce its task or keep extraction manual. Finished-email evaluation belongs in the separate QA calibration guide. Begin with four source cases, approve the fact card, and then create one Migma draft from the accepted brief.

Evidence

Sources behind this page

Claims remain tied to dated source review. Method and corrections stay public.

  1. S-01Migma brand setupdocs.migma.ai
  2. S-02Migma Visual Canvasdocs.migma.ai
  3. S-03Anthropic Haiku 5.5 launchanthropic.com