AI writing for ABM: distinguish better drafts from better meetings

Define a useful message

A useful enterprise message connects an observable business context to a question the recipient can answer. It offers relevant evidence and a clear next step. A message can be grammatically polished while being wrong about the company, the stakeholder or the product.

Before drafting, specify the role, decision stage, verified context, permitted claim and desired response. For an early conversation, the next step may be confirming whether a problem exists. Do not force every message to request a meeting when the evidence supports only a question.

What the writing experiment measured

Noy and Zhang randomly exposed half of 453 college-educated professionals to ChatGPT for occupation-specific writing tasks. The 2023 Science paper reported a 40% reduction in average task time and an 18% increase in assessed output quality.

These are task-level results, not reply, meeting or revenue effects. Our ABM application is to test whether assistance improves accepted message work under your constraints. Product accuracy, source support and relationship suitability need their own review because they were not established by the published headline percentages.

Explore the original methods and findings in Experimental evidence on the productivity effects of generative artificial intelligence.

The practical workflow

1. Specify the buyer decision
2. Freeze the message rubric
3. Review drafts without process labels
4. Measure full production effort
5. Test commercial response separately

Compare the approaches

Create a rubric before comparing drafts

Use criteria such as factual correctness, relevance to the role, usefulness of the question, strength of supporting evidence and clarity of the next action. Decide which errors cause rejection rather than a lower score. An incorrect company identity should not be compensated for by elegant phrasing.

Compare drafts against the same account evidence. Where feasible, hide whether a person or agent wrote the message. Include editing and approval in total production time. Record reasons for rejection so the team can improve the brief or workflow rather than merely regenerate more options.

Compare two fictional openings

A weak draft tells a fictional CFO that rapid growth means their costs are out of control. The company’s expansion announcement supports growth, but it does not support that diagnosis. A better draft asks which cost-to-serve assumptions need updating as the network expands and offers a relevant evaluation worksheet.

The second version changes the buyer task, not just the tone. A technical leader may instead need an integration constraint and a validation step. Generate variants around those decisions. Keep the evidence consistent while changing the question and depth for the recipient.

Keep citation review separate

Citation research shows why fluent language and supported language need different evaluation. A draft can include a real study while borrowing a result from an unrelated population. Check the material assertion behind each link before sending the message.

Once messages pass the quality rubric, test activation using a defined eligible account population. Track replies, attended qualified meetings and progression separately. Match starting conditions and preserve exclusions. A production-time gain and a commercial-response gain may both matter, but they require different denominators.

Explore the original methods and findings in Enabling Large Language Models to Generate Text with Citations.

Improve the brief before increasing volume

If most drafts need the same correction, inspect the input. Missing product evidence, unclear role selection or an unsupported account hypothesis will generate repeated editing work. Add the needed context or narrow the claim before raising throughput.

Keep the best accepted message and the reason it worked as a versioned example. Avoid treating it as a universal template. The next account may have a different constraint, relationship or decision stage that changes what is useful to ask.

For the next part of this decision, read Personalisation and trust: use context the buyer can recognise.

Your next-action checklist

• Writing time: Include editing and approval. Check the limitation: excludes downstream effects.
• Message quality: Add rejection conditions. Check the limitation: rubric can miss context.
• Reply rate: Classify reply relevance. Check the limitation: replies may be unqualified.
• Qualified meetings: Define attended and qualified. Check the limitation: requires comparable account groups.

Use the comparison to choose a bounded next step. Record the evidence, the responsible owner, and the review decision before extending the play to additional accounts.

Questions this guide answers