AI Guide: Case Studies

Table of Contents

    Financial Health

    Easier: CalMatters 'Giving Editor'

    Goal: Improve the effectiveness of fundraising appeals while reducing the staff time required to edit and refine them.

    How they did it

    Description: Sonya Quick, CalMatters’ director of membership, created Giving Editor, an AI fundraising coach that provides feedback on donor appeals and helps sharpen those messages.

    Potential Impact: Give staff access to consistent fundraising feedback without requiring a senior colleague to edit drafts, leaving more time for higher-value work. 

    One way to start: 

    1. Identify the specific job for your coach, such as “reviewing and improving fundraising emails.” 
    2. Create a Project in ChatGPT or Claude, and give it a goal: “Act as a fundraising editor and coach. Review my appeals for the strength of the ask, persuasiveness, audience fit, and alignment with our mission. Ask questions and suggest improvements rather than automatically rewriting the copy for me.” 
    3. Give it context about your publication, audience, and fundraising strategy. Include what you already know works from past campaigns (for example: Focus on the impact of our reporting). 
    4. Prompt your coach with a thorough request: “Here is my draft fundraising email to loyal supporters asking them to give $36 to celebrate three years of serving the community. Review it as a fundraising editor. What is working? Where am I losing focus or missing an opportunity to make the case for giving? Ask me any questions you need answered before suggesting specific improvements.” 
    5. Test the AI coach on several real appeals and refine its instructions when the feedback misses the mark. For example, if it consistently recommends copy that is too long for your audience, add an instruction to consider shorter, mobile-friendly appeals.

    Source: Hacks/Hackers conference (May 2026) presentation and notes

    Intermediate: Cityside’s “Reasons for Giving” Analysis

    Goal: Increase reader revenue by understanding what motivates donors and using those insights to improve fundraising strategy.

    How they did it

    Description: Cityside, the nonprofit that runs Berkeleyside, Oaklandside, and Richmondside, used AI to analyze almost 9,000 donor comments over four years to better understand why some of their supporters decided to give. They completed this with assistance from the News Revenue Hub and American Journalism Project. 

    Real-World Impact: Cityside discovered that donors were motivated more by the overall mission and community impact than individual beats and reporters. The finding pushed Cityside away from a complicated strategy built around individual coverage interests and toward broader audience engagement signals and simpler fundraising messages. Notably, the process took one to two hours once the data was ready.

    One way to start: 

    1. Make sure you have qualitative data about why people give. In this example, the News Revenue Hub donation form includes an open-ended “reason for giving” field. 
    2. Remove names, email addresses, and other identifying information, then upload the cleaned data to an AI tool that can analyze spreadsheets. Explain what each column represents and ask it to identify recurring themes in why people give, along with useful patterns in the language or tone of their responses.
    3. Ask the AI to suggest a manageable set of categories based on what actually appears in the responses. Review those categories yourself and combine or rename anything that won’t help inform a decision. For example, a broad category such as ‘liked our journalism’ may be less useful than distinguishing between motivations such as political coverage, local accountability reporting or community information.
    4. Add other non-identifying fields that could help explain the findings, such as gift amount, recurring versus one-time status, date, publication or newsletter engagement — but only when those fields are relevant to the question you’re trying to answer (for example, “Do donors give more at the end of the year?”)
    5. Once the data set is prepared for analysis, start asking the AI questions about it. Keep in mind that the prompts should give you something actionable. Some examples from Cityside include: “What percentage of donors give recurring and one-time gifts?” “What are the most common phrases in the ‘Reasons for giving’ field?”

    Source: Cityside via INMA and the American Journalism Project

    Advanced: The Seattle Times

    Goal: Increase advertising revenue by identifying and prioritizing promising new sales prospects.

    How they did it

    Job To Be Done: Increase advertising revenue by identifying and prioritizing promising new sales prospects.

    Description: The Seattle Times built an AI prospecting tool that helps advertising sales staff identify and research potential new clients. It evaluates prospects against criteria defined by the sales team and prepares background research and suggested questions so sales reps can decide which leads to pursue and arrive better prepared for conversations.

    Real-World Impact: Shortly after launch, a sales rep used the tool to identify an unexpected prospect that had not previously been on the team’s radar and converted it into a sale.

    One way to start: 

    1. Define what makes a strong advertising prospect. Look at successful current and past advertisers and identify useful signals: industry, geography, business size, advertising needs, past spending, or other characteristics. If you have structured sales data, an AI tool can help you look for patterns across those accounts.
    2. Create a Project and add relevant background materials, such as your rate card, advertising calendar, customer profiles, and sales criteria. If your organization permits it, you could also include an appropriately secured or de-identified CRM export.
    3. Tell the AI exactly how to evaluate a prospect. Create a simple scoring rubric (1-10, or 1-50)based on the criteria you identified above — for example, market fit, geographic relevance, likely advertising need, and compatibility with your products. Ask it to show the reasoning behind every score.
    4. When you open the project, ask the chat for a report. The prompt can be something like this: “I’m seeking potential advertisers for our morning news email. Using the criteria and client information in this Project, research 10 organizations that could be a strong fit. For each one, explain why it matches our criteria, assign a score using the rubric I provided, and cite the public information you used. Flag anything I should verify before contacting them.” 
    5. Validate any AI-generated suggestions against your internal records before you begin outreach. You should confirm that potential leads aren’t currently active in your CRM or already established advertising partners and double-check the specific data points used to validate a prospect’s rating.

    Source: The Lenfest Institute

    Journalistic Impact

    Easier: News in the Grove’s Automated Source Follow-Up

    Goal: Expand the reach of published stories by automatically following up with the people and organizations featured in them.

    How they did it

    Job To Be Done: Expand the reach of published stories by automatically following up with the people and organizations featured in them.

    Description: Chas Hundley, publisher of News in the Grove, uses Claude to scan newly published stories for people and organizations mentioned in the coverage and follow up with them by email. The workflow cuts down the time required to identify and contact everyone featured in a story. Hundley reviews this outreach as part of the publishing process.

    Potential Impact: While generating advertising revenue wasn’t the goal, Hundley says some people contacted about stories featuring them have subsequently bought advertising.

    One way to start: 

    1. Ask ChatGPT Work or Claude Cowork to review stories published on your website during a set period — for example, the last 24 hours. Have it create a table listing each story, its URL, and the people or organizations materially featured in it. Define what counts as “materially featured,” such as being quoted or discussed for at least one paragraph.
    2. If your organization permits it, connect your email account and ask the AI to look for contact information for the people and organizations identified in Step 1. Have it add any addresses it finds to the table. Verify the addresses before using them, and double-check that every person or organization on the list should receive an email. 
    3. Ask the AI to draft a short email for each recipient with a link to the story. You might use different templates depending on the circumstances: For example, Chas Hundley has different email text to share positive or community-oriented coverage versus a neutral notification for other stories. Review every message and recipient before sending.
    4. Once you’re satisfied with the workflow, schedule it to run automatically — daily, weekly, or on another useful cadence. Review the first several runs closely before relying on it as a routine.

    Source: News in the Grove

    Intermediate: Mountain State Spotlight’s AI-Generated Audio Stories

    Goal: Make reporting more accessible to readers with visual impairment.

    How they did it

    Job To Be Done: Make reporting more accessible to readers with visual impairment. In West Virginia, at least 1 in 30 people suffer from vision loss – a fact surfaced by the publication’s Innovation Fellow. 

    Description: Mountain State Spotlight uses Everlit, an AI text-to-audio tool available through Newspack, to create audio versions of select stories. The newsroom receives five free audio articles per month. Before publication, staff listen to each recording to catch errors and correct mispronunciations. 

    Potential Impact: Make reporting accessible in another format for people who have difficulty reading text, while also giving other audiences the option to listen instead of read.

    One way to start: 

    1. Start with a small number of stories rather than converting everything to audio. Choose stories that are likely to benefit from the format, and use the test period to evaluate accuracy and audience response.
    2. Prepare a story script for text-to-speech. Remove anything that should not be read aloud, such as photo captions, navigation text, or unrelated page elements. Make a list of names, places, and acronyms that may require pronunciation guidance. Be sure to instruct the AI to not change the reporting itself.
    3. Use a text-to-speech service such as Everlit or ElevenLabs to generate the recording. Before choosing a tool, make sure you can export or embed the audio in a format that works with your website.
    4. Have a journalist listen to the entire recording before publishing. Check names, places, acronyms, numbers, and other words the AI may pronounce incorrectly, as well as any places where the audio does not accurately reflect the published story.
    5. Tell your audience that the recording uses an AI-generated voice, and give readers a way to report audio errors.

    Source: Mountain State Spotlight

    Advanced: Technical.ly’s Layoff Alerts

    Goal: Surface story leads from regular reports and data sources.

    How they did it

    Job To Be Done: Surface story leads from regular reports and data sources. 

    Description: Technical.ly built a Claude Code automation that checks WARN and other layoff notices each week in states the publication covers. The system filters the results and emails relevant notices to the appropriate reporters and editors.

    Real-World Impact: Within a few weeks, information surfaced by the automation led to a story about potential layoffs at a defense contractor in Washington, D.C.

    One way to start: 

    1. Make a list of the public sources you want to monitor. Be specific about the agencies, geographic areas, or topics that matter to your coverage. For each source, record the agency or organization, URL, geography, and a short note explaining what information it publishes and why it matters.
    2. Give the source list to ChatGPT Work or Claude Cowork and explain exactly what you want it to look for. For example: “Check these sources for newly posted notices that could produce story leads for our newsroom. Tell me if you cannot access any source. Include only notices that meet the criteria below…” You’ll want to be even more specific about what you’re seeking, for example: “Include only notices in these counties.”
    3. Specify the deliverable and format, including fields such as date, organization, location, description, why it was included, and a link to the original source. For example, ask for the results in a spreadsheet.
    4. Test the process manually several times. Once you’re satisfied with the results, schedule it to run at an appropriate interval — weekly for slower-moving sources or more frequently for time-sensitive information. Ask it to include only information published or changed since the previous run.

    Note: A monitoring workflow that checks many websites frequently can consume a significant amount of tokens or AI usage. Start with a small number of high-value sources and a reasonable schedule, then expand if the results justify it. Depending on your plan, you may hit usage limits or incur additional charges for usage beyond what is included. Check your account’s usage and billing settings, and set spending limits or alerts when those options are available.

    Source: How LIONs are Using AI (June 2026)

    Operational Resilience

    Easier: New York Times’ Alt-Text Generator

    Goal: Make visual journalism more accessible to readers who rely on alt text, while reducing the staff time required to write it.

    How they did it

    Job To Be Done: Make visual journalism more accessible to readers who rely on alt text, while reducing the staff time required to write it.

    Description: The New York Times used Gemini to draft alt text for images, then developed and tested a workflow to bring those AI-generated drafts into its publishing process. Editors evaluated the results for accuracy, clarity, essential information, and sensitivity before the tool was piloted in the newsroom’s CMS. 

    Potential Impact: Make it easier to provide high-quality alt text consistently, without adding as much work for newsroom staff.

    One way to start: 

    1. Create a Project or, if your plan allows it, a custom GPT called “Alt-Text Assistant.” Give it detailed instructions, such as “Draft concise alt text that communicates the most important information in the image. Complement the photo caption rather than repeating it. Use clear, neutral language.” 
    2. Give it guardrails, such as “Do not infer details that cannot be established from the image” and “Flag anything you are unsure about rather than guessing.” Ask it to provide at least two options for a human to review.
    3. Give it examples of alt text your newsroom considers strong. Include a range of image types, if possible, and instruct the AI to use them as a guide for length and approach.
    4. Test the assistant on a batch of images before using it routinely. Review each result for accuracy, essential information, clarity, and inappropriate assumptions. If you see recurring problems — for example, descriptions that are consistently too long — revise the instructions and test again.
    5. Have a journalist or editor review and, if necessary, edit AI-generated descriptions before publication.

    Source: NYT Masterclass: Quantifying Feels A.I. Evaluation

    Intermediate: Enlace Latino NC’s AI Translation Workflow

    Goal: Expand the potential audience for reporting by making Spanish-language journalism available in English.

    How they did it

    Description: Although Enlace Latino NC primarily serves Spanish-speaking audiences, the publication wanted to make more of its reporting available in English. After an automated translation plugin failed to meet their standards, the team built a custom GPT with detailed instructions for translating news copy, including guidance on direct quotes and AP style.

    Real-World Impact: The workflow helped Enlace Latino NC produce more English-language journalism, including increasing the frequency of its English-language newsletter from once to twice a month.

    One way to start: 

    1. Create a custom GPT or Claude Project and give it a specific role, such as: “You are a Spanish-to-English translator for a news publication serving the Puerto Rican diaspora across the U.S. Accuracy is the highest priority. Preserve the meaning, tone, facts, and attribution of the original reporting. Do not add, omit, or infer information.”
    2. Create a context file with your publication’s style rules, preferred translations, proper names, place names, and guidance for handling quotes, titles, and terminology. If you have a stylebook, add that too.
    3. Test the workflow on a set of previously published stories that a fluent speaker can evaluate. Pay particular attention to quotes, names, numbers, idioms, and passages where a literal translation could alter the meaning. When you identify recurring problems, add the correct approach to your instructions or style file and test again.
    4. Have a journalist who is fluent in both languages review the translation against the original before publication.

    Source: How LIONs are Using AI (June 2026)

    Advanced: Moab Sun’s Virtual Ops Team

    Goal: Increase staff capacity by using AI to handle a range of recurring administrative and operational tasks.

    How they did it

    Job To Be Done: Increase staff capacity by using AI to handle a range of recurring administrative and operational tasks.

    Description: Moab Sun News publisher Maggie McGuire uses Claude Code as a kind of virtual operations team, connecting it to the tools she uses to run the publication. She has documented 24 recurring workflows as “skills” that help Claude handle tasks including advertising operations, newsletter production, scheduling, records tracking, and website management.

    Real-World Impact: McGuire says some routine tasks take substantially less time; assembling the weekly newsletter, for example, dropped from about 45 minutes to 15. The system has also helped her document 24 recurring workflows that previously depended on one person knowing how to do them.

    One way to start: 

    1. Pick one repetitive operational task that takes up more time than it should. Start with advertising: Which advertisers need a response? Which contracts are approaching renewal? Which leads haven’t been contacted recently?
    2. Ideally, keep this information in a spreadsheet or CRM that your AI application can access, such as Google Sheets or Airtable. You can also export the information as a CSV and upload it, but remember that an uploaded file is only a snapshot of the data at that moment.
    3. Write out the workflow step by step. For example: “Identify advertisers who have not replied in a week, check when they were last contacted, and draft a follow-up email.” If your AI tool supports reusable Skills, save the process as a Skill; otherwise, keep these directions as Project instructions or a reference document.
    4. Connect the tools the workflow needs. For example, you could give the AI appropriate access to your email or calendar so it can check when an advertiser was last contacted.
    5. Define the deliverable. For example: “Create a spreadsheet listing every client who needs follow-up this week. Include the client name, date of last contact, reason for follow-up, and a draft email for me to review.”
    6. Test the workflow several times and refine the instructions — for example, “only include clients who have advertised with us at least twice” or “use a warm, concise tone.” Review any proposed actions before allowing the system to send emails or change records. Once you trust the workflow, consider scheduling it to run once or twice a week.

    Source: Maggie McGuire, “What a Journalist and an AI Built Together,” February 2026

    Additional Reading and Resources

    Generative AI was used to refine the writing and check content on this page. All content was created, reviewed, and verified by a human, who takes full responsibility for its accuracy.