Building an AI recognition program from scratch

The AI recognition program was a four-week internal program designed to surface and celebrate AI innovation happening across a large marketing organization. There was no existing template. I built the program concept, mechanics, and technology from the ground up.

The core challenge was designing something that felt substantive, not performative. Anyone can run a "share your AI tip" campaign. The goal here was to create a credible, peer-reviewed evaluation process that produced defensible winners and made the underlying work visible to the whole organization.

Role

Program Lead: strategy, design, technical build, communications, and operations

Organization

Marketing and Communications function at a global professional services firm

Timeline

4-week campaign, August 2026. Built in parallel with full-time program delivery.

Team

Core team of 3: program lead (me), a PM handling submission intake, and a communications partner for post-campaign. 11 human judges coordinated externally.

Award structure

4 categories: Best AI Innovation, Best Agent, Biggest Time Saver, Best Team Submission. Scoring rubric: Impact 50%, Reusability 30%, Creativity 20%.

Platform constraints

Built on a new internal app platform that was still in beta, plus SharePoint, with significant technical limitations on storage, scripting, and layout.

The program had three distinct operational phases: submission, scoring, and announcement. Each had its own system of tools, stakeholders, and communications. Building the technical infrastructure while also running the program required clear sequencing.

Most of the decisions that shaped this program aren't visible in the final product. They were made upstream, under constraint, and they defined everything downstream.

1. Decision 01

   Structured rubric over popular vote

   Early designs considered a community voting format. I ruled it out because popularity and quality are different things, especially for technical AI work that a general audience might not be positioned to evaluate. A weighted rubric (Impact 50%, Reusability 30%, Creativity 20%) gave judges a consistent frame and produced scores that could be defended. The rubric also became the foundation for the AI judge's prompt design.

2. Decision 02

   Claude as a 12th equal judge, not an advisor

   Including AI as an advisory input would have been safe but uninteresting. The decision to give Claude an equal vote, scored against the same rubric with the same submission data, was deliberate. The goal was to test whether AI could apply structured evaluative criteria consistently across 85 entries, and to model collaborative human-AI decision-making in a real operational context. It held up: among the judges who submitted scores, Claude's sat closest to the rest of the panel on average, while still disagreeing sharply on individual entries. Those disagreements turned out to be the most useful part of the experiment.

3. Decision 03

   Building the scoring app instead of using a spreadsheet

   A shared Excel sheet would have worked. I chose to build a custom HTML application on the firm's new internal app platform, then still in beta, because it gave judges a single interface with all submission context visible while scoring: no tab-switching, no version confusion, no partial data. The app also allowed Claude's admin view to be gated behind a URL parameter, keeping the human and AI scoring operationally separate.

4. Decision 04

   File storage over database storage

   The beta platform's built-in database API was the obvious choice for persisting judge scores across sessions. After testing, it proved unreliable for shared cross-session state. I diagnosed the issue, identified the platform's file storage API as a working alternative, and re-architected the storage layer mid-build using JSON files keyed per judge. This required a full audit of key sanitization rules (alphanumeric, dots, hyphens only) to prevent silent save failures.

5. Decision 05

   Keeping winner selection with the judging panel

   An earlier version of the process added a leadership shortlist review before winners were finalized. I removed that step. Leadership did not score submissions; the 11-person human panel plus the AI judge already provided a strong, rubric-based signal. A separate senior review layer would have added scheduling risk, diluted judge authority, and blurred what the panel vote meant. Winners came directly from the scores. Clean, fast, defensible.

6. Decision 06

   Product testing a beta platform alongside its product manager

   When the beta platform's storage layer failed mid-build, I treated it as a product test, not just a blocker. I documented what broke, how to reproduce it, what I had already tried, and why each attempt failed, then brought it to the platform's product manager. We worked through it together, and the fix improved the platform for every team using it, not just this program. Diagnose, document, escalate with specifics, fix at the root: that is the difference between a workaround and a contribution.

7. Decision 07

   Personalized winner communications over a mass announcement

   Each winner received an individual email before the organization-wide announcement, with their specific judge rationale, their points, and a heads-up that the public post was coming. This sequencing was intentional: winners should hear it first, directly, in a way that felt personal rather than corporate. The rationale was generated inside the scoring app itself, using the Anthropic API and prompts refined to produce a single quotable sentence in a panel-judge voice.

All three applications were built without a development environment, external libraries, or dedicated engineering support. The technical work happened alongside program delivery, using HTML, CSS, and JavaScript on corporate platforms with real-world constraints, one of which was still in beta.

85 submissions across a four-week window represents strong internal participation for a program with no precedent and no mandatory requirement to enter. Getting there required a deliberate multi-channel promotion strategy designed around the organization's actual communication habits.

1. Channel 01

   Leadership activation

   Looped in senior leadership early and asked them to encourage their teams directly. Leadership-visible programs signal organizational priority: teams pay attention when their managing directors are paying attention. This required getting leadership buy-in before the campaign launched, not after.

2. Channel 02

   Internal social + email

   Ran coordinated posts through Viva Engage (internal social platform) and the firm's internal newsletter. Copy was written to lead with the award structure and recognition points, not just the campaign concept, giving people a concrete reason to act.

3. Channel 03

   Intranet hub presence

   Published the program on the marketing organization's SharePoint homepage, which serves as the team's primary digital hub. This created a persistent, findable landing point throughout the campaign rather than a one-time announcement that dropped out of view.

85

Submissions received across 4 weeks

12

Member judging panel: 11 human judges and Claude

55+

App versions shipped from zero to production

4

Category winners with full scoring documentation

• Delivered the full program end-to-end with a core team of 3 and no dedicated engineering team
• Published a live SharePoint leaderboard accessible to the full marketing organization, with search, filter, and submission detail views
• Introduced AI as an operational judge in an internal recognition program, a first for the organization
• Product tested a beta platform in a live program and partnered with its product manager on fixes that improved it for every team
• Executed a clean post-campaign handoff with winner spotlights, editorial content, and a structured comms transition
• Built a reusable internal tooling pattern (custom scoring app with integrated AI rationale, embedded in SharePoint) applicable to future programs