At 77-year-old promotional products company Gold Bond Inc., CIO Matt Price knew generative AI adoption wouldn’t come from rolling out a chatbot. Employees needed AI embedded into the work they already hated doing: messy ERP intake, document processing, and call follow-ups. Instead of pitching benchmarks, Price built a small group of “super-users” to surface Gold Bond–specific examples and train the rest of the org. They then wired Gemini and other models into high-friction workflows, backed by sandbox testing, guardrails, and human review for anything public-facing. The payoff showed up as behavior change, not hype: Daily AI usage rose from 20% to 71%, and 43% of employees reported saving up to two hours a day. “I wanted to bring everybody on the journey,” Price told VentureBeat. “After we reset some expectations, people started leaning towards it. Our adoption has taken off.” ERP streamlining, product visualizations Gold Bond, Inc. — not to be mistaken with the skin care company — is one of the largest suppliers in the $20.5 billion promotional products industry, producing custom swag and corporate gifts for 8,500 active customers. Orders, quotes, and sample requests arrive via the website, email, fax, and more — in every format imaginable. “So it gets very messy,” Price said. AI proved a natural fit. Previously, employees manually keyed order details into the ERP. Now, Google Cloud ingests incoming documents and normalizes them, while Gemini and OpenAI extract and structure the fields before pushing a completed purchase order into the system, Price said. From there, Gold Bond expanded into a pragmatic multi-model approach: Gemini inside Workspace, ChatGPT for backend automation, Claude for QA/reasoning checks, and smaller models for edge experiments. "We’re pretty agnostic on utilizing AI technology,” Price said. Gold Bond is largely set up as a Google shop, with implementation and change management led by Google premier partner Promevo.  Early wins included phone call summaries, email drafting, and contract review. A more advanced use case is AI-assisted “virtual mockups” of branded products; teams use Recraft to iterate on sample visuals before sending previews to customers, Price said. Employees also use AI to generate Google Sheets formulas (including Excel-style XLOOKUP logic), while NotebookLM helps build an internal knowledge base for procedures and training. Other ways Gold Bond uses AI internally: Presentations: Work that took four hours now takes about 30 minutes, Price said.  Code auditing: Developers run NetSuite scripts, then use two models to review them before moving to testing. Research: Tracking importer trends and tactics in response to tariffs. AI also compresses early-stage planning. “We go back and forth with AI and come up with a high level project that we can then build out for execution,” Price explained. “We get to concepts a lot quicker. We have a lot fewer meetings, which is great.” To quantify impact, Price’s team runs Kaizen events — short workshops that document baseline workflows and compare them with AI- and automation-assisted versions. To validate multi-LLM workflows, Gold Bond tests changes in a sandbox environment and runs QA scenarios before rollout. “Our technical team, along with the subject matter experts, sign off prior to shipping the changes or integrating to production,” Price said. Change management is a must Adoption wasn’t automatic — at a legacy company, change management was the work. “It's just apprehension a little bit, it's something different,” Price said.  Most users start with Gemini because it’s built into Workspace, then move to ChatGPT, Claude, or Mistral when they need different capabilities — or a second opinion. Price relies on a “small cool group” of about eight early adopters to test bleeding-edge tools; once they land a use case, they train the rest of the team. “You can't just look at something like a new piece of software," noted Promevo CTO John Pettit. "You really have to change people's thoughts and behaviors around it.” But even as Price's team is promoting widespread use, blind trust is not an option, he emphasized. Gold Bond added policies, DLP controls, and identity layers to reduce shadow AI use. It also uses LibreChat to centralize access to approved tools, enforce paid/approved usage, and block certain models when needed. Human-in-the-loop is mandatory: Public-facing content goes through approval, and outputs must be verified. “You have to set the right temperature of trust, but verify,” he said. Even with strong prompts, outputs still require verification. “You get the data back, you can't just blatantly take it and use it.” For instance, he’ll ask for sources and reasoning — “Give me all the work cited, where you are grabbing this data from” — and treats that verification step as part of the workflow, he said. Price also cautioned against overreach. “Agentic solutions can only go so far — there still need to be humans in the loop,” he said. “Some people have bigger visions than what the tech is capable of.” His advice for other enterprises: Don’t overwhelm yourself with the hype. Start simple. Start basic. “Provide detailed prompting, test it, play around with it.”