Inside Gold Bond’s IT-Led AI Rollout: Why Workflow Integration Beats Chatbot Hype

When a 77-year-old promotional products supplier pushes generative AI from experiment to everyday habit, it’s rarely because of a shiny chatbot. At Gold Bond Inc., a major player in the $20.5 billion promotional products industry, CIO Matt Price centered AI adoption on a different question: how do we embed AI into the work employees already dislike and do at scale—ERP intake, document processing, call follow-ups—instead of adding another destination app?

The result was not a viral internal tool but a series of IT-led workflow integrations across systems and roles. Those changes shifted behavior: daily AI usage jumped from 20% to 71%, and 43% of employees reported saving up to two hours per day. The story offers a concrete pattern for CIOs and IT leaders: AI adoption sticks when it is fused into operational workflows, governed centrally, and rolled out via disciplined change management—not when it’s pitched as a standalone assistant.

From ‘chatbot experiment’ to workflow-first AI strategy

Gold Bond, which serves roughly 8,500 active customers with customized swag and corporate gifts, faces a classic operations problem: high-volume, high-variance requests streaming in from every imaginable channel. Orders, quotes, and sample requests arrive via website, email, fax, and other paths, each with different formats and levels of completeness. Price characterized it simply: “So it gets very messy.”

Rather than start with an organization-wide chatbot or generic “AI copilot” push, Price anchored the strategy in these messy workflows. The guiding principle: use AI where manual effort is high, inputs are unstructured, and outcomes are repetitive but critical. That meant targeting:

• ERP intake and data normalization
• Document processing and contract review
• Call summaries and customer follow-ups
• Routine office productivity tasks like presentations and spreadsheets

To surface the right use cases, Price didn’t rely on benchmarks or vendor decks. He built a small cohort of “super-users”—about eight early adopters—to find Gold Bond–specific problems and test tools against them. This group became a bridge between IT and the rest of the business, converting vague AI interest into concrete workflows and then training colleagues once patterns proved useful.

Crucially, AI was not framed as magic. Expectations were “reset” early: models would assist, not autonomously run the business; outputs would be verified, not blindly trusted. That grounded stance helped employees move from skepticism to experimentation, and eventually to regular use.

How Gold Bond embedded AI into high-friction ERP and document workflows

The clearest example of IT-led integration is Gold Bond’s order and ERP pipeline. Historically, staff manually keyed order information into the ERP system from an array of formats and channels—a slow, error-prone process. AI fit naturally because the work was largely structured extraction from unstructured inputs.

Gold Bond’s current flow layers cloud services and multiple models:

• Document ingestion and normalization: Incoming documents are first ingested by Google Cloud, creating a consistent digital starting point regardless of source or format.
• Field extraction with LLMs: Gemini and OpenAI models extract and structure relevant fields—such as customer details, items, quantities, and pricing—into a clean, standardized schema.
• ERP integration: The system then pushes a completed purchase order into the ERP, removing the need for manual re-entry.

This is not “chat with your ERP.” It is AI wired into the critical path of order processing, with IT orchestrating the underlying services and guardrails. For CIOs, the pattern is notable:

• AI is invoked as part of a system-to-system workflow, not just a user-to-chat interface.
• Models are chosen and combined pragmatically—Gemini and OpenAI here, other tools elsewhere—based on task fit.
• The business impact is quantifiable in time saved and error reduction, lending itself to repeatable justification.

Beyond ERP, Gold Bond extended this model to other document-heavy processes. Contracts are reviewed with AI assistance, and internal documentation and procedures are being organized with tools like NotebookLM, which helps build a knowledge base for training and operations.

To validate these integrations, Gold Bond uses Kaizen-style events: short workshops that map baseline workflows and compare them to AI- and automation-assisted variants. This continuous improvement approach gives IT and business stakeholders a shared framework for measuring impact and avoiding one-off “AI pilots” that never scale.

Inside the multi-model stack: Gemini, ChatGPT, Claude, and more

Gold Bond has taken a consciously multi-model approach, avoiding lock-in to a single provider and assigning tools to tasks by strength. “We’re pretty agnostic on utilizing AI technology,” Price said. While the company is primarily a Google shop—working with Google premier partner Promevo on implementation and change management—it mixes several major models and some smaller ones.

The current stack includes:

• Gemini inside Workspace: Because it is embedded where users already work (Gmail, Docs, Sheets), Gemini often serves as the entry point for employees new to AI. This lowers friction and encourages initial experimentation on low-risk tasks like drafting and editing.
• ChatGPT for backend automation: Used behind the scenes for certain automation tasks, ChatGPT complements Gemini where different capabilities or behaviors are needed.
• Claude for QA and reasoning: Claude is used to perform quality checks and reasoning-heavy review, providing a second opinion or verification layer on outputs.
• Smaller models for edge experiments: Lightweight models are used where appropriate for niche or experimental use cases, reflecting an ongoing test-and-learn mindset.

LibraChat plays an important control role: it centralizes access to approved tools, enforces use of paid/authorized models, and allows the IT team to block certain models when necessary. That centralization is crucial for managing a multi-model environment, reducing shadow AI use, and ensuring that governance policies are consistently applied.

For CIOs, the takeaway is less about the brand names and more about the architecture: multiple models sit behind a controlled access layer, with IT curating which tools are available for which use cases. End users see a coherent experience, while IT can switch or augment back-end models as capabilities and requirements evolve.

Real-world use cases beyond chat: from phone calls to product mockups

Gold Bond’s deployment emphasizes concrete, repeatable use cases that directly relieve operational pain. Several have already become everyday tools for staff:

• Phone call summaries: AI-generated summaries of calls help teams track follow-ups, decisions, and customer needs without manual note-taking. This both saves time and improves continuity in customer interactions.
• Email drafting and responses: Employees use AI to draft responses and routine communications, speeding up back-and-forth with customers and internal stakeholders.
• Contract review: AI assists in scanning contracts, highlighting key clauses or areas that need attention. Humans still make the decisions, but they start from a more organized view of the document.
• Virtual product mockups: Using Recraft, teams generate AI-assisted visual mockups of branded products, iterating quickly on design concepts before sending previews to customers. This shortens the sample and approval cycle.
• Presentation creation: Work that once took around four hours now takes about 30 minutes, according to Price, as AI helps structure decks and draft content.
• Code auditing: Developers run NetSuite scripts and then use two separate models to review code before moving into testing, adding an automated sanity-check layer.
• Research and trend tracking: AI helps monitor importer trends and tactics, particularly in response to tariffs, giving the business a faster way to track and synthesize market information.
• Spreadsheet logic generation: Employees use AI to generate formulas in Google Sheets—including complex Excel-style logic like XLOOKUP—lowering the barrier for non-experts to build robust spreadsheets.

Collectively, these use cases explain why employees report significant time savings and why daily AI use has climbed. They cut across white-collar tasks, technical workflows, and customer-facing processes, reinforcing AI as a general-purpose assistant embedded in existing tools and systems rather than a separate destination.

Price notes that AI also accelerates upfront planning. Teams iterate with models to develop high-level project outlines before investing in detailed execution. The practical effect: faster concept development and fewer meetings, with humans still owning decisions and plans.

Change management: building super-users, resetting expectations

At a decades-old company, technology alone cannot deliver adoption. Price underscores that “change management was the work.” Employees were initially apprehensive; many saw AI as something different, potentially disruptive, or overhyped.

Most users begin with Gemini, simply because it’s surfaced where they spend their time: Workspace. From there, as they hit limitations or encounter new needs, they branch into other tools such as ChatGPT, Claude, or Mistral to get different capabilities or a “second opinion.” This gradual path eases people into a multi-model world without overwhelming them.

The “small cool group” of roughly eight early adopters has been central to this change strategy. These super-users:

• Trial bleeding-edge tools and workflows before broader rollout.
• Identify specific, high-value use cases grounded in real work.
• Train colleagues once patterns prove reliable and useful.

Promevo CTO John Pettit summarized the core challenge: You can’t treat AI like any other new software deployment. “You really have to change people’s thoughts and behaviors around it.” That includes addressing misconceptions about AI as a fully autonomous agent and reinforcing its proper role as an assistant that still requires oversight.

Price’s guidance for other enterprises is explicit: avoid being overwhelmed by the hype cycle. Start with simple, well-bounded use cases. Treat prompting as a skill to be learned—“Provide detailed prompting, test it, play around with it”—and give employees space to experiment inside guardrails. The goal is to normalize AI as part of daily work, not as a one-time transformation project.

Governance, guardrails, and the ‘trust but verify’ model

Gold Bond’s AI program is intentionally cautious on governance. While Price’s team encourages widespread usage, blind trust in model outputs is not an option—especially for anything public-facing or business-critical.

Several layers of control are in place:

• Policies and DLP controls: Formal policies define acceptable AI use, complemented by data loss prevention controls and identity layers to reduce shadow AI and protect sensitive data.
• Centralized access via LibreChat: LibreChat acts as a centralized entry point for approved tools, ensuring employees use sanctioned models under the right accounts and licensing. IT can restrict or allow models as requirements evolve.
• Sandbox and QA-first deployments: Multi-model workflows are tested in a sandbox environment, where the technical team and subject matter experts run QA scenarios before any production integration. Changes ship only after both groups sign off.
• Human-in-the-loop as a hard requirement: For content that faces customers or the public, human review is mandatory. Outputs are checked, edited, or rejected as needed.

Price describes the operating philosophy as setting “the right temperature of trust, but verify.” Even with strong prompts and seemingly authoritative outputs, staff are reminded that “you get the data back, you can’t just blatantly take it and use it.” Verification is treated as part of the workflow, not an optional extra step.

Practically, that means asking models for their sources and reasoning—“Give me all the work cited, where you are grabbing this data from”—and then checking that context. For CIOs, this reinforces a critical pattern: governance is as much about training people in healthy skepticism as it is about technical controls.

Lessons for CIOs: start where the pain is, not where the hype is

Gold Bond’s experience offers several concrete takeaways for IT and operations leaders planning or recalibrating enterprise AI rollouts:

• Anchor AI in painful workflows, not generic assistants. Focus first on high-friction tasks—like Gold Bond’s ERP intake and document processing—where AI can clearly remove manual drudgery.
• Use a multi-model stack, centrally governed. Assign different models to different strengths, but route access through a controlled layer (like LibreChat) to avoid fragmentation and shadow use.
• Build a super-user core. A small, trusted group of early adopters can translate technology into domain-specific workflows and drive peer-to-peer training.
• Institutionalize verification. Make human-in-the-loop review, sandbox testing, and QA scenarios standard practice, especially for external content and production workflows.
• Measure with before-and-after workflows. Use structured methods, such as Kaizen-style events, to benchmark time and process improvements rather than relying on anecdotal success stories.
• Right-size ambition. Price cautions that “agentic solutions can only go so far—there still need to be humans in the loop,” and that some visions outpace what current technology can reliably do. Keeping ambitions grounded helps sustain trust and avoid disillusionment.

Gold Bond’s numbers—daily usage up from 20% to 71%, with 43% of employees saving up to two hours per day—are not the result of a single flagship AI project. They come from a series of targeted, IT-led interventions that respect existing workflows, enforce governance, and treat AI as an evolving capability rather than a one-time rollout.

For CIOs, the message is clear: AI adoption succeeds when it’s integrated into the operational fabric of the business, not when it’s launched as a standalone chatbot in search of a problem.