Conversations from the trenches of enterprise AI adoption
Between keynotes & code lies the messy middle of AI adoption. BoxWorks expo hall revealed the gap between vision & daily implementation reality. Platform integrators debugging MCP servers. Financial services teams modeling AI unit costs. Government agencies training police officers. Box consultants building brand voice agents.
The conversations expose a truth: education is the bottleneck. Not model capabilities. Not infrastructure. Not even governance. The challenge is bridging what AI can do with what practitioners need to learn to make it work.
The Platform Integration Reality
Vishal Naik walked me through the MCP server differences between ChatGPT & Claude connectors. “The Box connector in ChatGPT doesn’t do the same level of functionality,” he explained. “It’s just using our public search APIs. The MCP connector lets you choose which specific API calls you want to wrap.”
The learning gap centers on explicit vs implicit triggering. “I used to work on Google Assistant. We always referenced whether it’s making the call based on implicit or explicit triggering,” Vishal noted. “You have to say ‘search in Box’ to trigger the connector.”
ChatGPT announced their MCP server the day we spoke. “You’ll be able to do that same functionality on the MCP server where you have those 10 or 11 different API calls that you can choose,” Vishal predicted.
Cost & security considerations shape usage patterns. “Box is known for security & robust practices,” Vishal observed. “But it’s kind of a black box of model training. There’s like, who owns my information? Where does it go?”
The solution integrates technical & educational approaches. “If you use an OpenAI model inside of Box AI, then you’re within the Box environment. Our data protection follows through.”
How do we build mental models for protocol selection that scale beyond individual developer knowledge?
Takeaway: Developers need conceptual frameworks for when different protocols activate, not just technical documentation. The mental model matters more than the API specification.
Financial Services: Precision Meets Consumption Economics
Cian Mulderrig demonstrated client onboarding workflows that span form submission, folder templating, document classification, metadata extraction & third-party verification via MCP servers.
“We can take that information & structure it in Box however you want,” Cian explained. “Hey, it’s a deal. I want this structure. Hey, it’s real estate. I want that structure.”
The educational challenge compounds rapidly. “No one knows AI,” Cian said. “What’s true today is not true two weeks from now. I don’t think anyone’s truly an expert. We can all be experts in how it works technically, but what it actually means for end users changes completely.”
Financial firms face a practical forcing function. “Even if you don’t want to do AI personally, you sort of have to, because if you don’t give it to your clients, they’re just going to download that file & upload it to public ChatGPT.”
Cost modeling becomes the new financial literacy requirement. “Between Box core product vs AI units vs APIs, but now add in my Salesforce agent, ServiceNow agent. All of these things are based on consumption,” Cian noted. “That’s going to be an opportunity for customer education.”
The shift from seat-based to consumption-based pricing demands new skills. Teams must understand thinking budgets, model costs per page, API call charges vs AI unit charges.
When does cost modeling shift from IT concern to business strategy requirement?
Takeaway: Financial literacy for AI consumption becomes as critical as technical implementation skills. ROI measurement requires understanding unit economics at the workflow level.
Government: Training at the Point of Need
Shawn Deines works public sector implementations. The Metropolitan Police Service in London uses Box for dash cam footage uploads. “Fill out this online form. Upload your dash cam footage. We’ll investigate it.”
The constraint: police officers have no time for comprehensive training. “When do they go on shift? 8:00 AM. Cool, we’ll be there at 7:30 before they go on shift. Quick 15 minutes.”
This works because it respects operational reality. “Police officers have no time & couldn’t care less about technology. How are you going to teach these police officers?”
The same pattern applies across government agencies. “Traditional workloads are still in the process of getting moved into the cloud. That’s where a lot of those initiatives overlap with where Box plays nicely.”
Integration challenges require consulting partnerships. “They’ll work hand in hand with the agency customers or city employees that are on the IT side. Sometimes our customers don’t necessarily know a lot about Box, but we work with partners who do.”
What workflows require just-in-time learning vs comprehensive training programs?
Takeaway: Education must be embedded in workflows, not separate from them. Micro-interventions at decision points beat hour-long courses for operational users.
Box Internal: Brand Voice Agents & Consulting Evolution
Eric Boozer from Box Consulting showed how they solve the “blank page” problem with internal agents. The brand voice agent addresses a specific pain point: “It took copywriters hours to go through top to bottom, every word, every punctuation” to ensure brand compliance.
Now: “I want to make sure this is in the Box brand voice. Look out for these things. Boom, it spits out an exact copy of either what’s wrong or what it should be. What used to take four hours takes seconds.”
Agent construction bypasses traditional development complexity. “In AI Studio, choose a model & give it custom instructions. You are an agent dedicated to reviewing. Here are our guidelines.” Simple system prompt. No coding required.
Scaling requires rethinking educational hiring practices. “We don’t always look for someone with technical background,” Eric explained about education consultants. “We can teach them how to use Box. They have this natural ability to understand how employees learn things over a period of time.”
The consulting team now includes agent development in statements of work. “Build up to three agents, or assign up to 50 hours to building agents.” Customers want this capability but need guidance on application selection.
How do we identify which processes benefit from agent automation vs human expertise?
Takeaway: Agent development becomes a collaborative consulting practice, not a technical implementation project. Success requires understanding learning patterns, not just prompt engineering.
Data Extraction: Overcoming the Blank Page Problem
Karen Hidalgo (Lyter) & Herman Sheynin demonstrated how the “blank page” problem paralyzes metadata strategy. “The number one question is, how do I even know what fields to put here?” Karen explained.
Their solution inverts traditional schema design. “Take 1,000-5,000 documents, put them in a Box Hub & ask Box AI what it thinks should be in metadata.”
This AI-assisted discovery approach eliminates the paralysis of starting with empty field definitions. Let AI show you what’s possible in your data, then add governance constraints.
Cost awareness becomes essential for project planning. “We charge during the extract process,” Herman noted. “Back of the envelope, it’s exactly one AI unit per page.” For 50 five-page documents, that’s 250 AI units. Understanding this math enables realistic scoping.
The educational pattern: Start with AI-assisted discovery, then move to structured templates. Don’t begin with complex field definitions.
What other “blank page” problems can AI-assisted discovery solve in enterprise workflows?
Takeaway: AI-assisted discovery transforms the blank page problem from paralysis into exploration. Show possibilities first, add constraints second.
The Federation Reality: Platform vs Product Strategy
These individual implementation challenges connect to a broader architectural shift. We’re building toward federated agent ecosystems where specialized agents coordinate across organizational boundaries.
Box’s positioning becomes clear in this context. “We’re a platform for content, not an AI company,” multiple Box team members emphasized. The distinction matters strategically.
In federated architectures, Box provides the content intelligence layer while agents from Salesforce, ServiceNow & others handle domain-specific workflows. Integration protocols like A2A & MCP enable this coordination.
The educational implication: teams need to think in terms of agent orchestration, not individual AI features. The unit of analysis shifts from “what can this model do?” to “how do these agents coordinate?”
Implementation Patterns That Scale
Start with AI-assisted discovery, then add constraints. Karen’s metadata field suggestion approach. Eric’s brand voice agent solving one specific problem well. Let AI show possibilities, then add governance.
Embed education in workflows, not separate training. Shawn’s 15-minute police officer training. Vishal’s explicit triggering patterns. Meet people at the point of decision, not in conference rooms.
Build financial literacy for consumption economics. Cian’s observation about AI unit complexity. Teams need cost modeling skills before they start building agents.
Scale through consulting partnerships, not DIY approaches. Government agencies working with implementation partners. Box consulting including agent development in SOWs.
Which implementation patterns require internal capability vs external partnership?
The Educational Infrastructure Transformation
Traditional corporate learning doesn’t match AI adoption patterns. Hour-long courses become obsolete when technology changes weekly. Feature documentation fails when people need conceptual frameworks.
What works:
Micro-interventions at decision points
AI-assisted discovery of possibilities
Cost modeling integrated into planning
Agent development as collaborative consulting
Just-in-time contextual guidance
What doesn’t:
Comprehensive upfront training
Feature-focused documentation without mental models
Separate AI education divorced from business workflows
Assuming technical background for all users
The maturity curve demands different educational interventions at different stages. Early adopters need exploration frameworks. Production teams need governance patterns. End users need workflow integration.
The mandate becomes clear: Organizations must build educational infrastructure that matches AI’s pace of change. Static training programs cannot serve dynamic technology adoption.
Bridge to Federation
The expo hall proved that BoxWorks isn’t just product announcements. It’s the hard work of implementation that enables federated agent architectures.
These aren’t isolated deployment challenges. They’re the building blocks of content intelligence at enterprise scale. The MCP server debugging, cost modeling complexity, training constraints & agent development patterns all contribute to the same outcome: making AI systems that organizations can trust & scale.
For developers: Mental models for protocol selection, cost optimization patterns, agent design principles that compose into larger systems.
For business users: Workflow integration approaches that survive technology change, human-in-the-loop patterns that scale, ROI measurement frameworks for consumption economics.
For executives: Strategic positioning in federated architectures, vendor evaluation criteria that prioritize interoperability, organizational change management for continuous learning.
The technology is ready. The protocols exist. The educational infrastructure is evolving to match the implementation reality.
Part 2 of 4 in the BoxWorks 2025 series. Next: VP panel insights on enterprise rollout strategies & organizational change management.
