Geoffrey Smolders

The through-line

Across these posts, Geoffrey Smolders keeps returning to one idea: the hard part of AI is never the model, it is what surrounds it, context, governance, permissions, and where the work actually happens. Whether he is talking about a Salesforce headless cockpit, an agentic discovery tool, or a Claude bot dropped into Slack, the argument is the same, capability was mostly already there, the unlock is the form factor and the trust layer around it [2][13]. This shows up first in his sales thinking (discovery decides deals, not top-of-funnel volume) [6], then increasingly in his product and delivery thinking as Biztory shifts from a Snowflake/Tableau analytics partner [1][16][17] toward building and shipping agentic tools on Claude and Anthropic through mid-to-late 2026 [7][8][9][11][12][13]. Over the arc of the sources there is a visible move from "we implement platforms" toward "we build governed agents that reason alongside consultants," with a parallel, more cautionary voice warning that AI making things easy to start does not make them safe to run [5].

On discovery

Smolders treats discovery, both in B2B sales and in enterprise AI deployment, as the actual site where value is won or lost, not the flashy part everyone optimizes for. On sales: "Discovery is different. It's where deals are actually won or lost" and "The best rep on your team asks at least 10-15 questions in a discovery call. The average rep asks maybe 2-3," while "We pour budget into filling the pipe" instead of fixing that gap [6]. He carries the same logic into how Biztory builds AI: "Agentic AI fails for one simple reason: it tries to do the work before understanding the business," which is why Biztory's own agents "interview in depth, trace your workflows, navigate your fragmented data silos" before any deployment [15]. This thinking is operationalized directly in Biztory Scout, launched for client discovery work with "overwhelmingly positive" feedback and its "first billable engagement" [7], and in the Agentic Enterprise Scout that interviews customers on their processes and feeds it into "years of Biztory experience" [11].

On governance and trust

A second recurring theme is that AI is only useful if it is governed, auditable, and bounded, otherwise it is a liability. His headless Salesforce experiment is explicit about this: "the interface isn't what matters, trusted context, identity, permissions and governed action are," and the reasoning running "inside Salesforce's Einstein Trust Layer... is what makes it enterprise-grade rather than a toy" [2]. He extends the same caution to the wave of AI-built internal tools, warning "AI did not make building easy. It made starting something way easier" and drawing a direct line to the Microsoft Access era of ungoverned shadow apps, concluding "The tools changed. The pattern did not," with a simple rule of thumb: "If it holds your customers, your money, or your regulated data, buy it" [5]. At Digital Workforce this becomes a formal method, agents are not deployed as "plug-and-play" tools but "onboard them like you would treat an employee," with "a salary," a "probation period," and graduation before scale [15].

On where AI actually sits

Smolders is preoccupied with placement, arguing that moving AI into the tool people already use, rather than a separate destination, is what changes adoption. Adding a Claude tag to Slack, he writes, "What changed last week wasn't the model. It was where it sits," removing the "friction" of "stop what you are doing, open another tab, copy your context across, come back," so that "The form factor is the unlock" [13]. He makes the identical claim about interfaces in his Salesforce headless build, "The Helm" running live demo data with role switching, Data360 for grounding, and Agentforce for reasoning, concluding twice, in near-identical language across two posts, "The future of Headless isn't around the corner. It's here" [2][3].

On naming and framing

Smolders is also attentive to how mislabeling a category distorts how people think about it. On embedded analytics, he argues the term itself is misleading and that it "should be called 'external analytics'," since it is really about sharing interactive dashboards with customers and suppliers, not a purely technical embedding exercise, done to measure engagement and open new revenue streams [16].

Takeaways

• When evaluating an AI tool, ask where it sits in the existing workflow before asking what model powers it, placement removed more friction than any capability upgrade did [13].
• Apply a governance filter before letting AI-built tools near production: if it touches customer data, money, or regulated data, buy rather than let it grow organically [5].
• Treat discovery as a deliberate discipline, not an afterthought, both in sales calls and in agentic AI rollouts, understanding the business before deploying anything is the difference between success and failure [6][15].
• Onboard AI agents with the same rigor as employees, defined scope, a trial period, and a graduation gate, rather than dropping them into legacy systems and expecting ROI [15].
• Don't assume "embedded" analytics is just a technical integration question, reframe it as "external analytics" aimed at customer and supplier engagement and new revenue [16].