Meeting Prep

Situation Read

BlueAlly is entering a two-track engagement with distinct risk profiles. Track one is guest-facing: parking guidance and line-queuing bots operating in physical space, in real time, in front of paying customers and their children. Track two is internal: a fish nutrition inventory system spanning procurement and animal behavior analysis, operating against health and welfare data with regulatory exposure (USDA/AZA-adjacent recordkeeping, though not confirmed in source material). These require different governance models, not one AI strategy.

The archived intelligence this month is dominated by a single argument from Nate B. Jones: a single-action agent that answers a question and exits "is not enterprise AI, it is demos." Both aquarium use cases risk landing in demo territory if scoped narrowly. The guest bot could become a scripted FAQ widget; the nutrition system could become a database lookup with a chat wrapper. Neither meets the bar Jones lays out for AI that compounds organizational value rather than plateauing at a novelty deployment.

Separately, the market is lowering the barrier to hosted agent infrastructure (Matthew Berman on Perplexity Computer) while enterprise voices (Jones, Emma/OpenAI, Miessler) are converging on the idea that the hard problem is no longer model capability but coordination, memory, and grounding. That reframing should shape how this is pitched: not "which model," but "what pipeline, what memory, what human gates."

Talking Points

A parking/queuing bot that only answers questions is a demo, not a deployment. Jones's nine-step pipeline (gather context → read source of truth → classify → bounded tools → draft/act → check → human gate → log → feedback loop) is the bar for anything BlueAlly proposes as "enterprise AI" here — Nate B. Jones, "Fix your AI pipeline or lose your budget."
The guest journey is a loop-of-loops problem, not three separate bots. Parking status, queue length, and guest notifications are recurring processes with shared state (where is this guest, what do they need next). Jones's "loop of loops" framing — recurring jobs that notice each other and share context, stopping before consequential action (e.g., before pushing a notification or a refund) — maps directly onto this, and he explicitly recommends starting with tedious, recoverable processes before anything financial or customer-facing — Nate B. Jones, "I Stopped Prompting AI One Task At A Time."
Guest memory will fragment across vendor tools unless owned deliberately. If parking, ticketing, and in-park messaging run on different platforms, none of them will share what the guest bot learned five minutes earlier. Jones argues memory architecture, not model choice, is the primary determinant of agent quality, and that platform-native memory is a retention mechanism for the vendor, not infrastructure for the buyer — Nate B. Jones, "How to build a 10-cent AI brain." This is a direct argument for BlueAlly owning the integration/memory layer rather than letting it default to whatever ticketing or parking vendor is in place.
The nutrition/procurement system is the stronger candidate for "true" enterprise AI because it has a real source of truth (feeding logs, vet records, inventory counts) to ground against — satisfying step 2 of Jones's pipeline — and lower reputational blast radius than a guest-facing failure.
Human gates are not a weakness to apologize for — they are the design. For a guest bot, any escalation path (lost child, refund, medical concern) should have an explicit, pre-specified human handoff, decided before build, not discovered after a failure — Nate B. Jones, "Fix your AI pipeline."
Hosted vs. self-hosted is now a real, named architecture decision. Perplexity Computer shows the hosted end of the spectrum: fast setup, hundreds of pre-built connectors, credit-based metering, no credential management burden — but Berman is explicit that organizations with data sensitivity requirements should not default to it — Matthew Berman, "Perplexity Just Built an AI That Does Everything." Guest PII (parking, payment) and animal health records both argue for scrutinizing this tradeoff rather than assuming hosted is the default answer.
Coordination between systems (parking vendor, ticketing, procurement, vet records) is the actual cost center, not the AI itself. Jones's "Open Engine" argument — that staff currently act as unpaid copy-paste middleware between disconnected tools — is a direct diagnostic question to ask the aquarium: how much staff time currently goes into manually relaying information between the parking system, the queue system, and guest services? — Nate B. Jones, "I Was The Only Thing Connecting Claude, ChatGPT, and Codex."
Whichever model vendor is recommended, model choice is not neutral. Karpathy's move to Anthropic, and Anthropic's simultaneous lead on revenue, zero founder attrition, and senior talent draw, is a three-signal viability indicator worth noting if BlueAlly is recommending a foundation model for a public-facing, safety-adjacent deployment — Matthew Berman, "This is absolutely CRAZY." Anthropic's safety-first posture in particular is relevant to a bot interacting with children in a physical, safety-relevant environment (parking lots, crowd queuing).

Relevant Themes

Agent pipeline depth vs. point-solution demos (governance)
Workflow/loop coordination across disconnected systems (parking, queue, procurement)
Memory as owned infrastructure vs. vendor-locked retention feature
Human-in-the-loop design as a trust mechanism, not a shortfall
Hosted vs. self-hosted architecture tradeoffs (data sovereignty)
Vendor/model selection signal-reading (talent, revenue, worldview)

What the Experts Are Saying

Nate B. Jones — enterprise AI requires the full nine-step pipeline (context → truth source → classify → bounded tools → draft → check → human gate → log → feedback loop); anything less is a demo, and companies measuring ROI against point tools are measuring the wrong baseline ("Fix your AI pipeline or lose your budget").
Nate B. Jones — "loop of loops": recurring jobs with memory that notice each other, share state, and stop before consequential action; the coordination layer between apps, not any single app, is where enterprise value and cost currently sit ("I Stopped Prompting AI One Task At A Time").
Nate B. Jones — memory architecture, not model selection, is the primary determinant of agent capability; platform-native memory is a vendor retention mechanism, and a portable, owned memory layer avoids compounding switching costs ("How to build a 10-cent AI brain").
Nate B. Jones — a shared ticketing/queue substrate (ticket, not prompt: context, constraints, definition of done, receipt) solves the human-as-middleware coordination tax between disconnected AI tools and systems ("I Was The Only Thing Connecting Claude, ChatGPT, and Codex").
Nate B. Jones / Ethan Mollick — AI proficiency is a proxy for management proficiency (goal-setting, feedback quality, delegation); organizations should audit training spend accordingly before assuming a tool-literacy gap ("You're learning AI wrong").
Matthew Berman — hosted agent platforms (Perplexity Computer) collapse setup time via pre-built connectors but are not the answer for organizations with data sensitivity requirements ("Perplexity Just Built an AI That Does Everything").
Daniel Miessler — a mature AI harness pre-solves context (identity, goals, current state) rather than wrapping tools; an "Ideal State Artifact" replaces ad hoc specs/PRDs as a living, versioned definition of "done" ("Rethinking AI Harnesses").
Matthew Berman (Karpathy/Anthropic segment) — model vendor selection carries worldview implications (safety posture, restriction philosophy) that should be modeled explicitly for safety-adjacent deployments ("This is absolutely CRAZY").

Customer Discovery Questions

Where does staff time currently go in manually relaying information between the parking system, queue/ticketing system, and guest services desk — is there a de facto "human middleware" cost today (per Jones's Open Engine argument)?
What counts as a "source of truth" for feeding schedules, vet records, and inventory today — spreadsheets, a dedicated system, paper logs? Any agent here must ground against something real, per Jones's pipeline step 2.
What guest situations absolutely require a human handoff today (lost child, medical event, refund dispute, upset guest), and who owns that escalation path currently?
Does the aquarium have a preference or existing policy on hosted vs. self-hosted data handling for guest PII (parking, ticketing) or animal health records?
Is there an existing integration between the parking vendor, queue/ticketing platform, and any CRM or notification system, or are these currently three unconnected systems?
Who on staff would supervise/manage these bots day to day, and what is their current comfort level giving feedback on AI output versus needing to check it line by line?
For the nutrition/procurement system: is there a regulatory or compliance recordkeeping requirement (health inspection, USDA, AZA accreditation) that any automated system must satisfy and log against?
What does "done" look like for each bot in the aquarium's own terms — has this been written down anywhere, or does it live only in stakeholders' heads?

Possible Workshop / Service Opportunities