Skip to content

Agentic AI Access: BNI's Thesis and Open Questions

[HYPOTHESIS] Strategic positioning doc. Articulates the through-line that ties together BNI Foundation's three AI-touching projects (MPowerUP Guardian AI, rlivn, Second Boot Module 04). Open for revision as Phase 1 pilots produce data. Last updated 2026-05-19.

The thesis in one sentence

BNI Foundation believes that agentic AI is becoming a category of literacy, agency, and economic participation as important as internet access was in 2000 — and that without intervention, access will follow money, leaving the underserved without the most powerful form of the technology.

Bringing agentic AI to vulnerable populations is therefore a primary mission goal, on equal footing with the existing BNI Foundation commitments around open-source software, accessible hardware, and privacy-first architecture.

What "agentic AI" means here

This thesis is specifically about agentic AI — not chat AI in general. The distinction is load-bearing:

Capability Examples (2026) Access for the underserved (2026)
Chat AI ChatGPT free tier, Claude free tier, local Ollama, Phi-3, Llama 3 Broadly accessible. The gap here is mostly literacy + hardware, not paywalled access.
Agentic AI Claude with tool use, ChatGPT Operator, Anthropic Computer Use, Cursor, GitHub Copilot, autonomous browsing/booking/researching/coding assistants Mostly paywalled. $20–200/month subscriptions, per-API-call pricing, professional-tier tools.

Chat AI helps a person think. Agentic AI acts on a person's behalf — applies for jobs, drafts and sends messages, books appointments, navigates benefits portals, writes code, researches options, files paperwork.

For someone with stable employment and disposable income, agentic AI is a productivity multiplier. For someone in recovery navigating reentry paperwork, applying for disability benefits, fighting an eviction notice, coordinating senior caregiving, or running a small program with one volunteer — agentic AI is the difference between systems being accessible to them or not.

The state of the gap (2026)

[EMPIRICALLY VALIDATED] Cost barriers for agentic AI tools are publicly documented and consistent across providers:

  • Claude Pro: $20/mo individual; Claude Code requires a Pro or higher tier; API tool-use billed per million tokens
  • ChatGPT Plus: $20/mo; Pro (incl. Operator): $200/mo
  • Cursor: $20/mo Pro, $40/mo Business
  • GitHub Copilot: $10/mo individual; $19/mo business

[HYPOTHESIS] These prices are well-targeted at knowledge workers with disposable income; less so at populations BNI Foundation serves. Whether the gap is growing is also [HYPOTHESIS] — it depends on whether free/open-source agentic AI catches up faster than paid tools advance. Worth tracking, not yet evidence-based.

[EMPIRICALLY VALIDATED] Local / open-source agentic AI exists but lags significantly behind paid tools as of 2026:

  • Models that fit on consumer hardware (Llama 3.2 3B, Phi-3 Mini, etc.) handle chat acceptably but struggle with the multi-step reasoning, tool use, and instruction-following required for reliable agentic workflows.
  • Larger open-weight models (Llama 3.3 70B, Mistral Large) approach paid-tool quality but require hardware most BNI participants don't own.
  • Open-source agentic frameworks (LangChain, LangGraph, AutoGPT, smolagents) exist but require developer-level integration work.

Bottom line: the affordability of chat AI has temporarily masked the inaccessibility of agentic AI. A working laptop with Ollama can converse with a participant; it cannot reliably navigate the West Virginia DHHR benefits portal on their behalf, draft and submit a cover letter to a specific job, or remind them about a medication and call their caregiver if they miss it.

How this thesis maps to BNI projects

Three BNI projects already operate inside this thesis, whether or not it was named:

Project Agentic AI role Population Status
MPowerUP Guardian AI (Phase 4.5) Plain-language translator, scam detector, service navigator, MPWR token advocate — acting on user's behalf inside a mutual-aid network Houseless, recovery, reentry Phase 4.5 planned; gated on 50+ active users and UX-integration design
rlivn Persistent warm companion managing reminders, medications, smart-home actions, family calling, caregiver escalation Elderly, dementia/Alzheimer's, neurodivergent Phase 1 scaffolding pending
Second Boot — Module 04 (AI Literacy) Curriculum module teaching cloud + local AI tools, including hands-on agentic workflows for job search, document drafting, learning Schools, recovery, reentry, seniors, incarcerated, all program audiences Curriculum drafted; first cohort pending

These three projects share more than a passing similarity — they're three different insertion points for the same thesis: agentic AI access for populations the market doesn't price for.

Where the thesis is on solid ground

  • [EMPIRICALLY VALIDATED] There is a real, present, paywalled agentic-AI access barrier. Subscription prices and capability comparisons are public.
  • [EMPIRICALLY VALIDATED] BNI's target populations (houseless, in recovery, reentry, elderly, economically marginalized) consistently have less disposable income than the population paid tools are priced for. This is not a controversial claim about poverty.
  • [EXPERT REVIEWED] The technical claim that small local models lag significantly behind frontier paid models at agentic workflows is well-documented in benchmarks (SWE-bench, GAIA, AgentBench, etc.) as of late 2025/early 2026.

Where the thesis has real gaps to close

[HYPOTHESIS] markers below. These are the four problems BNI signs up to engage with if this thesis is to mean anything operationally:

Gap 1 — Specificity: agentic vs. chat

External materials should distinguish agentic AI access from AI access in general. Conflating them weakens the thesis and exposes BNI to fair pushback (chat AI access is broadly available). The gap claim is most defensible when narrowly stated about agentic capability.

Commitment: Mission, grant, and partner-facing copy uses "agentic AI" precisely. We don't claim there's a gap in chat AI access where there isn't.

Gap 2 — Outcomes: "empower" requires evidence

[HYPOTHESIS] Bringing agentic AI to vulnerable populations improves their outcomes — employment, benefits access, recovery sustainment, health outcomes, social connection. This is plausible and motivates the work. It is not yet validated for the specific populations BNI serves with the specific tools BNI is building.

Commitment: External materials describe agentic AI tools as [HYPOTHESIS] interventions until pilot data is available. We say "we believe X improves Y" rather than "X improves Y." This applies to grant applications, partner communications, and user-facing copy.

Gap 3 — Safeguards: vulnerable users + agentic action = new research category

Agentic AI can take real actions in the world: send messages, spend money, book appointments, fill forms, share personal data. For users with full context, support networks, and high digital literacy, those actions have recoverable failure modes. For users in recovery, in reentry, with cognitive impairment, or with limited support — agentic failure modes are not symmetric.

A Guardian AI that misreads a help request and broadcasts a recovery-context disclosure to a wider Circle than intended causes a different category of harm than ChatGPT making up a fact. The MPowerUP project red-team analysis already touches the surface of this. The honest statement is: the field has not figured out vulnerable-population-specific agentic AI safeguards. BNI is signing up to invent some of them.

Commitment: Every agentic feature ships with explicit handling for failure modes specific to its target population. No "minimum viable agent" deployment to vulnerable users without an agentic-safety review documented per project. Phase 4.5 of MPowerUP is gated on this; rlivn's clinical / consent gating is gated on this; Second Boot Module 04's hands-on AI use is supervised in cohort and includes explicit failure-mode discussion.

Gap 4 — Economics: who pays for the agent calls?

This is the unsolved question that affects every concrete implementation. Tools that approach paid-tool capability require cloud inference at non-trivial per-call cost. The realistic stack for a BNI-served user is:

  • Hardware: Second Boot awarded laptop, 8–16GB RAM
  • Local AI: chat-quality (Phi-3 Mini or Llama 3.2 3B); insufficient for reliable agentic workflows
  • Cloud agent budget: required for capability — but the participant cannot afford it

Possible answers, each with its own [HYPOTHESIS]:

  • a) BNI Foundation pays. Grant-funded agent-call budget pool, drawn on by program participants. Sustainable if grants cover it. Not validated.
  • b) Sponsor pays. A corporate sponsor (Microsoft Philanthropies, Anthropic credits, Google.org) provides per-participant agentic-AI budget. Plausible; not arranged.
  • c) MPWR token economy pays. Inside MPowerUP, the MPWR token covers the participant's agent costs as part of the redeemable basket. Requires Phase 3.5+ to even exist.
  • d) Local-only. Accept the capability gap; ship the best local agent we can. Honest but means the gap remains — we're closing the access gap but not the capability gap.
  • e) Hybrid. Local for everyday use; cloud agent budget reserved for high-leverage actions (job applications, benefits navigation). Likely operationally correct.

Commitment: Each BNI agentic deployment publishes its agent-cost model up front — who pays, how much, what happens when the budget runs out, what the participant experiences.

What BNI is committed to validating

Concrete validation milestones, by project:

  • MPowerUP Guardian AI: Pilot with 10+ users; measured outcomes for at least one of: scam-detection-correct-rate, services-found-and-used count, benefits-application-completed count. [HYPOTHESIS] → [PILOT VALIDATED]. Requires Phase 3 hardening completed first.
  • rlivn: Pilot with 3+ client-caregiver pairs; measured outcomes for at least one of: medication-confirmation rate, caregiver-escalation-appropriateness rate, client-reported warmth/satisfaction. [HYPOTHESIS] → [PILOT VALIDATED]. Requires HIPAA / consent path validated.
  • Second Boot Module 04: First cohort with measured outcomes for: number of participants who complete one real agentic task during the module (cover letter to specific job, benefits-portal exploration, etc.), self-reported confidence change. [HYPOTHESIS] → [PILOT VALIDATED]. Requires first program partner committed (Phase 0 exit criterion).

Known unknowns

In keeping with the Epistemic Honesty directive:

  • Whether agentic AI access — at the quality level currently provided by paid tools — materially changes outcomes for the populations BNI serves. Plausible, not validated.
  • Whether closing the agentic AI access gap is the right Foundation priority versus the existing focus on safety + privacy + hardware infrastructure. Could be either complementary or competitive for limited team attention.
  • Whether the agentic-safeguards research BNI signs up for here is achievable by a 2-person team, or whether it requires recruiting domain experts (clinicians, social workers, recovery practitioners, formerly incarcerated technologists) into the design loop.
  • Whether the open-source agentic-AI ecosystem catches up to paid tools faster than paid tools advance — which would close the gap without intervention.
  • Whether "agentic AI for vulnerable populations" is a fundable thesis in the current grant environment. Mozilla Foundation, Ford Foundation, and similar mission-aligned funders might support; uncertain how to position for non-AI-specific funders.
  • Whether the populations BNI serves want agentic AI agency, or whether the framing imposes a tech-solutionist lens on what are fundamentally human-services problems. The honest answer is: ask them, in pilots.

What this thesis is NOT

  • Not an argument that AI fixes the underlying problems. AI is a tool; the problems (poverty, addiction, mass incarceration, elder isolation) have human causes that humans must address.
  • Not an argument that the underserved are less capable of using powerful AI tools. They are equally capable; they are currently denied access by price and adjacent barriers.
  • Not a commitment to deploy agentic AI to vulnerable populations regardless of safety. The Gap 3 / safeguards work has to happen first per project.
  • Not an exclusive Foundation mission goal. The existing commitments (open-source software, hardware frameworks, privacy-first architecture, accessibility, communities-as-stakeholders) remain. This thesis names a through-line that already runs across BNI's AI work; it doesn't displace the rest.

Decision log

Date Decision Status
2026-05-19 Thesis articulated and published [OPEN] for revision after first pilot data
TBD First validated outcome from any of the three project pilots —
TBD Agent-cost model selected for first deployment —
TBD Agentic-safeguards review template ratified —