Where the next useful enterprise AI systems might be built.
A map of industries, problems, technologies and commercial wedges—viewed through what Kubrick can credibly sell and what I am unusually positioned to build.
Executive Summary
The clearest opportunity is to extend Kubrick’s AI Business Partner model with an Agent Platform & Assurance layer. Kubrick already knows how to find workflows, ship lightweight tools and support adoption. The gap is the production machinery behind them: governed context, tool permissions, evaluation, evidence and platform engineering.
Lead with a Federated Agent Data Plane, not a generic “agent” proposition. Research indicates that organizational context is a binding constraint on sophisticated enterprise AI. Starburst and Databricks create a credible route from fragmented data to evaluated workflow; either platform can lead when discovery says it should.
This can solve placements because Kubrick has already demonstrated the commercial mechanism. Shell scaled to 165 consultants and retained more than 40%; AI Business Partners delivered 12 tools. A senior-led platform pod can create repeatable roles for data, platform, AI, governance and product consultants.
The market is asking for production discipline, not more demos.
Four findings change how the opportunity should be framed. They come from different samples and should not be combined into one forecast; together they identify the recurring design constraints.
AI is widespread; production agents are not.
The Stanford AI Index reports AI use at 88% of surveyed organizations in 2025, while agent deployment remained in the single digits across nearly every business function.
Implication: The scarce capability is no longer access to a model. It is turning a bounded workflow into a controlled production system.
Stanford HAI · 2026 AI Index ↗The more an agent does, the more organizational context it needs.
Anthropic found 77% of sampled enterprise API use followed automation patterns, but sophisticated tasks required longer inputs and were constrained by dispersed or tacit information.
Implication: Federation, data products, metadata and identity are not plumbing around the agent; they are part of the product.
Anthropic Economic Index · Sep 2025 ↗AI can improve adjacent tasks and still make the whole decision worse.
In a preregistered experiment with 758 consultants, AI users completed 12.2% more in-frontier tasks 25.1% faster, but were 19% less likely to be correct on an outside-frontier task.
Implication: A serious offer needs task boundaries, evaluation sets, evidence, human authority and a safe failure path—not just prompts and tools.
Organization Science · Mar 2026 ↗AI can compress the experience curve—but it does not remove the need for judgment.
A study of 5,179 support agents found a 14% average productivity gain and 34% for novice and lower-skilled workers. A 2026 Anthropic survey also found greater displacement concern among early-career respondents.
Implication: Kubrick can sell AI-enabled squads that transfer senior patterns to junior consultants while preserving review, learning and accountable ownership.
NBER · Generative AI at Work ↗Start from a frequent, consequential decision with an owner and measurable baseline.
Prove the minimum governed information and permissions the system needs.
Define correct, safe and useful with domain reviewers before production.
Sell a bounded outcome, then expose the repeatable roles needed to expand it.
The strongest proposition uses the overlap, not one résumé line.
The research points to a systems problem. That is useful because the combined experience in federation, lakehouse, cloud, governance and regulated delivery is more defensible than another general AI offer.
A stack for useful agents—not a logo collage.
The architecture begins at the workflow and works down. Starburst and Databricks overlap in federation; discovery should choose the simplest defensible access pattern rather than forcing both products into every sale.
When live access is the bottleneck
Many heterogeneous sources, limited appetite for movement, existing SEP estate, cross-domain consumers, read-only agent tools.
When building and learning is the bottleneck
Heavy transformation, retrieval, ML features, evaluation, monitoring, model serving and a lakehouse-centered workflow.
When the agent must cross the estate
Starburst exposes curated live data; Databricks builds and evaluates the application. Identity, cost, lineage and failure behavior require explicit design.
Four propositions deserve disproportionate attention.
The previous version gave twelve ideas almost equal visual weight. The research supports a narrower portfolio: two offers to package now, one banking workflow to prove and one energy proposition to develop with a sponsor.
Federated Agent Data Plane
Give one production agent governed, read-only access to the minimum data products it needs—without waiting for wholesale migration.
- Why it survives the research
- It directly answers the context bottleneck identified in enterprise AI research and uses the most distinctive overlap of Starburst, Databricks, governance and current delivery experience.
- First commercial wedge
- A 2–3 week Agent Data Access Lab: map one workflow, identify the minimum governed context, then prove a read-only path across three sources.
- Roles it can create
- Solution architect · Starburst engineer · Databricks/agent engineer · governance analyst · product analyst
- Kill it if…
- No sponsor owns a cross-system workflow, or the necessary data can already be served safely from one platform.
AI Evidence & Assurance Factory
Create a reusable evaluation-and-evidence pipeline that moves one high-impact AI workflow to a defensible go, remediate or stop decision.
- Why it survives the research
- The consulting evidence shows AI quality is task-dependent, while Databricks reports evaluation and governance are strongly associated with production deployment.
- First commercial wedge
- A 4-week Production Readiness Gate: instrument one agent, build an evaluation set, define release thresholds and automate its evidence pack.
- Roles it can create
- Assurance architect · evaluation engineer · ML engineer · domain reviewer · governance specialist
- Kill it if…
- The client has no agent approaching production or model-risk ownership is too fragmented to make a release decision.
Regulatory Change Impact Engine
Turn one regulatory change into traceable obligations, impacted controls, accountable owners, remediation actions and an audit trail.
- Why it survives the research
- It turns a familiar regulated-enterprise pain into a workflow with a buyer, measurable cycle time and an auditable human decision rather than another document assistant.
- First commercial wedge
- A 6-week change-to-control proof for one rule, one control domain and one source of evidence, with human approval at every consequence.
- Roles it can create
- Domain analyst · data engineer · graph/knowledge engineer · agent engineer · control owner
- Kill it if…
- The client cannot provide a recent rule change with known remediation cost, delay or control-coverage pain.
Operational Resilience Control Room
Fuse operational signals into prioritized cases, recommended actions and safe escalation paths with explicit human authority.
- Why it survives the research
- Kubrick has unusually strong energy delivery proof and this creates durable engineering demand, but operational safety and domain depth make it a partner-led second move.
- First commercial wedge
- A shadow-mode resilience pilot for one incident class: unify signals, recommend actions, record evidence and compare decisions with operators.
- Roles it can create
- Operational SME · streaming engineer · ML engineer · platform engineer · product lead · reliability engineer
- Kill it if…
- No operational sponsor will permit shadow testing against historical incidents or expose the required event and work-order data.
Move outward from credibility, not from industry enthusiasm.
This is a route-to-market sequence rather than a market-size ranking. Technical transferability matters, but references, domain ownership and procurement determine what is actually buildable.
Current credibility, known control environment and strong data-platform pain make this the lowest-friction place to earn a reference.
Best routeAgent data access · assurance · regulatory change
GateAvoid client-confidential reuse; sell a pattern, not Wells Fargo specifics.
The context, evaluation and control problems recur across industries and align with Kubrick’s partner portfolio.
Best routeAgent platform pod · data-product onboarding · evaluation factory
GateNeeds an executive owner and a live workflow, not an abstract architecture programme.
Kubrick’s Shell and JLR work demonstrates a route from senior-led delivery to large consultant cohorts and retained capability.
Best routeOperational resilience · asset decisions · governed data products
GateRequires operational SMEs and a higher reliability bar.
Investigation, claims and assurance patterns transfer, but domain and ethical risk are materially higher.
Best routeEvidence workbench · review agent · model assurance
GateStart only with human-reviewed scope and a credible domain partner.
Use cases are visible and repeatable, but procurement access matters more than technical fit.
Best routeAcquisition review · compliance evidence · cross-agency data
GateDo not invest ahead of a vehicle, prime partner or named agency sponsor.
Keep the breadth, but stop pretending every idea is equal.
Use this as a research backlog. “Lead” ideas have the strongest current evidence and right to win; “Develop” ideas need a sponsor or proof point; “Watch” ideas should not consume proposal effort yet.
Federated Agent Data Plane
Give one production agent governed, read-only access to the minimum data products it needs—without waiting for wholesale migration.
Evidence, buyer and first build
- Problem
- An enterprise wants agents, but the required data lives across warehouses, operational databases, legacy platforms and domain clusters with different permissions.
- Buyer
- Chief Data Officer · AI platform lead · domain CIO
- First build
- A six-week reference implementation for one agent, three governed sources and a measurable business workflow.
- Kubrick advantage
- Preferred Starburst implementation partner, Gold Databricks partner and cross-platform delivery model.
- My advantage
- Direct experience with federation, multi-domain SEP, GKE, access control, Databricks pipelines and production delivery seams.
- Placement engine
- 2-person architecture sprint → 5-person platform pod → domain teams for onboarding data products and agents.
- Hard truth
- The winning design may use Starburst, Databricks federation, or both. The proposition must be vendor-neutral enough to survive architecture discovery.
Regulatory Change Impact Engine
Turn one regulatory change into traceable obligations, impacted controls, accountable owners, remediation actions and an audit trail.
Evidence, buyer and first build
- Problem
- New rules arrive as documents, while policies, controls, systems, owners and evidence live in different tools and organizational silos.
- Buyer
- Chief Compliance Officer · operational risk · internal audit
- First build
- One regulation, one control domain, one evidence repository and human approval at every consequential step.
- Kubrick advantage
- Financial-services access plus Databricks, Neo4j and Collibra relationships can form a credible multi-vendor delivery pattern.
- My advantage
- Experience translating ambiguous requirements into infrastructure, ownership and implementation decisions in a regulated bank.
- Placement engine
- Compliance discovery team → graph/data engineering pod → embedded control-domain analysts and engineers.
- Hard truth
- It becomes boring document automation unless the offer owns the full change-to-control workflow and measures cycle time and control coverage.
AI Evidence & Assurance Factory
Create a reusable evaluation-and-evidence pipeline that moves one high-impact AI workflow to a defensible go, remediate or stop decision.
Evidence, buyer and first build
- Problem
- AI teams can prototype quickly but cannot produce consistent evidence for risk review, approval and production monitoring.
- Buyer
- Chief AI Officer · model risk · AI governance · product owner
- First build
- Instrument one agent, define domain-specific scorers, build an evaluation set and automate the release evidence pack.
- Kubrick advantage
- Kubrick already sells AI productionization and governance; the opportunity is a concrete factory with reusable tests and evidence packs.
- My advantage
- Strong fit with platform architecture, governance, delivery controls and the desire to build non-gimmicky agent systems.
- Placement engine
- Assurance architect + domain reviewer → evaluation engineering pod → recurring model/agent onboarding team.
- Hard truth
- Avoid generic responsible-AI consulting. The unit of value is a production release decision, not a framework presentation.
Connected Investigation Workbench
Build an evidence-first investigation workspace where agents assemble connections and recommend the next inquiry while humans retain the decision.
Evidence, buyer and first build
- Problem
- Fraud investigators manually connect transactions, identities, claims, cases, communications and documents across systems.
- Buyer
- Fraud operations · financial crime · program integrity
- First build
- One investigation type, a bounded entity model, read-only evidence access and a measured reduction in search time.
- Kubrick advantage
- The partner portfolio spans federation, lakehouse, graph and governance—unusually complete for an investigation system.
- My advantage
- Combines bank context, cross-source access, architecture and agent orchestration rather than reducing the work to a single model.
- Placement engine
- Data access pod + graph/ML pod + investigators-in-the-loop product team; strong multi-discipline placement potential.
- Hard truth
- High-value but crowded. Differentiation must come from governed live access and an auditable investigation trail.
Operational Resilience Control Room
Fuse operational signals into prioritized cases, recommended actions and safe escalation paths with explicit human authority.
Evidence, buyer and first build
- Problem
- Operators receive alarms, work orders, sensor signals, manuals, weather and incident reports through separate systems and respond under time pressure.
- Buyer
- Chief Operating Officer · resilience · asset operations
- First build
- One incident class, one operating region and a shadow-mode agent that recommends but never executes control actions.
- Kubrick advantage
- Energy access and proven aviation maintenance/decision-intelligence work provide a route to a real operational workflow.
- My advantage
- A stretch beyond current banking depth, but strong alignment with distributed systems, resilience, DR and complex dependencies.
- Placement engine
- Operational discovery → streaming/ML platform pod → long-lived regional rollout and reliability team.
- Hard truth
- The safety bar is much higher than office automation. Reliability, graceful degradation and operator trust are core product requirements.
Third-Party Risk Network
Create a living supplier graph that identifies changed obligations, hidden concentration and the cases requiring human review.
Evidence, buyer and first build
- Problem
- Contracts, obligations, supplier hierarchies, incidents, spend, performance and external risk signals are reviewed separately.
- Buyer
- Chief Procurement Officer · third-party risk · supply-chain resilience
- First build
- A single critical supplier category, contract obligations and two external/internal risk sources.
- Kubrick advantage
- Cross-sector reach and data/graph/document capabilities support reuse from banks to energy and government.
- My advantage
- Strong architecture fit, although the business domain would require deliberate learning to avoid a generic risk dashboard.
- Placement engine
- Procurement SME + data pod → graph and workflow build → category-by-category rollout.
- Hard truth
- Potentially broad but commercially vague. It needs a sharply defined event such as onboarding, annual review or incident response.
Federal Acquisition Intelligence
Build a governed acquisition workspace that classifies files, checks scope, drafts requirements and records human approvals.
Evidence, buyer and first build
- Problem
- Contracting teams review large document sets, classify requirements, draft solicitations and preserve compliance evidence manually.
- Buyer
- Agency procurement leader · contracting operations · CIO
- First build
- Mirror a published GSA use case with a narrow document class and measurable review-time target.
- Kubrick advantage
- Kubrick has public-sector delivery evidence, but its published procurement route is UK-centric; a US partner or vehicle is the gating asset.
- My advantage
- Good system-design fit and intellectually broad, but little direct federal procurement experience today.
- Placement engine
- Potentially large once procured, but slow route-to-market and eligibility constraints make it a partner-led bet.
- Hard truth
- Excellent inspiration, weak immediate wedge until US procurement access is proven.
Healthcare Waste Review Agent
Assist human reviewers by assembling policy-grounded evidence, network signals and a transparent recommendation for selected services.
Evidence, buyer and first build
- Problem
- Claims and prior-authorization reviews require policy interpretation, provider history, clinical evidence and fraud signals at scale.
- Buyer
- Payer program integrity · claims operations · public health agency
- First build
- One service category, shadow review only, explicit reason codes and reviewer agreement measurement.
- Kubrick advantage
- Data, AI and public/health capability can translate, but healthcare domain credibility and procurement access must be validated.
- My advantage
- Technically attractive investigation problem; larger domain-learning and ethical burden than the banking-adjacent options.
- Placement engine
- Large analytics and review-operations teams if validated; also unusually high regulatory and reputational risk.
- Hard truth
- A meaningful problem, not an easy first proposition. Start only with a domain partner and narrow human-reviewed scope.
Decision-Grade Operational Twin
Create a bounded operational twin that lets planners test disruptions, resource choices and maintenance actions before committing them.
Evidence, buyer and first build
- Problem
- Asset, production and supply-chain decisions rely on disconnected telemetry, plans, maintenance history and uncertain future scenarios.
- Buyer
- Operations strategy · manufacturing engineering · supply-chain resilience
- First build
- One asset family or constrained network, with scenario accuracy compared against historical disruptions.
- Kubrick advantage
- Adaptive Intelligence, aviation maintenance and supply-chain optimization are real proof points rather than a cold start.
- My advantage
- High novelty and systems depth, but farther from current expertise and less dependent on Starburst.
- Placement engine
- High engineering intensity and durable team demand, but longer discovery and domain-modeling cycles.
- Hard truth
- Avoid a visually impressive twin with no decision authority. Start from a recurring planning choice and work backward.
Cross-Domain Data Product Exchange
Package the architecture, controls and operating model for publishing and consuming governed data products across domain boundaries.
Evidence, buyer and first build
- Problem
- Domain platforms are isolated for ownership and resilience, but users and agents still need governed cross-domain consumption.
- Buyer
- Enterprise data platform · data-mesh lead · domain CIO
- First build
- Two domains, one published product, one consuming workload and explicit ownership, SLA, access and cost rules.
- Kubrick advantage
- Starburst’s preferred implementation partner can turn new product capability into a repeatable adoption offer.
- My advantage
- This is the closest adjacency to current multi-domain, Stargate, data-product, chargeback and governance work.
- Placement engine
- Architecture sprint → enablement pod → domain-by-domain product onboarding factory.
- Hard truth
- Very buildable, but it risks feeling like platform plumbing unless tied to a visible consumer such as an agent or risk workflow.
Federated Data Cost & Workload Router
Recommend where a workload should run and whether data should stay, federate, cache or move—using observed query and cloud-cost evidence.
Evidence, buyer and first build
- Problem
- Enterprises run overlapping query engines and move data without a clear view of workload fit, source cost, latency or domain ownership.
- Buyer
- Data platform leader · FinOps · cloud architecture
- First build
- Classify one month of workloads across two engines and validate recommendations with platform owners before automation.
- Kubrick advantage
- Multi-platform partnerships create permission to be a practical optimizer rather than a single-vendor advocate.
- My advantage
- Strong match with federation, chargeback, platform architecture and a desire to understand the economics of technical decisions.
- Placement engine
- Small diagnostic sprint can expand into platform engineering, FinOps and migration/remediation teams.
- Hard truth
- Commercially compelling if savings are measurable; technically difficult because comparable cost and performance data are messy.
Enterprise Agent Control Plane
Create a platform pattern for registering agents, controlling tools, evaluating releases, observing production behavior and escalating failures.
Evidence, buyer and first build
- Problem
- Teams are building agents in different platforms with inconsistent tool permissions, release tests, ownership, monitoring and kill switches.
- Buyer
- Chief AI Officer · enterprise architecture · platform engineering
- First build
- Bring two existing agents under one release, permissions and monitoring pattern instead of designing an abstract enterprise platform.
- Kubrick advantage
- Partner breadth across Databricks, Snowflake, Microsoft, Collibra and Starburst is an advantage if the design remains interoperable.
- My advantage
- Builds directly on platform, identity, multi-domain and agent-system interests while stretching into product architecture.
- Placement engine
- Central platform pod plus federated enablement teams; strong long-term demand if Kubrick lands the architecture role.
- Hard truth
- A large, competitive category. The wedge should be productionizing the first two agents, not selling a grand control-plane vision.
This is not a new commercial model. It is a sharper front door.
Kubrick’s own cases show that a bounded delivery outcome can expand into a large talent pipeline. The proposition should make that mechanism easier to start and more relevant to the agent era.
AI Business Partner, architect and domain owner define the decision, evidence, access and baseline.
One live workflow, minimum governed context, release evaluation and explicit human authority.
Reusable data products, tools, evaluation, observability and adoption around the first workflow.
Junior-heavy squads onboard workflows and data products while the client retains capability.
Avoid the boring middle and the empty frontier.
The work needs enough platform depth to stay interesting, enough business ownership to matter and enough control to be trusted.
Delivery without a proposition
- Generic cloud migration
- Dashboard backlogs
- Undifferentiated staff augmentation
- Chatbot over a document folder
- Governance decks with no runtime
Governed systems that move work
- Multiple systems and decision owners
- Agents with bounded tools
- Platform + data + workflow integration
- Human authority and measurable outcomes
- Reusable architecture and team shape
Interesting, but needs a partner
- Autonomous high-stakes decisions
- Robotics and control hardware
- Foundation-model development
- Clinical decision ownership
- Federal market without a vehicle
Signals worth following.
The recommendations are inferences, not market facts. These papers, reports, product documents and Kubrick cases establish the evidence underneath them and show where vendor claims require validation.
Adoption is high; agents are early
AI reached 88% of surveyed organizations in 2025, but agent deployment stayed in the single digits across nearly every business function.
Stanford AI Index 2026 ↗Context is an enterprise bottleneck
Anthropic’s enterprise API analysis found sophisticated tasks require more context and can stall where information is dispersed, tacit or inaccessible.
Anthropic Economic Index ↗The AI frontier is jagged
Consultants improved speed and output inside the frontier but became less likely to reach the correct answer on an outside-frontier task.
Organization Science 2026 ↗AI can accelerate novice performance
An NBER field study found a 14% average productivity gain and a 34% gain for novice and lower-skilled support agents.
NBER · Generative AI at Work ↗Starburst’s agentic layer is now tangible
The 480-e LTS release added MCP search and retrieval for data products; data-product sharing is GA and Data Products as Code is in preview.
Starburst 480-e release ↗Databricks now spans the agent lifecycle
MLflow 3 connects tracing, evaluation, human feedback and production monitoring; Unity Catalog governs data and AI assets.
Databricks MLflow 3 ↗Kubrick has already proven the scale mechanism
Shell scaled to 165 consultants with 42% reported cost savings and retained more than 40% as employees.
Kubrick · Shell case study ↗Kubrick has already proven workflow-level AI
AI Business Partners delivered 12 agents and tools, including reported reductions in manual effort and measurable quality improvements.
Kubrick · AI Business Partners ↗Extend Kubrick’s AI Business Partner model with an Agent Platform & Assurance offer. Lead with a Federated Agent Data Plane, prove it through one consequential banking workflow, then expand into evaluation, platform and domain squads.