A living strategy map
The vision: enterprise AI should compound organizational value over time. Successful agentic workflows become discoverable, reliable, reusable building blocks; meet users in the surfaces where they already work; progressively redesign and automate work; and improve through continuously maintained context. Underneath: three technical pillars, seven strategic tenets, one architecture for both enterprise-facing and customer-facing agents.
Created Aug 16, 2026 · Last updated Aug 21, 2026
From the handwritten notes, Aug 2026. The four commitments in those notes live on below as tenets 1 through 3 and technical pillar 1. What remains here is the shape of the journey: every workflow climbs an autonomy ladder, and every completed climb feeds the loop that makes the next workflow cheaper to build.
What has to exist, regardless of which agents get built on top.
Identify the structured and unstructured data sources, put a layer of tools above them, and make them accessible to every agent in the organization. The data does not move to the agents; the tool layer brings governed access to the data. The knowledge base is contextualized and personalized, and freshness is a managed property: decide explicitly when knowledge should be updated and when it should decay. An agent that cannot reach the right context at the right depth is an expensive chat window.
▸ structured + unstructured sources · knowledge graph · ontology · semantic layer · ingestion (chunk, vectorize, rank) · tool layer over data · freshness + decay
How do I manage all of these agents? What does my compute look like? What does it all cost? What do safety guardrails and security look like? How does auth work for agents? Governance answers these in code, not in policy documents: agent identity, compute allocation, cost controls, guardrail enforcement, and security boundaries, managed the way a fleet is managed.
▸ agent identity + auth · sandbox · security · guardrails · compute · cost controls · quotas
For developers and users alike. The bar is not "teams can build an agent." The bar is teams composing agentic loops out of agents that already exist, and end users getting surfaces that fit the job. Compound existing value instead of rebuilding it: every new agent should make the next team's agent cheaper to build.
▸ CLI · SDK · APIs · discoverability (agent catalog) · composable agent loops · native surfaces: terminal, email, dashboard, chat, embedded, background
The direction-setting rules that outlive any single project. Pillars are what has to exist; tenets are how you decide. Tenets 1 through 4 come straight from the vision notes; 5 through 7 are what running agents in production teaches.
The platform is measured by whether value stacks: agents built on agents, tools reused across teams, each investment raising the floor for the next one.
Identify the repeatable usage patterns of each function and meet it in its native surface. Developers live in terminals and CLIs, sales and marketing live in email, analytics lives in dashboards. The agent goes to where the work already happens, not the other way around.
Redesign the workflow rather than bolting a chatbot on top of it. Ask which triggers, rhythms, and rituals currently cause humans to act, and which of those the agent can own. The test for any copilot feature: does the recruiter need a summary to act on, or can the agent email the candidate directly, at least in the early stages?
Agents should need less and less prompting to make the right call. They run the floor and pull humans in at the right moments, rather than humans running the show and pulling agents in. The directionality flips: not humans using agents, but agents using humans, with the human as the API for judgment, taste, and authority.
The system gets better from its own operation: production outcomes feed back into better agents, better prompts, better routing. Improvement is a loop the platform owns, not a quarterly project.
Eval is not static. You need a constant process for evaluating the outcomes agents generate, and it plausibly deserves 50 percent of your bandwidth. If half the effort on an agent is not going into knowing whether it works, the agent is running on hope.
For every feature, decide how much automation is right by asking what failure costs. Name the risk and its vectors, then place humans accordingly: cheap-to-undo actions run free, expensive or irreversible ones get a human gate, and the dial widens as trust is earned.
Requests enter through the gateway; agents run inside the sandbox and reach everything through the tool gateway. Unstructured sources flow through chunk → vectorize → rank back into retrieval. Every run emits into the evaluation and observability plane.
The platform underneath is the same. What changes is who is on the other side of the screen. An employee will forgive a rough edge to get their work done; a customer will not. That single difference drives every row below.
| Enterprise-facing | Customer-facing | |
|---|---|---|
| Who it serves | Employees doing their jobs: analysts, support staff, engineers, ops | Paying customers who chose you and can leave |
| Example surfaces | Internal copilots, plain-English analytics (NL→SQL), document processing, background workflow agents | Support agents, onboarding guides, in-product assistants, voice agents |
| Speed expectation | Seconds are fine for hard questions; minutes are fine for background jobs | First words on screen in under a second, always streaming |
| Reliability bar | Mostly right, with the working shown, beats slow and perfect; users can verify | A wrong answer is a broken promise; verification is your job, not the customer's |
| Polish | Functional wins; rough edges are tolerated if the work gets done | Tone, brand voice, and edge-case handling are part of the product |
| When humans step in | Review before acting on numbers; gates at irreversible steps only | Seamless escalation to a person; the customer never feels stuck with a bot |
| Cost of a failure | Rework and lost time, contained internally | Lost trust, churn, screenshots on social media |
| What you spend on | Coverage: more workflows, more agents, more compounding | Reliability and polish: evals, guardrails, latency budgets on fewer surfaces |
What the research actually says, in plain words. Sources at the bottom of the section.
19 / 20
AI pilots never show up in the company's financials (MIT, 2025)
+34%
more issues resolved per hour by beginner support agents with AI; veterans gained almost nothing
56%
faster task completion for programmers with an AI assistant, in a controlled trial
2×
bought tools succeeded twice as often as home-built ones (67% vs 33%)
1 in 5
companies have actually redesigned a workflow around AI; redesign is the strongest predictor of profit impact
~5 hrs
of expert work is what a top agent can now finish half the time, doubling every few months
This is the analytics automation target: instead of filing a ticket with the data team, anyone asks "how did signups do last quarter in Europe?" and gets an answer. The catch: in classroom demos models score around 90 percent, but on real company warehouses (thousands of columns, messy names) they initially scored 10 to 17 percent. Every system that works in production closed that gap the same way:
51% → 90%
Snowflake's accuracy jump, achieved by giving the model a business dictionary, not a better model
40% → 90%
Pinterest's rate of finding the right table, achieved by documenting tables better
10 → 3 min
time to write a query at Uber, across 1.2M queries a month
MIT GenAI Divide 2025 · Brynjolfsson/Li/Raymond QJE 2025 · Peng et al. 2023 · Anthropic Economic Index 2026 · OpenAI GDPval (arXiv 2510.04374) · METR time horizons · McKinsey State of AI Nov 2025 · Stanford HAI AI Index 2026 · Spider 2.0 (arXiv 2411.07763) · bird-bench.github.io · Snowflake Cortex Analyst · Uber QueryGPT · Pinterest Eng · LinkedIn SQL Bot · Databricks Genie
What the major players publish as their enterprise AI frameworks, read against this page.
| Who | Their frame | What they add beyond this page |
|---|---|---|
| Databricks | Agent Bricks (Apr 2026) | An AI Gateway doing model routing, fallback, and cost optimization in one choke point; context grounded in catalog metadata, lineage, and data quality (claimed 70 percent accuracy lift over plain RAG); agents inheriting the invoking user's permissions; Lakebase for agent memory; the CLEARS eval rubric. |
| Gemini Enterprise + A2A | An agent gallery users can browse; a no-code builder tier; A2A as a vendor-neutral agent-to-agent standard (agent cards, task lifecycle, signed trust), now at the Linux Foundation, adopted by Azure and Bedrock. | |
| Microsoft | Frontier Firm + Entra Agent ID | Agents as first-class directory identities with a full lifecycle: inventory, classification, human accountability, conditional access, shadow-agent discovery. And the org finding: organizational readiness drives twice the impact of individual capability. |
| AWS | Bedrock AgentCore (Jul 2025) | Memory as its own service (session plus long-term semantic); a credential token vault for agent identity; browser and code-interpreter as separately governed sandboxes. |
| Anthropic | Building Effective Agents | Simplicity-first workflow-vs-agent taxonomy; tool design treated as a craft; guardrails that graduate with autonomy: spend caps first, then auto-review, sandboxing, per-workflow cost caps. |
| OpenAI | AI in the Enterprise + Agents guide | "Start with evals" as lesson one (agrees with tenet 3); a named guardrail taxonomy with per-tool risk ratings; explicit human-intervention triggers: failure thresholds and irreversible actions (absorbed into tenet 4). |
| McKinsey · a16z | Agentic mesh + CIO survey | Governed autonomy with defined autonomy levels; org design (AI council, lighthouse transformations); the procurement reality: 37 percent of CIOs run 5+ models, and price and security now outweigh accuracy in model selection. |
Compliance is not a document written after the fact; it is evidence the platform generates as a side effect of running. The two external anchors: the EU AI Act, whose high-risk obligations began applying in August 2026 (full phase-in through 2027), and NIST's AI Risk Management Framework (Govern, Map, Measure, Manage), the de facto US anchor. Both converge on the same demands, and the platform already produces most of the answers.
| What regulators ask for | Where the platform answers it |
|---|---|
| Risk classification of every AI system | Risk tier assigned at agent creation, in the registry, driving which gates apply (pillar 2 + tenet 4) |
| Human oversight proportionate to risk | Tenet 4: gates placed by cost of failure and risk vectors; escalation paths designed, not improvised |
| Record-keeping and traceability | Traces: every run documented in the evaluation and observability plane; the audit trail is a by-product, not a project |
| Accuracy and robustness evidence | Continuous evals (tenet 3): outcome-based, versioned, running before and after every change |
| Transparency and disclosure | Customer-facing UX: users know when they are talking to an agent and when a human takes over |
| Data governance | Pillar 1's semantic layer plus PII scanning and access inheritance at the gateway; agents never see more than their invoking user |
| An accountable owner | Every agent carries a named human owner in the registry, and one accountable body (platform + legal + security) owns the mapping above |
EU AI Act timeline: artificialintelligenceact.eu · NIST AI RMF: nist.gov/itl/ai-risk-management-framework
The questions to ask of every candidate workflow, from the same notes.