Kodak was, in effect, the first social network. People took pictures in order to share them, and the print was an early form of the social media post. Kodak did not lose its market to the digital sensor or the flat-panel display. It lost it to the cloud, which made sharing instant and free. Yet the cloud never appeared on Kodak's horizon, because the company defined its business by the steps it dominated, taking and printing pictures, and not by the outcome its customers valued: sharing them.
AI transformation exposes today's incumbents to the same risk. In this report, we map the horizons of five companies in industries where AI will reshape how value is delivered. Each diagram places at its center the lifecycle of the outcome customers rely on the company for, because customers do not buy products; they buy outcomes. On the surrounding ring sit the AI building blocks that will act on each step of that lifecycle. Each block on the horizon is a decision the company has not yet made: is it core to the business, is it a threat, or is it not relevant? And is now the right time to act? Regulations sit on the horizon too: a rule can make a block core, turn it into a threat, or set the date by which the choice must be made. For Kodak, the question was whether digital sensors were good enough to capture the pictures people wanted to take, share and keep. The block that matters most is the one positioned to take over the final outcome. That is where the incumbent's position will be decided.
The final outcome will go to whoever the customer trusts to deliver it. A customer delegates an outcome only to an agent it trusts: to move its money, settle its claim, restock its home or choose its therapy. The contest for the final step is therefore a contest for trust, and trust at scale rests on infrastructure that can be shown to customers, auditors and regulators.
Can I trust the outcome? The agent does the right thing, for the right amount, exactly once, in every case the business allows.
Is it safe to hand it the task? No one can turn the agent against its customer: no hacker, scam message or poisoned document can talk it into sending money, goods or data elsewhere.
Customers need both. A precise agent that can be hijacked is a liability, and a safe agent that makes mistakes is useless.
Every agent in this report runs on a shared platform that Raphaël Mansuy calls the enterprise agent chassis: identity, permission-aware context, governed tools, one gateway to models, and a trust layer that records what happened.14 AWS (Bedrock AgentCore), Microsoft (Foundry, Entra Agent ID) and Google (Gemini Enterprise Agent Platform) now each sell most of it. The platform is necessary, and it is no longer a differentiator. What the clouds leave open is where trust is won or lost. Their guardrails, such as prompt shields and content screening, protect the safety of individual calls. Precision across a sequence of steps is largely left to each company.
| Layer | Trust dimension | What a customer or regulator needs to see | What the clouds provide | What remains open | Cognitive Fab perspective |
|---|---|---|---|---|---|
| Identity and entry | Safety | Who is acting, for whom, under which mandate | Per-agent identity on all three clouds | Identity and policy stop at each cloud's edge; outside agents arrive through open protocols (A2A, MCP). Yogi Porla argues for a chain of custody from the user through every agent and service, a context-bound harness on each action, and an automated kill switch.26 | Verified claims and lineage checked at every step (in design) |
| Agent runtime | Precision | The process completes correctly, whatever happens | Managed runtimes with checkpoints | Durable state is not a correct process: nothing checks that a claim pays once or a unit sells once | polyflow: processes checked for every order of events |
| Knowledge | Precision | The agent applies the company's own expertise | Permission-aware retrieval and memory | Curated domain knowledge, the gap all three clouds leave open | polyx: expertise held as rules the agent is bound to |
| Policy and audit | Safety | Every action was allowed, and can be reconstructed | Policy as code, tracing, audit logs | Policy is not portable across clouds, and it gates single calls, not sequences of steps | polyaxion: gates on sequences, replayable ledger, one policy across clouds |
| Readiness | Precision Safety | The system holds under peak load and failure | Evaluations and A/B tests after release | Nothing rehearses peak load or failure before release | polytrion: rehearsal before go-live |
They combine the three traits that expose a business most to AI: a large share of knowledge and document work, high-volume decisions made under regulation, and enough scale that small efficiency gains are worth billions. McKinsey estimates generative AI's annual value at $200 billion to $340 billion for banking, $400 billion to $660 billion for retail and consumer goods, and $60 billion to $110 billion for pharma.13 In each, one company is used as a representative case.
In all five companies, the threat has the same shape: an agent that sits next to the customer and takes over the final outcome. It is a finance agent that manages money across banks, cover that pays itself from data, a shopping agent that decides where to buy, AI that chooses the therapy, or agents that let clients build and run their own systems. None of them attacks the step the incumbent is best at.
The incumbent keeps the customer if it fields its own agent on the final step first, and if it can prove that agent is trustworthy on two counts: precision, so customers can rely on the outcome, and safety, so no one can turn the agent against them. The agent platform itself is now a commodity sold by the clouds. Trust is not. Precision is built from stated expertise, processes checked against every order of events, and rehearsal before peak load. Safety is built from verified identity and gates on every action that moves money, goods or data. Both must be demonstrable to customers and regulators alike.
This report was prepared by Cognitive Fab from public information available as of October 1, 2026. Facts marked live are drawn from the sources listed and have not been checked against company filings. Lifecycles, threats, trust questions, readiness questions and recommended actions are Cognitive Fab's analysis, intended to frame the decisions each company must make. Cognitive Fab has no affiliation with the companies named, and they have not reviewed or endorsed this analysis. Content marked as the Cognitive Fab perspective describes our own products and is shown separately from the diagnosis.