Enterprise AI Agents: L0-L5 Maturity Model & 10 Commercialization Trends

On September 10, 2024, at the Bund Summit in Shanghai, the Shanghai Advanced Institute of Finance (SAIF) and Ant Digital jointly released the 2026 Enterprise AI Agent Commercial Evolution Top 10 Trends Insight . Unlike typical launches that tout model size or benchmark scores, this report asks four practical questions: how agents are delivered, how they monetize, how capital values them, and how enterprises govern them.

1. Definition & L0–L5 Maturity Model

The report defines an enterprise-grade AI Agent as a software agent system embedded in proprietary enterprise systems and business processes, possessing enterprise-level data security, permission isolation, and business execution capability . Three qualifiers act as veto criteria:

Embedded — not a side API; the agent lives inside the business system. A separate web page with manual copy-paste disqualifies it regardless of the underlying model.

Data security & permission isolation — personal agents risk a bad answer; enterprise agents risk a fund transfer. Alignment suffices for the former; architecture is mandatory for the latter.

Execution capability — not advisory. L1 tells you what to do; L3+ does it and owns the consequence.

Five necessary conditions for commercialization (sequential filter):

Condition 1: Enter real business processes, leave pure chat test interfaces. Condition 2: Obtain system-call and task-execution permissions to perform effective operations on underlying systems. Condition 3: Continuously complete measurable business tasks with stability for long, multi-step processes. Condition 4: Have a clear delivery mode with diversified pricing: subscription, API calls, one-time fee, outcome-based. Condition 5: Create measurable, attributable value.

Most "enterprise" products fail at Condition 2 (demo environment only, no production permissions). Survivors often die at Condition 5 (work done but cannot prove the savings).

The AI Agent Commercialization Maturity Model (L0–L5) :

L0 Demo: Conversation & content generation; concept/POC only.

L1 Assist: One-way reference advice; no direct system operation.

L2 Tool: Calls a single tool/API under human instruction; completes local, well-defined tasks.

L3 Production: Executes standardized processes across systems; enters real production environment — the commercialization watershed .

L4 Business Outcome: Tied to core KPIs (revenue, cost, efficiency, risk); forms sustainable outcome-based billing loop.

L5 Organizational: Acts as stable digital employee with defined role; integrated into org structure, permission management, and audit governance.

L2 and below are technical projects; L3+ become business projects with ROI discussions. The counter-intuitive insight: most enterprises stall at L1–L2, and the bottleneck is rarely the model but architecture . Stateless platforms (new session per interaction, manual confirmations, prompt-constrained credentials, fixed toolsets) accumulate "architectural debt" that makes every level-up a paydown.

Three product archetypes map to three business models:

Standardized Agent (general staff, preset skills/workflows) → subscription.

Dev Tools & Platform (enterprise IT, SDK, orchestration, monitoring) → platform + call fees.

Customized Agent (private deployment, deep integration around specific data/flows) → project implementation + ongoing service.

2. Supply & Delivery Layer: Three Trends

Trend 1: Entry-Point Embedding — Standalone Agent Apps Are Dead Ends

Embedding into production workflows and high-frequency entry points is the admission ticket. Salesforce Agentforce leveraged the CRM entry to convert rapidly to ARR; domestically, Alibaba Qwen, DingTalk Wukong, and Huawei HarmonyOS validated super-app/OS-level entry loops. Gartner predicts 40% of enterprise apps will embed task AI agents by end-2026 (up from <5% in 2025) and $234B of enterprise software spend will face "agent arbitrage" by 2030 — the UI you bought depreciates because the operator is no longer human. Yet only 17–31% of organizations have actually deployed agents to production. Embedding is the ticket, not the destination. Example: Salesforce's Claudeforce (Aug 26) made Claude the default reasoning engine inside Salesforce, wrapped in a 37-skill "harness" (AIforce) that feeds controlled data, actions, and workflows — sales reps can run a pipeline review without opening Salesforce. The interface is disappearing behind the agent.

Trend 2: Product System Industrialization — The Custom-vs-Standard Binary Is Broken

Vendors used to choose: standard product (good margin, can't handle complexity) or custom project (deliverable, not repeatable). The shift is toward a modular, combinable product matrix — decompose complex business into standard modules, let customers "Lego-assemble" on demand. Examples: Anthropic's five product-shape categories and Ant's Agentar Super Factory . The real moat: not a hit agent but a factory that sustainably produces, replicates, and evolves agents — a manufacturing mindset shift for software.

Trend 3: FDE Mode Deepening — "Going Deep" Is the New Competitive Edge

FDE (Forward Deployed Engineer) : engineers stationed on-site, co-defining and delivering the "last mile" of value, then feeding de-identified learnings back as reusable platform units. Palantir built its business this way; OpenAI and Anthropic are investing heavily. The win condition: whether field experience becomes reusable platform modules . If FDEs just write custom code per client, it's labor outsourcing with consulting margins. Only when each deployment adds a reusable module to the central platform does FDE turn from cost to asset — the fundamental difference between Palantir and a body shop.

3. Factor Monetization Layer: Two Trends

Trend 4: Data Elements Turn Dynamic — From Training Corpus to Real-Time Business Context

Two-year narrative: "feed data to train models." Report argues a fundamental shift: from static training corpus to real-time business context at agent runtime . Litmus test: can customer status, inventory, contract terms, and permission rules be fetched dynamically at the decision instant? That directly determines task success rate. As general reasoning commoditizes (models cheaper, easier to get), the model itself ceases to be the differentiator. Differentiation moves to "what it sees at the moment of decision." Same model tier: one sees live inventory and contract clauses, the other sees three pasted paragraphs — output quality diverges wildly. This mirrors the "organizational context battle": whoever controls runtime context routing controls the agent's decision ceiling. For buyers: your data governance level directly determines the agent level you can run. This cannot be outsourced.

Trend 5: Billing Shifts to Outcomes — Token → AWU → Outcome

Three-stage leap: Token metering → Agentic Work Unit (AWU) → Business Outcome . Market moving faster than the report:

Salesforce: Aug 26 announced Agentforce billing tied to revenue growth or cost savings. Earlier Help Agent at $2/autonomous resolution with strict definition: ≥2 conversation turns, no explicit negative feedback, no human escalation — any violation = free. Self-service site handled 4.3M consultations, ~70% resolution (vendor-reported, unaudited). Acquired Fin (ex-Intercom) for ~$3.6B; Fin's entire identity was $0.99/resolution. Latest: Agentforce ARR >$1.5B, +240% YoY. Sierra: Pay only on successful ticket resolution. Price band formed: HubSpot $0.50, Intercom Fin $0.99, Zendesk $1.50, Salesforce $2.00 — market accepts the model but 4× spread remains.

OpenAI CFO's candid take: swap scorecard from "cost per token" to "useful intelligence per dollar" — AI should be measured by work completed, not compute consumed.

Cold water required:

AliPartners analysis of 65 enterprise software cos: only 4 fully adopted outcome billing ; >50% still seat-based. 2026 H1 buyer survey: 43% prefer usage-based, only 27% prefer outcome-based — most contracts land in hybrid.

Outcome billing shifts risk to seller but also hands seller the definition of "completion." Salesforce's "≥2 turns, no negative feedback, no escalation" is self-written; silence = acceptance.

Attribution is a minefield. Stripe warns: sales conversion may stem from product changes, marketing, seasonality — not the software.

Cost structure: estimates show AI app companies spend 20–40% of revenue on inference + per-customer fine-tuning, compressing margins to 50–60% vs. traditional SaaS 70–85%. Outcome-based model comes with a far uglier gross-margin sheet. Hence vendors talk Outcome but sign "base platform fee + variable usage" hybrids.

Actionable takeaway: Before accepting "pay for results," nail down three things: how to define completion, how to attribute, whose data counts. Without those, outcome pricing just renames an uncertain bill.

4. Industrial Capital Layer: Two Trends

Trend 6: Value Chain Stratification — Goodbye "Model Is King," Harness Orchestration Layer Becomes New Growth Pole

Industry chain splitting into three layers:

Foundation models: price war accelerates commoditization; non-SOTA pricing power collapses.

Vertical apps: highly fragmented; only deep engagement in high-value scenarios survives.

Middle orchestration/governance (Harness): emerges as new value center .

Whoever captures enterprise workflows and context routing, and guarantees stable long-horizon task delivery, builds differentiation atop base models. The "dumbbell" ecosystem's hollow middle is being filled by encapsulation engineering (Harness) and orchestration platforms . A founder's vivid description: the valuable asset isn't the API (the button-click actions) but the harness wrapping it — skills, docs, rules encoding how top experts run workflows and extract value from systems . Salesforce calls AIforce a harness for this reason. Our earlier Claude Code teardown reached the same conclusion: as model capability converges, differentiation migrates to "unsexy" engineering — session management, fallback, context assembly. Now the report confirms it commercially and assigns value ownership: the most profitable segment is neither the model builder nor the scattershot app maker, but the middle "dispatch hub."

Critical limitation: universal harness hits a 40-year unsolved problem — centralized org knowledge management has never succeeded at enterprise scale. Salesforce conquered sales and stopped; SAP conquered procure-to-manufacture and stopped; Workday conquered HR and stopped. No one became the company's "top leaf." Agents inherit these silos verbatim: Claudeforce is a CRM agent, not an enterprise agent. Root cause: half technical, half organizational — companies ship along their org chart, so no one owns end-to-end flows. AI gets bolted onto each silo instead of re-stitching cross-silo workflows. Near-term winners will likely be teams that own both the harness and the outcome metric in a well-bounded, high-value domain — a narrower, more winnable problem.

Trend 7: China–US Pricing Divergence — US Anchors to Labor Replacement, China to Industrial Value-Add

2026 H1 global AI agent funding already exceeded 2025 full year — a capital super-cycle. Clear pricing logic split:

US: ARR priced at FTE replacement rate × wage level . Precondition: high US labor costs + mature SaaS pay habits. An agent replacing an $80k role supports corresponding pricing; sky-high P/S ratios reflect this narrative. Exit path: defensive M&A by giants.

China: Four-way co-investment (financial VC, platform CVC, industrial CVC, state funds) pricing anchored to industrial value-add . Not "how much wage saved" but "how much incremental industrial value created." Chinese buyers want agents to help business earn more, win more orders — not just cut headcount. Therefore copying the US "sell standard software" playbook fails: if your product cannot enter the production line, the store, the specific supply-chain link and generate auditable increment, your pricing has no anchor.

5. Organizational Governance Layer: Three Trends

Trend 8: Task Collaboration Bureaucratization — Multi-Agent from Flat Emergence to Enterprise Hierarchy

Flat, peer, emergent multi-agent collaboration looks great in demos; in production it exposes three failures: deadlock, divergence, accountability vacuum (Agent A waits for B, goal dilutes across hops, no one owns the screw-up). Observed shift: "supervisor + executor + compliance verification" layered division , with human-in-the-loop approval and full-chain audit to satisfy governance rigidity (controllable, auditable, accountable). Examples: Salesforce's layered trust controls, Alibaba Cloud AgentTeams' gatekeeping and e-seal mechanism. Digital employees are being formally inducted into org accountability systems. That means you must give them positions, reporting lines, job descriptions — the entire management apparatus you apply to humans.

Trend 9: Enterprise Organization Hive-ization — Agents Swallow Invisible Coordination Costs

First wave of agent impact is not role replacement but consumption of hidden coordination costs : information search, cross-department chasing, progress tracking, standardized reporting. Data points: Klarna agents handle workload equivalent to 700 FTE customer service ; 添可 (Dreame) cut customer queue time from 3 minutes to 8 seconds . Report's restraint: headcount hasn't collapsed, but growth logic is rewritten — hiring freezes, layer flattening, budget shifting to high-value roles. Future org: "super individuals + hive combat units," few core experts driving multi-professional agent clusters; "human efficiency" replaces "headcount" as the competitive metric. More credible than "AI replaces humans" — it doesn't promise mass unemployment, it promises organizations won't need as many middle coordinators. If your job is mostly passing info, chasing progress, summarizing reports, the most valuable part of your role gets swallowed.

Trend 10: Security Governance Infrastructure — KYA Becomes Org-Level Commercial Infrastructure

KYA (Know Your Agent) — agent identity authentication. When agents hold real permissions (pricing, ordering, fund transfer), "who is it, what can it do, who's liable" shifts from philosophy to mandatory exam. Two divergent paths: US favors agent passports + liability insurance for market-based risk sharing; China relies on agent identity auth, national standards, and human-takeover "five questions" mechanism for pre/during/post control. Graded autonomy mainstream: low risk full auto, high risk mandatory human approval.

Major move at the forum: KYA national standard initiation ceremony . The standard "Blockchain and DLT — Agent Identity Management Requirements" first systematically establishes a blockchain-based agent identity management model covering architecture, identity ID, establishment, verification, cross-chain interop, and interface specs. Report's closing line: KYA compliance is no longer a brake pad but the passport for agents to call real system permissions. No identity system → no real permissions → stuck at L2 forever.

6. Core Conclusion: China's Moat = "Factory Product Matrix + FDE Deep Service" Dual Flywheel

Rooted in "industrial value-add" not "labor replacement," Chinese enterprise agent commercialization cannot rely on a single SaaS standard product. The real moat is a new delivery paradigm:

Wheel 1 — FDE On-Site Co-Creation: Engineers who know tech and live on-site crack the "thousand enterprises, thousand faces" heterogeneous scenarios. Responsible for "drilling in." Hub — Reusable Unit Deposition: Field experience and de-identified knowledge deposited back to platform as module inventory. Responsible for "spinning." Wheel 2 — Factory Product Matrix: Modular assembly, on-demand combination, continuous production, replication, evolution. Responsible for "replicating out." Scaled delivery lowers cost, which fuels more on-site scenarios; flywheel accelerates sustainably.

Only FDE, no factory: every delivery starts from scratch — revenue scales linearly with headcount, margin declines with complexity. Not a software company; a labor outsourcer wearing an Agent coat. Scale = danger.

Only factory, no FDE: production line churns out standard agents, but Chinese enterprises are heterogeneous — data compliance, demo-to-production, legacy ERP interfaces differ per client. Standard pieces can't crack these scenes → can't enter core flows → can't get runtime context → can't achieve high success rate → can't reach L4.

These two wheels are mutually prerequisite , not parallel. Factory provides scale/efficiency floor; FDE provides path into heterogeneous scenes; FDE's field experience feeds back as factory inventory. Once spinning, latecomers must replicate not a product but the production line plus the accumulated reusable assets co-built by engineers. That is the shape of the barrier.

Project lead, SAIF researcher Wu Yujun : 2026 is the inflection year from "capability competition" to "business closure competition." Future enterprise digital competition will be about building self-executing, continuously evolving human-machine collaborative digital labor systems. He urges decision-makers and investors to seize the 2026 commercialization window, shifting from compute arms race to workflow embedding and digital asset accumulation.

Three Self-Check Questions

Self-Check 1: Is your agent written into the business process SOP with a clear KPI? If it's only on the website saying "we integrated a large model," it's still in the last era — regardless of which model. Self-Check 2: What L-level is it at? Who signed off on that rating? The watershed is L2→L3. Translating "we can do it" to "it runs cross-system unattended and settles by outcome" spans an entire architectural debt. Self-Check 3: If billing by outcome, who defines "completion" — you or the vendor? If this isn't settled, all the beautiful narratives about outcome pricing are just renaming an uncertain bill.

In H2 2026, enterprise AI agent competition isn't on stage. It's in the places no one photographs — ERP interface docs, permission approval forms, resident engineer weekly reports, and the attribution confirmation sheet nobody wants to sign.