joinwell52

Posted on Sep 3 Originally published at joinwell52-ai.github.io

Global Real Digital Workers & SaaW Commercial Landscape 2026-2

#ai #opensource #programming #architecture

Models, Governance, Pricing, and Capability Tiers

Products may all write reports, change code or handle business processes while differing in delivery quality, failure handling and the final bill. Assessing readiness for a role requires more than a feature list: how does the model decide, how does the runtime constrain actions, and who accepts the result?

This volume uses our SaaW (Software as an Agent Worker) concept: software continuously delivers work under explicit responsibilities, permissions and governance rules. See From SaaS to SaaW. Volume 1 compares products and roles. Here we ask: Which model makes decisions? How are model errors contained? Are software and model costs separate? How do these capabilities support D1–D5 classification?

1. Models Are Not Interchangeable Processors

Business facts and fixed tools can mask model differences in short generation, single functions and simple support. Continuous decisions amplify those differences in repository engineering, open research and long cross-system tasks.

The project team previously encountered severe unsupported fabrication in model testing. Such experience cannot permanently condemn newer versions, but it motivates one rule:

Supporting a model does not automatically qualify a digital worker for a role. A configuration with a history of serious fabrication must pass real role evaluations again.

The same model behaves differently in a chat window, a coding agent or a runtime with terminals, tests and checkpoints. The evaluation unit is:

Model
+ runtime
+ tools and sources of truth
+ context compression and memory
+ permissions and recovery
= worker configuration

1.1 Representative Product Model Transparency

Transparency concerns disclosure of model choices and routing, not model quality or role capability. CodeFlowMu is our self-developed collaboration application/runtime and appears last alongside other products.

D1 denotes assistance; D2 predominantly preset automation; D3 autonomous task execution; D4 a sustained bounded role; D5 stronger authorization, acceptance and recovery.

Product	Public Model / Routing	Transparency	Model-Cost Separation	Tier
Salesforce / Agentforce	Multi-model stack; exact routing varies by tenant and release and is not disclosed as one fixed model.	Medium	Model cost is bundled into credits or flat access, not transparently separated.	D3
Microsoft / Copilot Studio / Microsoft 365 Copilot	Primarily GPT-family models with selectable models in some scenarios and external services.	Medium	Platform licensing plus credit meters; premium reasoning may add meters, so costs are partly separated.	D3
ServiceNow / Autonomous Workforce / AI Specialists	Now LLM plus third-party models; the platform is positioned as model- and cloud-agnostic.	Medium	Model and platform costs are generally bundled in enterprise contracts.	D4
Workday / Sana Agents / Agent System of Record	Gemini became the default model for Sana for Workday in 2026; other agents may use multiple models.	High	Credits meter agent work; model cost is not fully itemized.	D3
Glean / Glean Agents / Independent Agents	Supports broad frontier-model choice; routing is task- and tenant-dependent.	Medium	Generally bundled into enterprise contracts; model cost is not public as a separate line.	D4
Oracle / Fusion Agentic Applications	Supports Llama, Cohere, external industry models, and partner models.	High	The application platform and model/cloud consumption are partly separated.	D4
Google Cloud / Gemini Enterprise Agent Platform	Primarily Gemini models, with support for customer and open models.	Medium	Clearly separated.	D3 (Platform)
Cognition / Devin	Core Devin routing is not fully disclosed; the product ecosystem exposes multiple frontier models.	Medium	Model, compute, and software capabilities are bundled in quotas/ACUs.	D4
Sierra / Sierra Agents / Horizon Agents	Underlying model stack and routing are not disclosed.	Low	Model cost is fully bundled into outcome pricing.	D4
Factory / Droids / Missions	Multi-model router supporting Claude, Gemini, Kimi, MiniMax, BYOK, and local models.	High	BYOK enables clear separation of software and model costs.	D4
Harvey / Harvey	Publicly uses models from OpenAI, Anthropic, Google, Mistral, and others.	Medium	Models and platform are bundled.	D3
Bland AI / Voice AI Agents	Proprietary voice/dialog stack; the company says production calls do not rely on third-party frontier models.	Medium	Fully bundled.	D3
Laiye / Laiye Worker / WEP	Model-neutral routing lists DeepSeek-V4, Qwen3-Max, Kimi K2, Doubao 2.0 Pro and GLM-5.1; selection considers task, budget, latency and compliance.	Very high	Platform subscription, credits and enterprise-owned/private models can be partly separated.	D4 candidate (sustained role unverified)
Tencent / WorkBuddy	August 2026 official list includes Hy3, GLM-5.2/5.1, MiniMax-M3, Kimi-K2.7-Code/K2.6 and DeepSeek-V4-Flash/V4-Pro; custom OpenAI, Anthropic and Gemini integration.	Very high	Enterprise/custom models and private deployment allow model costs to be separated from software licensing.	D3
Alibaba / DingTalk / Wukong	Exact model names are not disclosed; premium tiers expose model configuration parameters.	Low	Seat and compute credits are partly separated.	D3
Baidu AI Cloud / Keyue Digital Employees	ERNIE / Baidu AI Cloud model stack.	High	Model and product package are bundled.	D2–D3
Deepexi / DeepWorks / FastAGI	Configurable multi-model stack with FastAGI as runtime/guardrail layer.	Medium	Model and platform costs can be partly separated by deployment.	D3
Langboat / LangClaw	Mengzi model family with configurable models.	Medium	Potentially partly separated; contract terms are not public.	D3
SAP / Joule / Joule Agents	SAP Business AI multi-model stack; different use cases can use different foundation models.	Medium	Software subscription and premium agent usage are partly separated.	D3
Dust / Dust Agents	20+ models including GPT, Claude, Gemini, Mistral, and DeepSeek.	High	Credits bundle model and work usage; not fully separated.	D3
Mistral AI / Agents API / Enterprise Workflows	Primarily Mistral models with enterprise hybrid deployment.	High	Platform and model calls can be metered separately.	D3 (Platform)
Relevance AI / AI Workforce	Multi-model with BYO LLM/API keys.	Medium	Explicitly separates agent Action fees from model/tool Vendor Credits.	D3 (Platform)
Torq / Socrates	Exact models are not fully disclosed; the platform meters AI Credits.	Medium	Models and agent actions are bundled into fixed credits.	D4
causaLens / Digital Worker Factory	Model-agnostic, deployable on customer-selected models and infrastructure; workflow, trusted facts, causal verification and deterministic gates are central.	High: model strategy is explicit, without a fixed default model	Runtime and model costs can be separated on customer infrastructure.	D4
CodeFlowMu / Author-developed product	Models reason; the runtime organizes team collaboration and work state. The current complete model/routing list was not verified.	Architecture responsibilities are public; current configuration list not uniformly disclosed	Current quotation and model-cost separation unverified.	D4 (software engineering role)

1.2 Four Commercial Model Structures

Figure 05. Structural relationships only. Widths do not encode prices; deployments can combine patterns. Source: the author's synthesis of the corresponding analysis in this article and its listed references.

Structure	Meaning	Examples
Fully bundled	Customers pay per minute, conversation, result or plan without a visible model-cost line	Sierra, Bland, Intercom, Shulex
Seat + credits	Seats and agent usage are separate, while credits still bundle model costs	Microsoft, WorkBuddy, Wukong, Salesforce
Platform actions + model cost	Agent execution and model/tool consumption are metered separately	Relevance AI, Google Agent Platform
Bring your own model	Customer keys/private/local models can shift model costs to the customer; additional action/runtime fees depend on the plan	Factory, Relevance AI, enterprise-custom WorkBuddy, some UiPath/Laiye deployments

Relevance AI illustrates the distinction: action fees cover agent execution, while Vendor Credits cover model/tool costs, with supplier-cost metering and customer keys described in its documentation. Separately explaining software value and model cost is a useful design for CodeFlowMu; actual charging still depends on the product plan.

1.3 Why Real Engineering Amplifies Model Differences

A digital software worker must preserve:

Repository structure and architectural constraints.
Consistency across files.
Tests, builds, dependencies and CI state.
Explicit user prohibitions.
Failure causes and prior repairs.
Git, worktree and branch state.
Completion criteria and supporting evidence.

Engineering task	Where differences appear
Write a function or explain an error	Leading models may be close
Change several files compatibly	Global consistency matters more
Failure → diagnosis → repair → retest	Recovery and causal judgment diverge
Large refactoring/migration	Locally correct changes can damage the whole system
Multi-hour unattended work	Drift, forgotten constraints and repeated side effects dominate

Factory Missions offers one response: acceptance contracts, feature lists, research notes and runtime rules are shared files; Workers start with fresh context, Validators check output, and an Orchestrator organizes continuation or rework.

Cognition also compares model cost with real engineering rollouts. In the study's August 2026 materials, Devin Desktop/CLI offered GPT-5.6 Sol and compared score against cost on tasks such as FrontierCode, rather than relying only on launch benchmarks.

Worker capability depends jointly on the model, runtime, tools and facts, state and recovery, and governance. A standalone model score does not measure role delivery.

Why Long Tasks Amplify Small Differences

An earlier error can enter later context as a supposed fact and influence planning and actions. Checks, retries and recovery can interrupt that propagation. Long-task evaluations must observe error accumulation, detection and correction rather than one benchmark score.

Models therefore belong in role-admission decisions.

2. RPA, Agents and Digital-Worker Boundaries

RPA (Robotic Process Automation) usually executes predefined interface or system operations. Classification depends on the division of responsibility between models, workflows and runtimes.

Type	Who decides the next step?	Capability	Typical risk
Traditional automation	Programmer-defined	Executes stable fixed processes	Interface/rule changes break it
LLM + workflow	Model interprets; workflow determines most paths	Natural-language entry and limited exceptions	A workflow bot may be marketed as an employee
Agentic execution	Model plans, selects tools and adapts	Completes a complex task	Drift, repeated side effects, confident errors
Digital worker	Model/runtime maintain role, tasks and state	Sustained cross-cycle work	Model changes alter role capability
High-trust software worker	Work governance contains errors	Can be authorized for production business work	Complete public evidence remains scarce

RPA is useful in high-risk writes:

Model: understanding and planning
+ rules / business systems: facts and boundaries
+ automation / APIs: deterministic execution
+ humans: exceptions and accountability decisions

The question is whether RPA acts as the worker's hands or determines the entire product's behavior.

3. How Commercial Products Contain Hallucination

Models can generate unsupported content. Enterprise systems must reduce errors, detect them and contain their effects on business actions.

Defense	Techniques	Purpose
Grounding	CRM, ERP, knowledge, retrieval, business APIs	Constrain business claims with records
Scope	Roles, topics, skills, tool allowlists	Restrict questions and actions
Identity/least privilege	Role permissions, access control, agent identities	Limit authority despite model intent
Action schemas	Parameterized tools, business objects, deterministic processes	Validate parameters, permissions and rules
Call-time approval	One-time tokens, human approvals, escalation	Separate capability from current permission
Runtime evaluation	Simulation, regression, output checks, injection detection	Find deviations before and during operation
Observability/audit	File, command and API logs; cost/outcome dashboards	Diagnose, assign responsibility, stop and improve
Recovery/rollback	Checkpoints, versions, compensation, idempotency	Avoid duplicate writes and lost responsibility

3.1 What Enterprise Governance Already Does

Salesforce constrains models with CRM facts, permissions, actions, monitoring and audit.
ServiceNow discovers, observes, governs and disables misbehaving agents, recording files, commands and API calls.
Glean provides request-level checks, identities, scoped credentials, checkpoints, evaluations, approvals and rollback.
UiPath combines injection/sensitive-data defenses, mandatory human intervention and deterministic robot execution.
Oracle supplies roles, approval hierarchies, transactional context, shared memory and action history.
Torq governs security tools, policy, approvals and remediation audit.

These are real controls, mainly over assets, security, access and runtime behavior.

3.2 Why Work Governance Adds Further Questions

Open knowledge work must also establish:

Which source supports a market number, legal conclusion or technical judgment?
Is it a source fact, inference or unproven assumption?
How are conflicting sources preserved?
Who reviews independently and accepts the conclusion?
Does formal responsibility survive model changes or process restarts?

Agent governance addresses safe software operation. Work governance also addresses why a conclusion is valid and who is responsible for it.

3.3 Five Reliability Layers

Figure 06. Different defenses answer different questions. This is not a vendor score or a claim that one product implements all five layers. Source: the author's synthesis of the corresponding analysis in this article and its listed references.

Layer	Techniques	Question addressed	Remaining limit
1. Prompt guardrails	System prompts, prohibited terms, formats	Reduces obvious boundary violations	Confident fabrication remains possible
2. Grounding	Retrieval, enterprise knowledge, CRM/ERP, APIs	Grounds business facts outside model memory	Does not guarantee correct reasoning
3. Agent governance	Identity, permissions, allowlists, approval, audit	Controls actions and authority	Does not prove delivery claims
4. Decision verification	Trusted Facts, structured claims, deterministic gates, judges	Checks whether decisions follow trusted facts and rules	Does not establish complete organizational accountability
5. Work governance	Formal tasks, delivery, fact checking, observation, acceptance authority, continuity	Establishes responsibility, accepted results and valid state	Requires organizational adoption and institutional fit

causaLens describes Layer-4 mechanisms: Trusted Facts retain system-computed evidence; Structured Decision Claims organize facts, rules and recommendations; a Deterministic Verification Gate checks rules before model evaluation and returns failed outputs to execution.

This adds checks on decision support beyond retrieval-augmented generation (RAG) and prompts. Fact checking and independent evaluation already appear in commercial technical designs.

TMPA (Textual Multi-Agent Process Architecture), used by CodeFlowMu, also asks how checked judgments become traceable formal work results through tasks, delivery, acceptance and responsibility facts.

3.4 Comparing Governance Mechanisms

Product	Grounding	Action authority	Human approval	Runtime evaluation	Deterministic verification	Independent Judge/EVAL	Durable responsibility facts
ServiceNow	Strong	Strong	Strong	Very strong	Partial	Evaluation system	Weak/platform-local
Glean	Strong	Per-request/action checks	Strong	Strong	Partial	Alignment/evals	Weak
Relevance AI	Connector-dependent	Edge configuration	Strong	Task View/logs	Weak	Weak	Task-level
causaLens	Very strong	Strong	Strong	Strong	Very strong	Judge Agent	Medium
Factory	Code/test facts	Tools/sandbox	User takeover	Validator	Validation contract	Validator	Project-task level
SAP	SAP business facts	Enterprise permissions	Strong	Enterprise monitoring	Business rules/processes	Partial	Platform-process level
CodeFlowMu	Tasks, reports and engineering evidence	Role permissions and governance decisions	Human approval entry	Public regression and implementation cases	Fixed-input state/rule checks	Separate QA; EVAL observes alongside execution	Persistent tasks, reports, reviews and recovery records

These are qualitative comparisons from public mechanisms, not uniform test scores. CodeFlowMu evidence includes public collaboration and V1.8.0 results in I1.0. Role authorization checks do not prove call-time authorization for every external action.

4. Why Support and Open Research Differ in Delivery Difficulty

Support and enterprise transactions have clear facts, bounded actions and measurable completion: orders, logistics and refund eligibility come from systems and rules; operations include lookup, address changes, refunds, tickets and escalation.

Open research has no single truth database. It requires source selection, quality assessment, conflict handling, fact/inference separation and responsibility for conclusions. Retrieval establishes what was found, not whether a conclusion is correct.

Work	Fact certainty	Action space	Completion criterion	Fit for current agents
Support/orders/employee self-service	High	Small	Clear	Very high
Sales/procurement/contract triage	Medium	Medium	Partly measurable	Medium-high
Software engineering	Medium	Large	Partly verified by tests/review	Medium-high, model-sensitive
Business research/strategy	Low	Large	Hard to verify automatically	Low
High-risk finance/legal decisions	Medium	Medium	Major liability	Human final authorization required

5. Pricing Units Reveal the Delivery Unit

Figure 07. Units are grouped into access, activity, capacity and outcomes. Arrows describe different delivery objects, not a universal migration or proof that outcome pricing dominates. Source: the author's synthesis of the corresponding analysis in this article and its listed references.

Pricing Unit	What the Customer Buys	Examples
Seat	Access for a human user	M365 Copilot, Wukong, WorkBuddy
Credits	Agent/model capacity	Microsoft, WorkBuddy, SAP, Torq
Action	One tool/business action	Salesforce, Relevance AI, SAP
Conversation	One customer interaction	Salesforce, Ada, Decagon
Resolution	A validated solved issue	Zendesk, Intercom, Yellow.ai
Lead	One researched/contacted prospect	11x, Artisan
Minute	Voice labor	Bland, PolyAI, Parloa
Agent Compute Unit	Planning, tools, code, and runtime resources	Devin
Outcome	Retention, sale, resolution, collection, etc.	Sierra

Pricing Changes Highlighted in the 2026 Study

HubSpot: Customer Agent at $0.50/successful resolution; Prospecting Agent at $1/recommended outreach lead.
Intercom Fin: $0.99/outcome (resolution, procedure handoff or disqualification); successful sales qualification $9.99. Failure or a user-requested human handoff does not incur an outcome fee.
Zendesk: From May 2026, Automated Resolution/Resolution Allowance meters successful AI resolutions without escalation, with complexity tiers.
Relevance AI: Platform action fees and model/tool Vendor Credits are separate. The cited documentation lists $80/additional 1,000 Actions, supplier-cost pass-through and customer keys.
Laiye Worker: RMB39/199 plans, credits and enterprise governance/private deployment illustrate China's hybrid structure.
WorkBuddy Enterprise: RMB198/316 per user/month, credits and private deployment show continued seat pricing plus agent usage.

Software access
→ human productivity
→ agent actions
→ digital work capacity
→ verifiable business outcomes

Outcome charging is not yet universal. Many products still use seats, credits and enterprise contracts.

6. Capability Tiers and Sample Totals

The same Volume 1 comparison contains 55 entries: 54 original commercial samples plus CodeFlowMu. Capability and paid commercial status are separate; D2–D3 boundary cases retain their labels.

Tier/label	Count	Examples	Conclusion
D1	0 in main table	Pure Q&A/search excluded	Does not imply absence from the market
D2	0 standalone	Boundary entries appear below	Boundaries are not forced into D2
D2–D3 boundary	9	Baidu Keyue, Robin, Juro, LayerX	Planning versus preset workflow needs product-level assessment
D3	36	Salesforce, Microsoft, UiPath, SAP, Zendesk, Intercom, WorkBuddy, Wukong	Largest sample category, including 3 platforms
D4	9	ServiceNow, Glean, Oracle, Devin, Sierra, Factory, Torq, causaLens, CodeFlowMu	Public-evidence initial assessments; CodeFlowMu is limited to software engineering
D4 candidate	1	Laiye Worker	Execution/routing evidence exists; sustained role needs verification
D5	0	No award	Public evidence does not fully cover the report's conditions
Total	55	54 commercial samples + 1 self-developed product	Mutually exclusive labels

Counts describe sampling and classification, not market share. Platform configurations may occupy different tiers.

6.1 Role Capability Versus High-Trust Governance

Devin spans repositories but needs large tasks decomposed and ambiguous architecture clarified.
Sierra spans days/weeks, with limited model, evidence and recovery disclosure.
ServiceNow/Oracle have strong governance but focus less on evidencing open knowledge conclusions.
Factory supports long tasks and multiple models, while acceptance is mainly engineering validation.
Glean has identity, credentials, evaluations and rollback; customer-built agents vary.
Torq acknowledges nondeterminism and recommends fixed workflows when determinism is required.
causaLens has trusted facts, deterministic verification and judges, but public evidence is incomplete for cross-task formal responsibility, side-effect receipts and final governance-state reconstruction.

7. Model Admission for Digital Workers

Figure 08. Role risk → capability profile → model/host/tool/runtime evaluation → allow, restrict, require review or reject. This is a proposed framework, not completed benchmark results. Source: the author's synthesis of the corresponding analysis in this article and its listed references.

A runtime should maintain role requirements and configuration admission, rather than only model = X.

Role risk and task type
→ Required capability profile
   Facts / instructions / long tasks / tools / recovery
→ Evaluate model + host + tools + runtime
→ Allow / restrict / require independent review / reject

Admission should produce actionable conclusions rather than only an aggregate score:

Configuration A
Information organization: allow
Content drafts: allow with review
Business research conclusions: restrict
Production code changes: independent tests and review required
High-risk external actions: reject

7.1 Model Capability Profile

Dimension	Real evaluation question
Factual reliability	Does it admit uncertainty or invent sources, files and API results?
Long-task stability	Does it drift after 30 minutes, two hours or multiple days?
Tool use	Are tool choice, parameters and order correct?
Instruction retention	Does it retain prohibitions on editing/publishing and requirements to verify?
Recovery	Does it resume correctly after tool failure, host restart or model change?
Self-checking	Does it detect contradictions rather than defend earlier conclusions?
Evidence discipline	Does it separate external facts, tool results, inference and assumptions?
High-risk actions	Should it receive write authority or only propose actions?

Example minimum role requirements:

Support lookup: moderate reasoning + strong facts
Software development: repository understanding + tools + tests + recovery
Business research: source discipline + conflict handling + independent review
Critical finance/legal action: strong model + deterministic rules + human final authorization

Model substitution must not be confused with equivalent capability.

8. Volume Conclusion

The competitive questions are who can:

Constrain model uncertainty within facts, tools and authorization.
Admit different models only to suitable roles.
Explain software costs, model costs and delivery clearly.
Reconstruct accountability after failure rather than merely retry an API.
Turn completion claims into evidenced, independently reviewable delivery.

Volume 3 examines public implementations of identities, task ledgers, pre-action authorization, recovery, phone supervision and installation, then develops CodeFlowMu's product direction.

9. References

Product and pricing information follows the original research scope; current vendor terms govern actual purchases and configurations. Tiers and qualitative comparisons are not uniform product tests.

Original English article · 中文版

Research Center

DEV Community