Service-as-Software: A New Model for AI Agents

Key takeaways

1. Date / source. December 3, 2025, Thoughtworks blog (Insights / Blog / Generative AI section). Author: Matt Kamelman.
2. Pivot concept. Service-as-Software (SaS) — succession to SaaS in the age of AI agents.
3. Canonical formula. "SaaS sells tools, SaS sells outcomes." ### The SaaS vs SaS distinction | Dimension | SaaS (Software-as-a-Service) | SaS (Service-as-Software) | |-----------|------------------------------|---------------------------| | Sold | Tools / capabilities | Outcomes / impact | | Action | Enable humans to solve problems | Automate reasoning itself | | Pricing | Seats, features, usage | Demonstrated alignment + impact | | Mechanics | Software executes human instructions | System operates on dynamic goals | | Memory | Often stateless | Memory across interactions | | Coordination | Often intra-application | Autonomous, cross-tool, cross-API | | Cognitive contract | Humans instruct the machine | Humans + machines collaborate | ### The three capabilities of an agentic SaS system 1. Dynamic goals — operate on the basis of objectives, not fixed workflows. 2. Memory across interactions — retain context between sessions. 3. Autonomous coordination — orchestrate tools and APIs without human intervention. ### The Cognitive Contract — three principles | # | Principle | Kamelman's definition | |---|----------|---------------------| | 1 | Interpretable and auditable | "Users need to be able to understand why the system made a decision." | | 2 | Aligned with human goals | "The system's objectives must match human intent and ethical boundaries." | | 3 | Trained and iterated in real time | "Systems continuously refine behavior based on feedback." | ### The Cognitive Orchestrator — three functions | # | Function | Description | |---|----------|-------------| | a | Feedback loop design | Review step by step, instrument learning loops | | b | Managing uncertainty with guardrails | Business rules + circuit breakers to contain drift | | c | Measuring alignment | Quantifiable alignment score (agent KPI) | ### Three archetypal SaS agent examples | Domain | Agent type | |---------|--------------| | Marketing | End-to-end campaigns (briefing → creative → media → measurement) | | Finance | Forecast modeling | | Operations | Request triage | ### The structuring historical analogy ` Mainframes ↓ Client-server ↓ Web/Cloud ↓ (cognitive contract remains: humans instruct machine) SaS / Agentic systems ↓ (cognitive contract evolves: humans + machines collaborate) ` > "The cognitive contract remained the same: humans had to instruct the machine." — every past generation upheld this contract. The agentic rupture is the first to renegotiate it. ### Connection to the watch dossier #### "Billable hours / per-seat → outcome-based" convergence
4. Kamelman/Thoughtworks. (2025-12-03): SaaS seats and features → SaS demonstrated alignment and impact. Canonical Anglo-Saxon conceptual formalization.
5. Sternfels/McKinsey. (OfficeChai 2026-01-14): "migrating away from pure advisory work to outcomes-based model". Application to consulting.
6. VoxComm/Mandese. (March 2026): "billable hours are dead, AI killed them", 4-shift outcome framework. Application to agencies.
7. Bain Rule of 40. (April 2026): outcome-based pricing models could shift revenue from fixed seats to labor/operations economics. Application to SaaS.
8. Bain cross-system labor. (May 2026): accumulated execution data as moat. Application to enterprise.
9. → Convergence: Kamelman supplies the conceptual vocabulary (SaS) that chronologically precedes the sector manifestations (consulting, agencies, SaaS) and unifies them. The Kamelman fiche is the conceptual pivot of the "outcome-based services" cluster. #### "Cognitive orchestrator / agent supervisor" convergence
10. Kamelman. cognitive orchestrator (feedback loop + guardrails + alignment scoring).
11. Osmani Agent Harness Engineering. (2026-04-19): agent = model + harness, ratchet principle.
12. Mornati. What is a Developer When We Use Coding Agents? (2026-03-14): Agent Supervisor.
13. Wescale Augmented Software Factory. (2026-05-03): Strategic Judge + Agent Manager.
14. Habert PROJ-AI. (2026-05-05): six zones (DOCS/IDEAS/DR/OUT/DOCTRINE/AGENT) + Decision Records, 7 dimensions.
15. Frizzo. (2026-05-05): "the new bottleneck is supervision".
16. → Cross-cutting convergence: a new organizational role is emerging, named differently by each author but with three invariant functions: feedback design, guardrails, alignment measurement. #### "Cognitive contract" convergence
17. Kamelman. cognitive contract (interpretable + aligned + iterated).
18. Talisman Ontology Pipeline Refresh. (2026-05-04): AI augment trained engineers, not replace.
19. Karpathy. (2026-04-29): outsource thinking but not understanding.
20. Osmani Cognitive Surrender. (2026-05-05): Cognitive Offloading (healthy) vs Cognitive Surrender (toxic).
21. Anthropic Skill-Formation. (cited by Osmani): 17% lower comprehension among junior engineers.
22. → Convergence: the human-machine cognitive contract is being renegotiated across multiple axes (services, taxonomies, code, design, learning). #### "Interpretability + auditability" convergence
23. Kamelman. principle 1 of the cognitive contract.
24. DORA ROI 2026. (2026-04-21): AI is an amplifier — measurement and governance are prerequisites.
25. Habert PROJ-AI. Decision Records, 7 dimensions.
26. Anthropic political bias. (2025-11-13).
27. → Convergence: auditability has become a structural prerequisite for agents in production. #### FR / European position vs. Anglo-Saxon
28. Kamelman/Thoughtworks. Anglo-Saxon conceptual formalization.
29. Wescale. (FR): Augmented Software Factory → equivalent FR terminology with an industrial-military framing.
30. Habert. (FR): PROJ-AI + doctrine.
31. Tatsyi/Raiffeisen Bank Ukraine. organizational testimony.
32. → Connection: SaS is the canonical Anglo-Saxon vocabulary; Wescale and Habert propose alternative or complementary FR vocabularies. For FR executive-committee presentations, leverage SaS (international English) and anchor it in FR vocabulary. ### Limitations to flag
33. Conceptual article. with no empirical validation — no client case studies, no commercial figures, no actual pricing model (which metrics of alignment/impact, measured by whom).
34. 3 agent examples. (marketing / finance / operations) not detailed: no named actors, no results, no timeline.
35. "Demonstrated alignment and impact" as a pricing metric. Kamelman provides no measurement tools. Yet this is precisely the critical operational workstream needed to make SaS commercially viable.
36. No engagement with obstacles. regulation, legal liability in the event of agent error, cybersecurity, memory storage (GDPR), tools/API interoperability.
37. Thoughtworks blog source. ≠ peer review; obvious institutional positioning (Thoughtworks sells agentic services, so SaS serves its commercial interest).
38. Author Matt Kamelman. a sparsely documented public profile outside Thoughtworks — brand authority more than authorial authority.
39. SaS concept = a remix of pre-existing ideas. outcomes-based pricing (Sternfels, January 2026), value-based models (Mandese, March 2026), outcome-based pricing as tailwind (Bain, April 2026) — Kamelman precedes them chronologically (December 2025), but it is not certain the SaS term will prevail over outcome-based services, agentic services, or other competing formulations. ### To be leveraged for
40. Executive committee / board presentations. 2026 strategic vocabulary — Service-as-Software as a canonical term to frame pricing shifts.
41. AI architects / product strategists. the 3 agentic capabilities + 3 cognitive contract principles + 3 cognitive orchestrator functions = a design grid for agents in production.
42. Pricing strategy / CFO. the "no longer pay for seats, pay for alignment + impact" argument — a structured pricing shift.
43. HR / org design. the cognitive orchestrator role as a new role to formalize (job description, training, performance measurement).
44. Connecting the outcome-based 2025-2026 cluster. Kamelman = chronological conceptual pivot preceding the sector manifestations; to cite when introducing analyses of the "billable hours killed by AI" cluster.
45. FR / European connection. SaS as the canonical Anglo-Saxon term; anchor it in FR vocabulary (Wescale Augmented Software Factory, Habert PROJ-AI) for FR executive committees.