We write the way we build: sourced, reproducible, and fair to Agentforce. The series examines the platform job by job, asking what works, where completion breaks, and why. All of it is free on Substack.
MOTA's founders have a long history of building enterprise platforms in highly regulated industries. With the arrival of LLMs in the enterprise, they grew uneasy as "hallucination" became an accepted explanation for wrong answers that carry real cost. The academic and industry research below says it formally: LLMs cannot plan or verify their own work, so both must live outside the model. It made that case on puzzles; MOTA brings it to live enterprise workflows.
| Source | What it shows |
|---|---|
| Kambhampati et al., ICML 2024: LLMs Can't Plan, But Can Help Planning in LLM-Modulo Frameworks · PDF · slides · talk | LLMs cannot plan or self-verify on their own. The prescription is an external planner and sound verifiers around the model. |
| Valmeekam et al., NeurIPS 2022: PlanBench | The standard benchmark for LLM planning. Progress across model generations has been, in the authors' words, surprisingly slow. |
| Valmeekam et al., NeurIPS 2023: On the Planning Abilities of LLMs | GPT-4 solved roughly a third of simple block-stacking problems, and close to none once the object names were disguised. The plans looked plausible and were wrong. |
| Stechly et al., 2024: On the Self-Verification Limitations of LLMs | Asking a model to check its own work does not rescue performance. Verification has to come from outside the model. |
| Valmeekam et al., 2024: LLMs Still Can't Plan; Can LRMs? | Reasoning models improve on PlanBench but degrade as plans grow longer, at far higher cost. The gap narrows; it does not close. |
| Apple ML Research, 2025: The Illusion of Thinking | On controlled puzzles, reasoning models collapse to zero accuracy past a complexity threshold, even when handed the algorithm. |
| AIGuys (Medium): LLMs Still Can't Plan and Reason | A widely read practitioner walkthrough of the evidence, for readers who want the argument without the notation. |
| FireCompass: Why LLMs Cannot Plan | A practitioner's summary of the same evidence, from a team building agents in production. |
| Salesforce AI Research, 2025: CRMArena-Pro | The enterprise confirmation: agent accuracy falls from 58% to 35% when CRM flows go multi-turn. |
Our reading of the field, offered in good faith to encourage debate. The researchers cited here are independent of MOTA and do not endorse it. If new evidence shows models planning reliably, the Completion Index will report it.
Yes, for many jobs. Not yet for all. This essay sets the framework the rest of the series follows.
Sales operationsWhere single-record automation shines, and where cross-record sweeps lose the thread.
RevenueRoll-ups are easy. Reproducible forecast logic is not.
Sales developmentMulti-turn conversations are precisely where published benchmarks show agents drop hardest.
Account managementLong-running flows across many objects, and the quiet cost of one wrong join.
Availability, scalability, security: Agentforce passes the classic list. Reliability, durability and auditability were never on it, until agents arrived.
The benchmark · LinkedInWhat Salesforce's own benchmark shows, job by job: the 58% to 35% split is architectural, not accidental.
FoundationsA 5-step workflow at "90% accuracy" per step completes 59% of the time. A short lesson in compound reliability.
FoundationsHow runtime enforcement and execution traces make AI systems more auditable than human processes.
We also publish on LinkedIn and Medium, where new analysis appears as it ships.
Across 2025 and 2026, enterprise IT coverage converged on one finding: AI workflows break in production without deterministic control, and the burden of adding it has landed on customers.
| Coverage | What it reports |
|---|---|
| Salesforce Engineering, Oct 2025: Toward guided determinism | Salesforce's SVP of AI Research: its "most sophisticated customers are struggling" to keep autonomous agents on-topic in critical workflows. The prescribed fix is deterministic structure over agent execution. |
| CIO.com, Jan 2026: Agentforce recalibration raises costs and complexity for CIOs | Analysts at Greyhound Research and Avasant on the shift to Agent Script: autonomy without guardrails is unscalable, ROI timelines stretch, and the work of workflow mapping, testing and governance now falls on customer teams. In regulated industries, deterministic automation "is not optional; it is essential." |
| VentureBeat, 2026: The workflows breaking enterprise AI | Salesforce's own framing at the Agentforce Operations launch: many enterprise workflows were designed around human judgment and break when agents follow them literally. |
| CIO.com, 2026: Why agentic architecture is still so puzzling | IT leaders rolling out dozens of agents still lack the architecture underneath them; most agents stall at pilot on workflow redesign, measurement and security. |
Our reading: the market has agreed that determinism is required. The open question is whether every customer should build it themselves, one script at a time. MOTA's answer is a completion layer that does the planning, gating and audit as a product.
| Source | What it establishes |
|---|---|
| CRMArena-Pro, Salesforce AI Research (arXiv:2505.18878) | Agent accuracy falls from ~58% single-turn to ~35% multi-turn on realistic CRM flows; near-zero default confidentiality awareness |
| Gartner press release, June 2025 | Over 40% of agentic AI projects predicted canceled by end-2027; costs, unclear value, inadequate risk controls; "agent washing" |
| Forrester, Predictions 2026: Automation at the Crossroads | <15% of firms to enable agentic features; deterministic automation as the backbone of reliability and compliance; process intelligence to rescue 30% of failed AI projects |
| IDC, Agentic AI Is Breaking Your ROI Model | New agent metrics: autonomous resolution rate, decision quality, escalation accuracy |
| IBM, Straight-Through Processing | The enterprise standard for flow completion: end-to-end, zero-touch |