Skip to main content

White paper no. 3 · August 2026

Enterprise agentic processes and tomorrow's interfaces

While AI produced answers, conversation was interface enough. Once it executes processes, the useful interface becomes a control room, and reliability becomes a question of architecture before it is a question of model.

This white paper no. 3 (FR + EN) sets out the method: the distinction between a skill and an agentic workflow, the reliability cost of nondeterminism measured on benchmarks, the four software engineering gestures that make a process sustainable, and the three tiers of use with their five conditions for moving up.

Version française →

What you will find inside

Key figures

0.5

the minimum trigger rate a skill must reach over three runs to be accepted by the official test procedure.

Source: Agent Skills open standard, consulted 30 August 2026

36 % / 86 %

the end-to-end success of a twenty-step nondeterministic chain, against three steps, at 95 % success per step.

Source: White paper calculation, independence assumption

73.1 %

success three times in a row on Toolathlon, where step independence would predict 52.4 %: failures are correlated.

Source: Anthropic, System Card Claude Opus 5, 24 July 2026

26.0 %

of tasks completed by the best model on real business flows spanning 47 applications, against under 10 % in April 2026.

Source: AutomationBench (Zapier), July 2026 reading

0 out of 240

unsafe tool actions with a metadata policy barrier, against 82.50 % under naive execution, with no loss of utility.

Source: NetInjectBench, 11 July 2026

21 %

readiness on business processes, last of the seven dimensions measured: the least ready dimension needs neither a licence nor a server.

Source: Deloitte, 11 August 2026 (501 executives)

Access white paper no. 3

Two ways to access the white paper, whichever you prefer:

Read the online version

The full content of the white paper, structured into chapters with anchors, freely available. No form.

Read online →

You can also write directly to contact@paulantoinetual.fr.

Frequently asked questions

What is the difference between a skill and an agentic workflow?
A skill is a reusable task entrusted to the model's judgement: it captures know-how and can be revised by the business, without guaranteeing the moment it fires. An agentic workflow is a sequence decided by code, whose loops, filters and retries are written once and versioned: what it guarantees is the sequence itself, replayable and testable. The Claude Code documentation puts it this way: what is reusable is “the instructions” for a skill and “the orchestration itself” for a workflow.
Why does reliability fall when nondeterministic steps are added?
Because success rates compound. At 95 % per step and under an independence assumption, a twenty-step chain succeeds end to end in 36 % of cases against 86 % for three steps, with no model having changed. The independence assumption is false and the white paper documents it: on the Toolathlon benchmark, the best model of July 2026 succeeds on 73.1 % of tasks three times in a row where independence would predict 52.4 %. Multiplying probabilities therefore remains a worst-case heuristic rather than a law, but the direction of the slope depends on no assumption.
Do I need a tool or a budget to start?
No. The first deliverable is a numbered list of the steps of a real process, each carrying three questions: does it admit a single correct answer, does it require reading unstructured material or synthesising, does it trigger an irreversible action. That work costs half a day, with no tool and no budget, and it shows how many nondeterministic draws the company is about to stack up. The Deloitte study of 11 August 2026 confirms the order of priorities: across seven readiness dimensions, business processes come last at 21 %, the one dimension that requires neither a licence nor a server.
How does this differ from white papers no. 1 and no. 2?
AI Maturity of French SMEs answers “where do I stand?” and From PoC to Industrialisation answers “how do I industrialise?”. This third volume covers what comes next: how a process executed by agents holds up over time, which software engineering gestures make it reliable, and which interface that execution calls for once conversation is no longer enough.