Abstract
AI agents now answer business questions by calling tools that wrap existing systems. Every system can hold correct data and the agent can still give a wrong answer, because a value loses part of its meaning on the way to the decision: its unit, its scope, its status, its date, or the authority behind it. This paper presents Actor-Network Analysis and Design (ANAD), a framework for finding and closing those losses, one decision at a time.
ANAD adapts Actor-Network Theory (ANT), which was built to explain how networks of people and things hold together, into a design procedure. It adds Network Fidelity Engineering (NFE), our proposal for specifying and testing what a value must carry to a decision. The framework has four phases, thirteen steps, four signed gates and eleven documents. It has a chart notation adapted from data-flow diagrams and an optional knowledge graph.
We apply ANAD end to end to a fictional insurer's claims-payment server, and as analysis only to the Citigroup trading incident of 2 May 2022. Desk checks found 8 problems in the specification; an independent re-run found 30 more. Twenty fixes resolve 30 of them; 8 remain undecided. What the framework does is shown in these cases; it is not compared with other methods.
1. Introduction
1.1 Correct data, wrong answers
A claims handler asks an AI assistant whether an insurer can pay a contractor's invoice of £18,000 today. Every system the assistant consults is accurate. The right answer is to pay £12,500, and only after a claims manager approves it. In an illustrative trace, not a recorded run, a first design of the assistant answers that the full £18,000 can be paid.
Nothing in that design is broken. The invoice system reports the gross total, including VAT the insured can recover. The policy system returns the policy version in force today, not on the date of the loss. The claims ledger reports payments for one claim, not for the policy year, and leaves out a pending payment. The answer reaches the handler as a recommendation that reads like an approval. Each step passes a value on, and each changes what the value means. The data is right; the meaning that reaches the decision is not.
Three established practices each cover part of this:
- Data quality checks whether a source is accurate.
- Component testing checks whether each call is well formed.
- Evaluation checks whether an answer is right, for the cases someone thought to test.
None of them follows a value's meaning across the whole chain from source to decision. In agentic systems that chain is long. Its last automated reader, a language model, can read labels in their everyday sense: "amount" as the amount payable, "available" as available to use, an empty result as "none exist".
1.2 Why Actor-Network Theory
ANT treats an outcome as the effect of a network of people, things and texts, and asks where meaning changes as it passes between them. It distinguishes intermediaries, which pass meaning on unchanged, from mediators, which transform it. It treats inscriptions, such as a field name or the wording of a rule, as actors that shape what others do. These ideas fit agentic data flows closely: a tool description is an inscription, and a field that looks like a copy can be a mediator.
ANT was built to describe and explain, not to design. Its founders offered principles and rules of method, never a procedure (Section 2). Using its concepts as design steps is therefore our adaptation. Earlier information-systems work used ANT mainly to explain outcomes, though Hanseth and Monteiro (1997) drew design implications from it.
1.3 Contributions
The paper makes four contributions:
- A framework. ANAD combines a network layer from ANT with a fidelity layer from NFE. It specifies an ontology, a glossary, eleven documents, a chart notation, a lifecycle with signed gates and four cycles, and the roles that own each record (Section 4; Appendices B–H).
- Two worked applications. A fictional insurer's claims-payment server is taken through all thirteen steps on paper; the steps that need recorded runs could not be completed. The Citigroup trading incident of 2 May 2022 is analysed from the regulator's Final Notice and stops, correctly, at the analysis gate (Section 5; Appendix I).
- An evaluation of the specification itself. Two desk checks and an independent re-run tested whether the specification could be applied as written, and recorded where it failed (Section 6).
- A narrow statement of the gap. Design-oriented uses of ANT exist, as do related practices in socio-technical design, safety engineering and data quality. What none combines, for one decision, is named records, signed gates, a test against an independently computed answer, and hop-by-hop tracking of named loss modes through software agents (Section 3).
1.4 What the paper does not claim
The agent runs in the worked cases are illustrative traces, written to show reasoning. They are not recorded results. NFE and its terms are our proposal, not a standard. The framework is shown working in cases; it is not claimed to outperform other methods, and that comparison has not been made. The paper claims a contribution, not a first. It was drafted with an AI model (Claude, made by Anthropic) working under the author's direction; Section 6.4 says which checks the model ran.
1.5 How the paper is organised
Section 2 introduces ANT and what the framework borrows from it. Section 3 reviews related work. Section 4 presents the framework. Section 5 applies it to two cases, with a third as context. Section 6 reports how the specification was tested. Section 7 discusses criticisms and limits, and Section 8 concludes. The appendices hold the full specification, the full history of ANT behind Section 2, and the complete case records.