AI Power

How S-I-C-T helps enterprises build smarter generative AI solutions

An Evaluative Heuristic for Enterprise Generative-AI Use-Case Selection

The Structure–Information–Cohesion–Transformation (S·I·C·T) Lens Applied to the Google Cloud 101-Blueprint Corpus

Róth Miklós Roth Complexity Lab · CRS AI Marketing & SEO Ügynökség

Working paper — conceptual / position paper. Not peer-reviewed. Version 1.0 · 2026

Epistemic status of this paper. This is a position paper, not an empirical study. It proposes an interpretive lens and a research programme; it does not report validated results. Every substantive claim carries one of four calibration badges, defined in the legend at the end. The single most important honesty constraint, stated up front: the S·I·C·T stability relation S + C ≥ I + T is advanced here as a heuristic ordering claim, not as a measured inequality or a derived law. Sections 6–7 make the conditions under which it could become testable explicit.

Abstract

Enterprise adoption of generative AI is constrained less by model capability than by use-case selection: which process to automate first, where human oversight is non-negotiable, and how to recognise when a deployment is stable enough to scale. Catalogues of "use cases" answer what is possible but not what is prudent to attempt next. This paper introduces a deliberately lightweight evaluative lens — Structure (S), Information (I), Cohesion (C), Transformation (T) — and applies it to a public corpus of 101 architectural blueprints published by Google Cloud (Seroter & Sanin, 2025). We make three contributions. First, we describe a faithful, source-grounded mapping of the corpus into ten operational clusters (§2, §4). Second, we use the S·I·C·T lens to surface recurring control motifs across clusters — patterns such as "provenance-as-structure," "human-in-the-loop as cohesion guard," and "variant generation as transformation requiring structural counterweight" (§5). Third, and most importantly for intellectual honesty, we calibrate the epistemic status of each claim (§6) and set out a falsifiability programme that converts the informal heuristic S + C ≥ I + T into operational, pre-registerable hypotheses (§7). We argue the lens earns its place as a communication and triage device, while explicitly declining to claim predictive or law-like status for it.

1. Introduction

A recurring observation from practitioners is that the bottleneck in enterprise generative AI is not "can the model do X" but "of the hundreds of things the model could plausibly do, which one should we attempt first, and how will we know it is safe to scale?" Public use-case catalogues — including the Google Cloud corpus examined here — are inspiration artifacts: they enumerate possibilities and reference architectures, but they do not rank candidates by deployment risk, nor do they offer a shared vocabulary for the failure modes that recur across otherwise unrelated industries.

This paper does not propose a new model, benchmark, or empirical result. It proposes a lens — a compact conceptual scaffold intended to do two jobs well: (i) give cross-industry deployment patterns a common name, and (ii) act as a triage heuristic during the pilot-selection stage of a programme. We are explicit throughout that a lens of this kind is valuable in proportion to its clarity and parsimony, not to any claim of having discovered an underlying law of complex systems. Where the framework has previously been stated in slogan form, this paper's job is to expose its seams: what it asserts, what it cannot assert, and what would have to be measured to move any of its claims from heuristic to grounded.

[HEURISTIC] The central wager of the paper is that a four-term decomposition is the right size for this triage task: rich enough to separate the common failure modes, coarse enough to be remembered and applied in a meeting.

2. The corpus and its provenance

[GROUNDED] The primary corpus is the public Google Cloud article 101 gen AI use cases with technical blueprints (Seroter & Sanin, 2025), published 21 August 2025. It contains 101 architectural blueprints organised into ten industry groups, each blueprint pairing a stated business challenge with a Google Cloud reference stack and a data-flow pattern. It is described by its authors as the technical complement to a separately maintained list of 600+ customer-inspired use cases.

[GROUNDED] This paper, and the Hungarian-language practitioner derivative that motivated it, are independent works. They are not Google publications, claim no partnership or endorsement, and reuse the corpus only as an enumerated problem set under fair description, with attribution. The derivative re-groups the blueprints into ten locally meaningful clusters and, in two instances, relocates a blueprint to a more natural cluster (e.g., the "compare vendor proposals" pattern is grouped with manufacturing procurement rather than retail). These are editorial regroupings; the underlying problem statements are preserved.

[GROUNDED] A capability audit of the corpus finds the stated AI functions to be conventional and appropriately bounded: document extraction, demand forecasting, semantic/vector search, multimodal inspection, summarisation, and retrieval-augmented question answering, each typically gated by human review in the high-stakes blueprints (healthcare, finance, safety). We found no capability that is overstated relative to publicly demonstrated practice as of the corpus's publication date.

3. The S·I·C·T lens

[HEURISTIC] The lens decomposes any candidate deployment into four terms:

Term Reading in a deployment context Typical instrument Structure (S) The stable scaffold a system is measured against: schemas, brand rules, protocols, permissions, provenance, the "single source of truth." Canonical data model, policy, access control, citation requirement Information (I) The inflowing signal the system must absorb: orders, sensor readings, tickets, queries, market moves. Event volume, novelty, ambiguity of inbound data Cohesion (C) The quality of connection between parts — between channels, between machine output and human judgement, between message and audience. Hand-off integrity, consistency across surfaces, reviewer coupling Transformation (T) The rate and reach of change the system performs: variant generation, translation, content production, automated action. Throughput of generated/changed artifacts, automation depth

[SPECULATIVE] The framework additionally advances an ordering slogan:

S + C ≥ I + T — "the structure and cohesion a system carries should at least keep pace with the information it ingests and the transformation it performs."

We label this SPECULATIVE deliberately and without hedging. As written it is not a mathematical inequality: S, I, C, and T are not defined on a common, dimensioned, measurable scale, so the relation cannot be evaluated, falsified, or "validated" in its literal form. Its legitimate use is as a design intuition — a reminder that scaling input volume (I) or automation reach (T) without a matching investment in canonical structure (S) and human/contextual coupling (C) is the recurring shape of brittle deployments. Section 7 sets out what would be required to upgrade it from slogan to hypothesis.

4. Method: mapping the lens onto the corpus

[HEURISTIC] For each of the ten clusters we read every blueprint through the four terms and recorded (a) which term the blueprint's value primarily lives in, and (b) which term, if neglected, produces its characteristic failure. The method is interpretive and non-blinded; it is a structured reading, not a measurement. Its output is a vocabulary of motifs (§5), not a score.

The ten clusters, after regrouping: (1) commerce & e-commerce; (2) marketing, media & content; (3) automotive, transport & logistics; (4) finance, insurance & fintech; (5) healthcare & life sciences; (6) telecommunications, IoT & customer experience; (7) travel, hospitality & mobility; (8) manufacturing, industry & energy; (9) public sector, education & non-profit; (10) technology, enterprise operations & AI governance.

5. Findings: recurring control motifs

The lens is useful to the extent that it makes the same failure mode legible across industries that do not normally share vocabulary. Five motifs recurred across clusters.

[HEURISTIC] 5.1 Provenance-as-structure. In every blueprint where an AI answer feeds a regulated or high-trust decision — compliance Q&A, clinical research search, enterprise knowledge retrieval, policy-document analysis — the value collapses without source attribution. In lens terms: the answer is a Transformation of underlying Information, and citation is the Structure that preserves Cohesion between answer and evidence. The practical rule the motif yields ("never return an answer without a traceable reference") is independently sound; the lens's contribution is to explain why the same rule recurs in clinical, legal, and corporate settings.

[HEURISTIC] 5.2 Human-in-the-loop as cohesion guard. Across healthcare notes, loan underwriting, insurance claims, and crisis triage, the corpus consistently assigns the machine to preparation and the human to decision. The lens reads the reviewer as the load-bearing Cohesion element between high-throughput Transformation and accountable action. The motif predicts where removing the human is most dangerous: precisely where T is high and the cost of an incorrect C-link is irreversible.

[HEURISTIC] 5.3 Variant generation needs a structural counterweight. Campaign creative, ad variation, product descriptions, and multi-channel content are Transformation-heavy by design. The corpus repeatedly pairs them with a brand/style/approval constraint. In lens terms, raising T (many variants) without raising S (brand rules) and C (editorial coupling) is the shape of off-brand drift. This is the clearest informal instance of the S + C ≥ I + T intuition — though, per §3, still only an intuition.

[HEURISTIC] 5.4 Single-source-of-truth resolves channel incoherence. Omnichannel retail, unified customer view in telecom, and catalogue de-duplication are all, in lens terms, Cohesion failures caused by competing Structures (systems that disagree about the same fact). The motif's prescription — converge on one canonical record rather than synchronising many — is a structural, not a modelling, fix.

[HEURISTIC] 5.5 Observability as the structure of the system's own state. The governance blueprints treat monitoring/evaluation as the Information a system holds about itself, which becomes the Structure needed to steer it. The motif generalises the "measure before you scale" discipline that the other four motifs each presuppose.

[GROUNDED] What these motifs are not: they are not predictions, and they do not exceed the plain managerial advice already implicit in the corpus. Their contribution is organisational and mnemonic — one vocabulary for failure modes that are otherwise re-discovered separately in each vertical.

6. Epistemic calibration

This section states plainly what the paper does and does not establish.

[GROUNDED] The corpus exists, is accurately sourced and dated, and its capability claims are conventional and bounded.
[GROUNDED] The mapping in §4 is reproducible as a structured reading: a second reader given the same definitions would produce substantially similar term-assignments.
[HEURISTIC] The S·I·C·T lens is a useful triage and communication device. It earns this status by parsimony and by making cross-industry motifs legible — not by any demonstrated predictive power.
[HEURISTIC] The five motifs (§5) are interpretively supported by the corpus but are not quantified and could be re-described without the lens.
[SPECULATIVE] The relation S + C ≥ I + T is, as written, a slogan. It is not currently falsifiable. Claiming it as a validated stability law would be an epistemic error, and this paper does not make that claim.

A reader who discards the lens entirely loses nothing factual; a reader who adopts it gains a shared vocabulary at the explicit cost of remembering that the vocabulary is a lens, not a law.

7. Toward falsifiability: operationalising S·I·C·T

[HEURISTIC] The honest path from heuristic to grounded is to specify measurements under which the ordering claim could fail. We sketch one such operationalisation as a research programme, not as a result.

Proposed per-deployment proxies (illustrative, requiring validation):

I (information load): inbound event rate × a novelty/ambiguity index (e.g., share of inputs falling outside the training/reference distribution).
T (transformation reach): rate of artifacts generated or actions taken × an irreversibility weight.
S (structure): coverage of a canonical schema/policy (e.g., proportion of decisions traceable to a single source of truth) × constraint enforcement rate.
C (cohesion): integrity of hand-offs (e.g., human-review coverage on high-irreversibility actions; consistency rate across channels).

Candidate falsifiable form. Define a normalised stability margin m = (Ŝ + Ĉ) − (Î + T̂) on standardised proxies. The heuristic predicts that deployments with m < 0 exhibit a higher rate of a pre-specified failure outcome (off-brand output, unrecovered error, rollback, trust incident) than deployments with m ≥ 0, holding model and domain fixed. The claim is falsified if, across a pre-registered sample, no monotone association between m and the failure rate survives controls.

Why this is honest rather than rhetorical. This formulation can lose. It specifies the proxies in advance, names the failure outcome in advance, and admits a result of "no association," which would downgrade the slogan to decorative. A pilot of this kind is well suited to a pre-registered design with timestamped hypotheses (e.g., an OpenTimestamps-anchored pre-registration on the lab's existing pilot infrastructure) to prevent post-hoc reinterpretation. Until such a study is run and reported, the ordering claim remains SPECULATIVE and should be presented as such.

8. Limitations

Single corpus, single author of the lens. Findings derive from one curated catalogue read through a proprietary framework; both choices import selection effects.
Interpretive, non-blinded mapping. §4 is a structured reading, not a measurement; inter-rater reliability is asserted, not yet tested.
No outcome data. The paper observes prescribed control patterns in a reference catalogue, not their realised effect in deployment.
The lens may be redundant. §5's motifs are restatements of sound practice; the lens's marginal value over plain checklists is itself an open empirical question.
Proxy risk. The §7 proxies are plausible but unvalidated; a poor proxy could make the ordering claim trivially true or trivially false for reasons unrelated to the underlying intuition.

9. Related framings (analogy, not derivation)

[HEURISTIC] Two established ideas rhyme with the S·I·C·T intuition and are offered strictly as analogies to locate it, not as foundations from which it is derived.

Requisite variety (Ashby, 1956): a regulator must command at least as much variety as the disturbances it controls. The S + C ≥ I + T slogan is in the same family of "a system's regulatory capacity must keep pace with its load" intuitions, but S·I·C·T is neither a formalisation nor a special case of Ashby's law, and we do not claim the inequality inherits Ashby's mathematical standing.
Information theory (Shannon, 1948): the term "Information" in the lens is used in an informal, business sense (inbound signal to be absorbed), not in Shannon's precise entropy sense. Conflating the two would be a category error.

Naming these analogues is itself a calibration act: it shows where the intuition comes from while refusing the borrowed authority of formal results.

10. Conclusion

We have presented S·I·C·T as a parsimonious lens for triaging enterprise generative-AI use cases, applied it to a faithfully sourced 101-blueprint corpus, and surfaced five cross-industry control motifs. The lens's honest value is communicative and organisational: it gives one vocabulary to failure modes otherwise re-discovered per industry. Its central ordering claim is, at present, a design heuristic and not a law; §7 specifies what a falsification attempt would look like. The framework should be deployed — in client work, in pedagogy, in deployment reviews — with its epistemic status visible, never as settled science.

Calibration badge legend

[ESTABLISHED] — Settled, consensus knowledge. (Not used in this paper.)
[GROUNDED] — Well-supported by a specific, checkable source or directly verifiable observation.
[HEURISTIC] — A useful, defensible device or interpretation; not predictive or proven.
[SPECULATIVE] — A conjecture or slogan; not currently testable or validated as stated.

References

Seroter, R., & Sanin, A. (2025, 21 August). 101 gen AI use cases with technical blueprints. Google Cloud Blog. https://cloud.google.com/blog/products/ai-machine-learning/real-world-gen-ai-use-cases-with-technical-blueprints
Ashby, W. R. (1956). An Introduction to Cybernetics. Chapman & Hall. (Cited as analogy only — Law of Requisite Variety.)
Shannon, C. E. (1948). A Mathematical Theory of Communication. Bell System Technical Journal, 27. (Cited as analogy only — informal vs. formal "information".)

S·I·C·T is a proprietary conceptual framework of the Roth Complexity Lab; it has no peer-reviewed standing and is presented here as a heuristic lens.

Create it