Foundations

Every analytical decision on record: why AI verdicts need a paper trail

Most of what AI does in a communication platform is not writing. It is judging: is this article about us, which topic, paid or earned, which journalist fits. Why those verdicts need versioned instructions and a record, and how Acontema keeps one for each.

, 6 min read, 2 sources

In short

  • Analytical AI decides what ends up in a report long before anyone reads the report.
  • Adoption runs far ahead of governance: most teams can produce with AI, few can show how a given result came about.
  • Acontema keeps the instructions for each kind of decision as numbered versions and stores every verdict with the version, model, input and reasoning behind it.

1AI judges more than it writes

The public debate about AI in marketing is about drafts. In a platform that monitors media and reports to management, most AI calls are verdicts instead. Is this article really about the brand, or does it only mention it? Is it editorial coverage or a distributed press release? Which topic does it belong to, which product does it name, is that product ours or a competitor’s? Is this comment a question, a complaint or praise?

Each answer is small. Together they decide what a coverage chart shows, how sentiment moves, how much media value is reported and which journalists end up on a pitch list. A report is only as defensible as the thousands of verdicts underneath it.

State of Martech 2026 describes how teams take up AI in steps: analysis first, then generation, and AI that acts on its own a distant third. The authors read it as a gradient of trust rather than of capability. Each step asks for more confidence in what the AI does.(State of Martech 2026, p. 33)

The models are capable of more than most organizations are ready to hand them. The bottleneck is governance infrastructure, org design, and executive air cover, not the technology.
Scott Brinker and Frans Riemersma, State of Martech 2026, p. 33

2The governance gap

The same survey finds the controls lagging in every category where AI production is widespread. In content, AI copy production is used by 91 % of respondents, checks on authenticity and AI detection by 37 %. In data, compliance and governance, lineage, and privacy and consent management sit around half, well below the coding and automation use cases next to them.(State of Martech 2026, p. 34)

Production runs ahead of control

Share of respondents using AI for the use case, 2026; n = 208

AI copy production91 %
Data compliance and governance50 %
Data lineage49 %
Customer privacy and consent management47 %
Content authenticity and AI detection37 %

Producing with AI is common. Being able to show how a result came about is not.

Source: State of Martech 2026, p. 34

Policies are not the problem. The report counts 73 % of respondents with a formal policy for generative AI, up from 52 % in 2024, and cites SAS research in which only 8 % of organisations are fully confident in their wider AI governance readiness. A policy says what should happen; it does not show what did.(State of Martech 2026, p. 10)

Why governance keeps losing the budget discussion, according to the authors: it has no immediate return, and its value lies in a crisis that has not happened yet. With the EU AI Act in force and disclosure rules spreading, they expect that to change.(State of Martech 2026, p. 34)

Everyone’s learned to make the sausage with AI. Almost nobody’s bought a labeling machine.
State of Martech 2026, p. 35

The view from the top is similar. In EY-Parthenon’s CEO Outlook of October 2026, only 16 % of the 1,200 CEOs have detailed, real-time visibility into what AI costs and high confidence in judging its return.(EY-Parthenon CEO Outlook, October 2026, chapter 2)

3What a decision record holds

In Acontema each kind of analytical decision has its own set of instructions, kept in the platform as numbered versions: the system instruction, the template the input is placed into, the model and its settings. When a verdict is made, it is stored together with everything needed to explain it later.

One verdict, one record

Is this article about the brand?

Decision type: coverage relevance

What the model saw

Input
The article text, as it was read
Fingerprint
A hash of the exact instructions and text

Under which instructions

Prompt version
Coverage relevance, version 4
Model and settings
The model chosen for this kind of decision

What it decided

Verdict
Mentions the brand only in passing
Reasoning
The product is named once, in a list; the article is about a competitor

What it cost

Tokens and cost
Per decision
Duration
How long the model took
Stored once, never edited. A better prompt becomes a new version; this record keeps pointing to the one that made the verdict.

The fields stored with each analytical AI decision. The example values are schematic.

Two details matter in daily work. The input is stored as the model saw it, with a fingerprint of the exact instructions and text, so a verdict can be compared with what was actually on the page. And where the same input must always get the same answer, for example whether a mention is linked to the right product, the fingerprint lets the platform reuse the earlier verdict instead of asking again. Anything a user asks to regenerate is run again.

4Explaining a verdict months later

The same applies to classification further down the chain. Whether a copy counts as a distributed press release or as editorial coverage decides the paid-against-earned split in a board report; whether a product mention refers to the brand’s own product or a competitor’s decides the product view. Each of those verdicts carries its own record.

5Where the record is kept

More than a hundred kinds of AI step in Acontema run on versioned instructions. The analytical ones, those that classify, match or score, keep a record of each verdict. Among them:

  • Coverage. Whether an article is really about the brand and how central the brand is to it; whether a copy is a distributed press release; which country and which event it belongs to.
  • Topics. Which cluster a topic belongs to, whether two topics are the same, which funnel stage a topic serves.
  • Products. Whether a mention names the brand’s own product or a competitor’s, and which catalogue entry it matches.
  • People and comments. How a journalist’s byline is written; which subject and which kind of question the comments under a post are about.
  • Research. Whether a candidate is really a competitor; whether a suggested content gap is relevant; whether a question in Prompt Research is on topic and really asked.

Where a cheaper method decides first, the record covers that too. A new topic is matched to a cluster by similarity of meaning; only uncertain cases go to a model, and both kinds of assignment are logged.

6What to ask of any AI platform

Whether or not you use Acontema, these are the questions that separate a platform you can defend in front of a client or a works council from one you have to trust blindly.

  1. 1

    Are the instructions versioned?

    Every kind of decision should have instructions that are kept, numbered and changed only by creating a new version.

  2. 2

    Is each verdict stored with its reasoning?

    A label without a reason cannot be checked. The reasoning is what lets a person agree or disagree.

  3. 3

    Is the input kept as the model saw it?

    Pages change and articles are edited. The record has to show what the model actually read.

  4. 4

    Is the cost known per decision?

    Tokens, cost and duration per verdict make the price of analysis visible, which few leaders can say about AI today.

    EY-Parthenon CEO Outlook, October 2026, chapter 2

  5. 5

    Can a person overrule it?

    Records do not replace judgement. In Acontema topics can be moved between clusters by hand, and such moves are marked as manual; entities can be renamed, retyped and merged; drafts go through review and approval before anything is published.

Sources

Quotes are verbatim. Page numbers refer to the printed page of the PDF. Statements about the platform describe what it does today.

See it on your own brand.

Start free trial

14-day free trial. No credit card required.