AI may know a lot. But who gets to decide?

Introduction: a plausible recommendation is not yet authority

An AI detects that a customer has exceeded its credit limit and recommends blocking the next delivery. The data in the ERP system seem clear. The recommendation seems plausible. But sales knows about a special agreement, accounting is expecting a payment, and management has deliberately approved the order. Who gets to decide now?

This is where it becomes clear what using artificial intelligence in a company is really about. Evaluating information is not the same as judging a situation as a business. And a convincing recommendation does not yet confer decision-making authority.

The article “AI Shouldn’t Make the Decision, So How Can Context Tell Us Who Should?” by Progress Data Platform takes up this distinction.^[1] It draws attention to a question that easily recedes behind the enthusiasm for autonomous agents: does the system know not only what should be done, but also who is allowed to decide?

For us at PRODVIS, this is a central design task of digital transformation. We neither want to limit AI to the role of a non-binding adviser nor give it blanket authority to act. What matters is combining its capabilities with a reliable ERP data foundation, clear processes and unambiguous responsibilities. Only then can we sensibly determine where it should act on its own – and where it must support the right person with the right information.

1. A correct analysis is not yet a correct decision

Let us return to the example from the beginning. The AI agent’s analysis – credit limit exceeded, invoices overdue – could be entirely correct. Even so, automatically blocking the delivery might be the wrong action.

Perhaps a payment has already been announced but not yet matched. Perhaps there is an approved special agreement. Perhaps the delivery concerns a contractually sensitive deadline. Or management has deliberately decided to accept a temporarily higher risk for this customer.

The example shows the difference between a data finding and a business decision. The finding describes a situation. The decision also takes into account goals, obligations, exceptions and responsibilities. An AI that ignores this difference does not just automate a work step. It tacitly assumes that a recognised pattern already amounts to authority to act. That is exactly what we should prevent.

2. Context is more than additional content for a language model

Progress rightly criticises a narrow understanding of context: more documents, better search and additional data access can improve answers. But they do not yet answer the question of who is allowed to act.^[1] For our work connecting ERP, processes and AI, a threefold distinction is therefore helpful.

• Business context: Which customer, which order, which agreement and which data source are affected?
• Process context: What stage is the case at? Which steps have already been completed, and which processes depend on it?
• Decision context: Who is responsible? Which limits apply? When is an approval required – and who takes over if the responsible person is unavailable?

These layers belong together. Understanding a customer file does not mean being allowed to change it. Recognising a payment bottleneck does not mean being allowed to withhold a payment.

Context should not only improve the answer. It must also define the limits of the action that follows.

3. No accountable decision without reliable ERP data

In the mid-market, all three layers of context share a common origin: the ERP system. That is where customer master data, credit limits, open items, orders, deliveries, conditions and postings live. That is where it is documented what status a case is in and who last changed it. And that is where – ideally – roles, permissions and approval limits are stored.

An AI can only act as responsibly as this data foundation allows. If an incoming payment has not yet been posted, the customer appears overdue. If a special agreement is recorded only in an e-mail rather than in the customer master, it does not exist for the agent. If there are duplicates in the customer master, a risk is spread across several records and remains invisible. The analysis would then be formally correct – and factually wrong.

Data quality in ERP is therefore not IT hygiene but a prerequisite for any delegation to a machine. Gartner puts the cost of poor data quality at an average of at least USD 12.9 million per year per organisation – an estimate from 2020, before agentic AI.^[3] With AI agents, the stakes rise: errors no longer stay in a report but can turn directly into actions.

We see four requirements for a reliable ERP data foundation:

• Timeliness: postings, incoming payments and status changes are recorded promptly so that the AI sees the actual state, not yesterday’s.
• Unambiguity: master data is maintained, duplicates are cleaned up, and terms such as “overdue”, “blocked” or “approved” are defined consistently across the company.
• Completeness: agreements, exceptions and approvals are documented in a structured way in the system – not only in mailboxes or in individual employees’ heads.
• Traceability: origin, change history and permissions are visible, so every recommendation can be traced back to its data basis.

Progress describes the same principle for manufacturing: an AI system investigating a quality problem should not draw conclusions from a single test result alone, but take into account the specification valid at the time, the supplier batch, process parameters and earlier deviations.^[5] Applied to ERP, this means: the AI needs the connected business transaction, not the isolated record.

There is another point that is often overlooked. An agent must be able to recognise when the data foundation is not sufficient. If information is missing or two sources contradict each other, that is not a reason to improvise but a signal to hand over. A cloud ERP with a unified data model, a clean role concept and documented workflows – as we implement it in our projects with Exact Online, for example – is therefore not just an administrative system. It is the backbone of responsible automation.

4. The human in the loop needs a clear mandate

“A human will check it at the end” sounds reassuring. As a process description, however, this sentence is insufficient. Which human? With what expertise? With what authority? On what information basis?

The Progress article makes clear that people in decision processes are not arbitrarily interchangeable.^[1] For the mid-market, this results in a concrete design task: an AI system should not pass a case to any available employee, but to the responsible role – with a traceable justification and the information relevant to the decision. In its governance guide, Progress similarly states that AI decisions must be routed so that people can review, escalate or challenge them.^[2]

For a discrepancy between purchase order and invoice, this could mean: purchasing assesses the agreed conditions, goods receipt confirms the quantity actually received, and accounting evaluates further invoice processing. The AI can bring these perspectives together. But it should not hide their differences behind a seemingly clear overall recommendation.

The European AI Act sets a comparable requirement for high-risk systems: the persons entrusted with oversight must be able to interpret the output correctly, remain aware of the tendency to over-rely on the system – “automation bias” – and be able to override the result.^[4] Even where an ERP process does not fall into this category, it is a good benchmark. A good decision template therefore shows not only what speaks in favour of a measure. It also makes visible which information is missing and which uncertainty remains.

5. Design decision paths, not just AI workflows

From a PRODVIS perspective, this is where the real change of perspective lies. An AI workflow describes how a system processes information and uses tools. A decision path also describes when an insight may become a binding action. Before any automation, five questions should be answered:

1. Which information – and from which sources in the ERP – must be current and complete for this decision?
2. Which missing or contradictory data prevent automatic execution?
3. Which role is accountable for the decision?
4. Within which explicitly defined limits may the system act on its own?
5. How are execution, exceptions and handovers documented traceably?

This means autonomy is no longer a blanket property of an agent. It becomes a limited authority within a specific process. A system could, for example, independently assign a case and prepare a reply, while a change to payment terms still requires approval. What matters is not whether the agent can technically do both. What matters is whether the company deliberately allows both.

6. Processes, data, visibility: a shared design task

This idea connects directly with our PRODVIS guiding principle. Processes clarify how work and decisions are connected. Data – above all master and transaction data in the ERP – creates a reliable foundation and makes its origin and limits visible. Visibility shows what the system has recognised, proposed or executed – and where a decision is still pending.

Our aim is to design these three perspectives together. A powerful model alone neither resolves an unclear question of responsibility nor a contradictory approval process – and it does not repair incomplete master data. Conversely, governance should not be understood as an additional control layer that is placed on a finished AI application after the fact. It belongs in its design.

Conclusion: progress does not mean replacing people everywhere

We do not read the Progress article as a rejection of autonomous AI. Rather, it is a call to define autonomy precisely. Not every action needs manual approval. But every automated action should be based on a deliberately granted mandate and a reliable data foundation. Where information is missing, rules collide or authority ends, the system should not improvise but hand over in an orderly way.

Because that is exactly where responsibility begins: not with the quality of an answer, but with the question of who may draw which consequences from it. For us at PRODVIS, this means: do not introduce AI in isolation, but connect it with the processes, ERP data and responsibilities that make a company capable of acting. Who may approve an exception? Which information is still missing? When must a case be stopped? And how will it remain traceable later why a measure was carried out? These questions are not an obstacle to innovation. They describe the design task.

Our goal is therefore neither full automation at any price nor manual approval of every single step. It is deliberately designed collaboration: AI prepares, brings information together and acts on its own where its mandate is clear. People decide where judgement, special authority or entrepreneurial assessment are required. Processes give this collaboration structure. Reliable ERP data gives it a sound foundation. Visibility makes decisions and their consequences traceable.

That is how we want to measure progress: not by how rarely a person has to intervene, but by how well the interplay works. An AI that recognises its limits and hands the right case with the right information to the responsible person is not less capable. It fulfils its mandate.

The decisive question is not: how much autonomy is technically possible? It is: which autonomy makes sense in this process – and who is accountable for it?

References

1. Progress Data Platform (2026): AI Shouldn’t Make the Decision, So How Can Context Tell Us Who Should? In: Practical AI: Data to Decision, LinkedIn newsletter, issue 10, published 30 October 2026. LinkedIn newsletter “Practical AI: Data to Decision”
2. Progress Software (2026): AI Ethics & Governance – From Principles to Production. Progress Data Platform Insights. Accessed 31 October 2026. Progress: AI Ethics & Governance
3. Gartner (n.d.): Data Quality: Why It Matters and How to Achieve It. Gartner Data & Analytics Topics (figure based on Gartner research from 2020). Accessed 31 October 2026. Gartner: Data Quality
4. European Union (2024): Regulation (EU) 2024/1689 (AI Act), Article 14: Human oversight. AI Act Service Desk of the European Commission. Accessed 31 October 2026. EU AI Act, Article 14
5. Progress Software (2026): How Progress Data Platform Turns Fragmented Manufacturing Data into Connected Intelligence. Progress Blog. Accessed 31 October 2026. Progress Blog: Fragmented Manufacturing Data

Note: all online sources were last accessed on 31 October 2026. The Gartner figure is based on a survey of reference customers of data quality solution vendors and should be understood as an order of magnitude.

Further reading at PRODVIS

• Agentic AI in ERP for the mid-market
• AI agents are the new customization
• Cloud ERP with Exact Online for the mid-market
• Deutsche Fassung: KI darf viel wissen. Aber wer darf entscheiden?

Created with AI support – reviewed by humans.