AI has already entered many businesses, but it is still frequently limited to individual assistants, occasional experiments and manual copy-and-paste. The useful management decision is no longer whether to test AI, but which uses should remain individual and which should become controlled business workflows.
What has just happened
On 14 October 2026, the European Commission published a Eurobarometer survey of 12,826 SME and mid-cap executives. It found that 42% already use AI or algorithmic management systems: 30% in a limited or experimental way and 12% extensively. Customer service is the leading current or planned use, followed by performance monitoring and planning. In France, the 2026 France Num Barometer reports that 53% of SMEs use AI, while only 11% of small businesses say they automate tasks with it. The operational issue is therefore no longer adoption, but the transition from individual use to an integrated process.
Why this concerns you — or does not
Your company is directly concerned if employees repeatedly use an AI assistant to process similar emails, prepare reports, summarise documents, qualify requests or generate information that is subsequently copied into a CRM, ERP or ticketing system.
In that situation, the business may be paying twice: first for the AI tool and then for all the manual work surrounding it. It also depends on individual practices, including personal prompts, inconsistent checks, limited traceability and results that other employees may struggle to reproduce.
Industrialisation is not necessarily justified when:
- the task is occasional, variable and heavily dependent on expert judgement;
- the output is not written into a business system and is reviewed before being shared;
- the input consists of public or appropriately protected information;
- a potential error can be detected and corrected without customer, financial or contractual consequences.
An individual assistant may remain the right solution. Automating too early would turn a flexible tool into a workflow that is more expensive to operate and maintain.
A workflow becomes relevant when the trigger can be identified, the input is sufficiently structured, exceptions can be described and a business owner can define what an acceptable result looks like. The surrounding process is often more important than the power of the model itself.
What this changes in practice
Companies should distinguish between three operating models.
| Model | Suitable use | Operational consequence | Required deliverable |
|---|---|---|---|
| Individual assistant | Research, drafting, summarisation or occasional analysis | The employee initiates the request, provides context, reviews the output and transfers it manually | Usage rules, an approved tool and data-handling instructions |
| AI-assisted workflow | Recurring queues, comparable documents, classification or response preparation | The workflow collects data, calls the model, applies business rules and submits the result for approval | Process map, acceptance criteria, exception handling and rollback procedure |
| Controlled AI agent | Multi-step journeys involving variable paths and several tools | The agent chooses among authorised actions, with restricted permissions and usable activity logs | Permission matrix, action log, stopping rules and an accountable owner |
In customer service, moving to a workflow could mean automatically classifying incoming requests, retrieving relevant customer information and drafting a response. An employee would still approve the answer before it is sent and recorded in the CRM.
For an administrative team, AI may extract and check information from a document, but the workflow must determine where the file is stored, how missing information is reported and who handles the exception. In operations, AI may prepare a planning proposal without making a sensitive assignment decision on its own.
The budget comparison should not be limited to subscription versus development costs. It should include licences, API consumption, connectors, business acceptance testing, monitoring, error correction, maintenance and supplier exit costs. Our article on the full lifecycle cost of automation provides a suitable budget structure.
The decision rules are:
- keep an individual assistant for variable tasks whose quality must be assessed by a specialist;
- choose an assisted workflow when the task is recurring, measurable and connected to a business system;
- consider an agent when the path varies but its tools, permissions and stopping points can be bounded;
- do not automate a process whose ownership, source data or rules are constantly changing.
For an incoming email queue, our AI email agent deployment playbook explains the technical and operational controls to implement.
What to do by 31 January 2027
- By 20 November 2026 — executive management, finance and IT: export AI subscriptions, API invoices and tools enabled within collaborative suites. Assign each expense to a department, an accountable owner and a declared use. The deliverable is a usable inventory rather than a list of product names.
- By 4 December 2026 — business managers: select a recurring task that currently involves manual copy-and-paste. Document its trigger, input data, human decision, exceptions and target system. The resulting process sheet should make it possible to decide whether the use should remain individual.
- By 18 December 2026 — IT, security and the DPO where applicable: classify the data, verify access rights and specify what may be sent to the provider. When personal data is processed, GDPR principles continue to apply, as explained by the CNIL guidance for AI professionals.
- By 15 January 2027 — sponsor and project manager: compare the individual assistant, assisted workflow and controlled agent models. The decision paper should address total cost, integration, reversibility, human controls, acceptance criteria and the shutdown procedure.
- By 31 January 2027 — executive committee: issue a go, no-go or remain-individual decision. If the project proceeds, appoint the business owner, build a test set from real situations and record the starting performance before automation. The profitable AI workflow validation playbook can structure this pilot.
What you can ignore for now
You do not need to turn every prompt into an agent. A well-controlled individual use may remain more economical and reversible than an integrated workflow.
You can also ignore constant model comparisons when they do not change expected quality, total cost, data location or integration options. Replacing the model will not repair a process without an owner or reliable source data.
Using generative AI does not automatically require a high-risk compliance programme. Classification depends on the system’s intended purpose, the decisions involved and the company’s role. However, measures supporting AI literacy have applied under Article 4 of the AI Act since 2 February 2025, with the relevant enforcement framework applying since August 2026. The Commission’s official AI literacy questions and answers explain the requirement.
An AI policy does not replace workflow controls. It establishes general rules, while each project must translate them into permissions, approvals, logs and incident procedures. Our analysis of the France Num corporate AI policy guidance covers that distinction.
Our view at D1 Consulting
The unit of value for enterprise AI is not the prompt, model or number of licences. It is the business process delivered with observable quality, controlled costs and clearly assigned accountability.
We retain general-purpose assistants for variable, lower-risk activities. We recommend a workflow when the trigger, data, human approval and write-back to the target system can be defined. Agents are reserved for cases where autonomy creates genuine value without unnecessarily broad technical permissions.
Our method covers:
- documenting the current process and expected outcome;
- designing human approvals and exception handling;
- prototyping with n8n, Make, Power Automate or the selected stack;
- testing against real business situations;
- managing deployment, knowledge transfer and supplier decisions.
This work is delivered through our Process Automation and Optimisation service. When several departments, systems or suppliers are involved, our IT Project Management service provides the required project framing, acceptance management and production decision process.
👉 Book a free 30-minute diagnostic to determine whether your AI use should remain individual, become an assisted workflow or evolve into a controlled agent.
