Use-case owner
A named person is accountable for each use: authorizes the scope, reviews outputs and can stop it.
Applied AI
Artificial intelligence on processes that are already understood
AI does not fix a disorganized operation: it accelerates it, mistakes included. That is why we apply it after the process is ordered and accountability for the result is defined.
- We do not deploy systems that decide autonomously about clients, people, payments or contracts.
- We do not guarantee results, savings or improvement percentages from using a model.
- We do not introduce tools whose output the company cannot check or explain.
Governance
Governance: who decides, and under which rules
Before the first trial we write down how AI use is governed in the company. Without those rules, any deployment is left to individual judgment.
Written usage policy
Which tools are allowed, for which tasks, with which data, and what is expressly forbidden.
Stop criteria
It is defined in advance which signals force suspension and who may do it without asking permission.
Periodic review
AI use enters the management agenda as often as any other operational risk.
Oversight
Human oversight
AI prepares work; the decision remains a person's. The level of oversight matches the impact of the case, and never disappears where third parties are affected.
- Mandatory prior review when the output reaches a client, a supplier or an accounting record.
- Documented periodic sampling for low-impact internal tasks.
- Explicit instruction that the team may reject any output without a technical justification.
- A record of human corrections, since they are the best source of improvement for the use case.
Privacy
Privacy and data handling
- Prior classification of information: what is public, internal, confidential and personal.
- Minimization: only the data needed for the task enters the tool.
- Explicit ban on entering personal or confidential data into unauthorized tools.
- Review of vendor terms regarding retention and use of the information submitted.
- Preference for anonymized or sample data during trials.
Traceability
Traceability
An AI use that cannot be reconstructed cannot be audited or improved. We keep enough record to answer three questions: what was asked, what was answered and who approved it.
- Use-case record: purpose, tool, data used and owner.
- Retention of instructions and relevant outputs according to the impact of the case.
- Evidence of human review and of the changes made before the output took effect.
- Version control of instructions, so a change in result can be traced to a change in input.
Criteria
Use-case selection
Not every case qualifies. We apply the same criteria to all of them and drop a case at no cost when one is not met.
- Stable process
- The task is done today in a repeatable, described way.
- Available data
- The needed information exists and may legitimately be used.
- Verifiable output
- A person can check whether the result is correct.
- Bounded impact
- An error is detected and reversed without relevant harm.
- Enough frequency
- The task repeats often enough to justify the work.
- Owner available
- Someone inside the company can review and decide.
When we say no
- The process changes weekly or depends on constant exceptions.
- The result cannot be checked before it takes effect.
- It requires personal or confidential data with no basis for processing it.
- The decision involved requires human professional or legal responsibility.
Principles
How we apply it
Process first
Before automating, define the input, the output and the quality criterion. Without that there is no way to tell whether the model helps.
With a responsible person
Any model output affecting a customer, a payment or a relevant decision has a named human review.
Bounded scope
Start with a small, measurable and reversible case, not with a company-wide platform.
Data under control
Define what information may enter an external tool and what information never leaves the company.
Application
Common uses in small and mid-sized companies
Operational documentation
Drafting and maintaining work instructions from what the team already explains verbally.
Information preparation
Ordering, summarizing and classifying texts, emails or reports so management reads less and decides better.
Service and response
Draft replies to repeated enquiries, always reviewed before they go out.
Analysis of existing data
Exploring information the company already holds to detect concentrations, deviations and repetitions.
Support for continuous improvement
Grouping incidents and complaints by cause to prioritize where to intervene.
Transparency
Limits we explain before starting
- A model can produce incorrect information that looks correct.
- Quality depends on the information available inside the company.
- Automating a poorly defined process multiplies the problem.
- It does not replace legal responsibility or management decisions.
Next step
Start by knowing where your company stands
The Business Maturity Assessment evaluates six management dimensions in about ten minutes and returns a result per dimension with priority areas.