You have an idea for using AI in your business. Perhaps Claude could prepare customer enquiries or turn notes into action lists. But do you already have what you need for a test, or does something else need attention first?
You do not need to make your entire business “AI-ready”. You need a suitable framework for the particular task you want to test. A small pilot using approved sample data requires something different from an agent acting independently across several systems.
This self-check helps you identify what is available, what is missing and when not to start. It does not require a technical assessment, just answers to six practical questions.
This self-check is not tied to Claude. We use Anthropic’s Claude as a consistent practical example. The six questions apply equally if you choose ChatGPT or suitable models from China, such as Qwen or DeepSeek. You are checking the prerequisites for the task, not your choice of vendor.
When switching, reassess quality using the same test cases, available features and integrations, permissions, costs and the terms of the specific deployment. A different model does not automatically resolve unclear data permission or missing expert review.
In this article
- Limit the assessment to one task
- Check six prerequisites using actual evidence
- Start, prepare further or stop
- Understand three typical outcomes
- Turn the assessment into a next step
1. Assess a workflow, not the whole company
“Are we ready for AI?” is difficult to answer when it simultaneously means quoting, accounts, customer communication and internal knowledge search.
Begin with one sentence: “We want AI to support this step: …” If you do not yet have a specific task, start with AI in a small business: Which problem should you tackle first?.
For this self-check, we use enquiry preparation at Klarwerk Service, a fictional technical-services business with twelve employees. AI should summarise an enquiry and flag missing information. Prices, appointments, sending messages and binding commitments stay with a person.
Do not turn the following areas into an overall score. Mark each separately:
- Established: there is concrete evidence, not just an intention.
- Still open: a clearly identified preparation step is missing.
- Critical: the proposed test would not be responsible under current conditions.
This is a preparation worksheet, not a certification or scientifically validated maturity assessment.
2. The self-check: Six questions for your pilot
A. Is the task clearly bounded?
Can a colleague explain what goes in, what should come out and where the task ends? “Automate the office” is not enough. “List missing information in a sample enquiry” is testable.
Evidence can be simple: a short process description identifying the trigger, input, output and excluded actions. Also record a simpler alternative, such as an enquiry form.
The task is still open if you cannot yet describe what is going wrong. Observe several cases before choosing software.
B. Can you access and legitimately use the required data?
Are the proposed inputs available, understandable and approved for the chosen system? Do you know which information is confidential or relates to identifiable people?
You need evidence of: the source, a suitable test dataset and agreed use. Completely fictional examples can avoid uncertainty about real customer data in an initial experiment. They do not establish that the later live process will work.
The situation is critical if sensitive documents are to be uploaded without clarified permission or reviewed terms. A business subscription alone does not resolve every question about data use and permitted processing. Involve the appropriate privacy or legal advisers where needed.
C. Are there reliable reference documents?
If AI is expected to refer to services, processes or rules, is there a current, approved source? Who maintains it?
For the first experiment, evidence can be: a small, clearly labelled document with an owner and version date. At Klarwerk, this might be an approved list of information required for service enquiries.
The area is still open when several contradictory documents circulate. First agree on the authoritative version. Adding more files does not automatically resolve the contradiction.
Claude Projects can bring documents and instructions together as context. You still need to establish that the material is current and accurate. [1]
D. Can a person reliably assess quality?
An output is not useful just because it reads clearly. Who can tell whether information is missing, an assumption is wrong or a statement exceeds the assignment?
You need evidence of: a knowledgeable reviewer and short acceptance criteria. For enquiry preparation: the request is described accurately, missing details are flagged, no information is invented and no commitments are added.
Anthropic recommends defining specific, measurable success criteria in advance. For this small experiment, we translate that into a simple checklist; it is not a quality seal for the model. [2]
The situation is critical when nobody can assess correctness but the output will directly determine decisions or customer interactions. Choose another task or arrange expert review first.
E. Are ownership and boundaries agreed?
Who owns the experiment? Who may use the output? Could a wrong answer trigger an action before anyone sees it?
You need evidence of: a named owner, an approval step and a manual fallback. At Klarwerk, drafts remain in the test document. Initially there is no automatic sending or write access to the customer system.
Starting with broad actions but no way to intervene or reverse them is critical. An agent is not a prerequisite for your first experiment. Anthropic distinguishes predefined workflows from more dynamically operating agents and recommends justifying the additional complexity. [3]
F. Have you allowed time and defined a basis for evaluation?
Who reviews the test cases and records rework? Is there a comparison with the current approach? What ongoing effort would be acceptable?
Initial evidence can be: a small tracking sheet for handling time, corrections, unusable outputs and unresolved questions. Set a review date and name the person responsible.
This area is still open if the team is supposed to “try it sometime alongside everything else”. Schedule setup, checking and evaluation. Without that work, an experiment may happen but is unlikely to support a sound decision.
3. How to interpret your self-check
No critical areas, and the limited experiment’s prerequisites are established? You can prepare a small test with human review. This is not approval for an automated live process.
Some areas still open? Name a specific preparation step for each: update a document, appoint a reviewer or create suitable sample data. Complete those steps and reassess.
At least one critical area? Do not start the proposed experiment unchanged. Resolve the permission question, remove uncontrolled actions or choose a lower-risk task.
Good documentation does not compensate for missing data permission. An inexpensive tool does not compensate for missing expert review. That is why these areas are not combined into a reassuring average score.
4. Three examples from a small business
Enquiry summaries: Suitable for a limited test
Klarwerk creates fictional enquiries. An experienced colleague checks summaries against an agreed list. There is no connection to live customer systems and no automatic sending.
Assessment: a controlled experiment can be prepared. Its results initially apply to those examples, not to every future customer enquiry.
Quote drafts: Clarify the foundations first
The team also wants help drafting quotes. However, old service lists contradict each other and nobody has agreed on the authoritative version.
Assessment: clarify the scope and source first. Otherwise, a well-written draft would merely disguise existing uncertainty.
Independent customer replies: Do not start unchanged
A proposed agent would access real emails and reply directly. Access rights, approval and responsibility for errors are unclear.
Assessment: do not start this version. A smaller alternative could produce drafts from appropriate test data, with mandatory human review.
These examples are fictional. They illustrate different start decisions, not measured customer outcomes.
5. Turn the self-check into a small action plan
- Name one task: write down its input, output and boundaries.
- Review the six areas: involve someone who actually knows the work.
- Add evidence: link the document, checklist or ownership agreement instead of simply writing “done”.
- Assign open items: every preparation step gets an owner and a date.
- Resolve critical items first: do not dilute them with positive ratings elsewhere.
- Record the decision: begin a limited experiment, prepare further or deliberately work without AI.
Copyable assessment template
Proposed task:
Expected output:
Excluded decisions/actions:
A – Task clearly bounded:
Status / evidence / missing preparation:
B – Data and permitted use established:
Status / evidence / missing preparation:
C – Reliable reference material available:
Status / evidence / missing preparation:
D – Expert review possible:
Status / evidence / missing preparation:
E – Ownership and boundaries agreed:
Status / evidence / missing preparation:
F – Evaluation scheduled:
Status / evidence / missing preparation:
Decision and reasoning:
Next step / owner / date:
Before and after: An idea becomes a decision
Before: “We should try Claude for our office.” Nobody knows exactly which data, tasks and checks are involved.
After the self-check: there is a limited assignment, suitable inputs and someone who can assess the result. Alternatively, it is clear which prerequisite needs attention first.
“Not yet” is a useful result too. You have identified a concrete preparation step instead of an unclear AI project.
Conclusion: Ready for one task, not for everything
You do not need a company-wide overhaul to learn with AI. A useful first experiment does require a clear task, suitable data, reliable reference material, expert review, ownership and time for evaluation.
Start where you can establish those conditions. More system access and independent actions are separate future decisions, not automatic next steps.
Ask your AI tool / LLM / agent
Describe the workflow in abstract terms, without confidential information:
- “Which prerequisites are missing for this limited test? Ask questions rather than making assumptions about our data.”
- “Which errors should a human reviewer be able to recognise in this workflow?”
- “Turn these open items into a preparation plan. Separate necessary prerequisites from optional improvements.”
Further reading and sources
- AI in a small business: Which problem should you tackle first?
- Why an AI tool is not yet AI transformation
- 1: Anthropic – What are projects?
- 2: Anthropic – Define success criteria and build evaluations
- 3: Anthropic – Building effective agents
This self-check is an editorial worksheet. It does not replace technical security assessment or privacy, legal or specialist advice. The decision to start remains with the responsible people in the business.
