An AI readiness assessment is not a maturity badge. It is a way to decide where to start, what to fix first and what not to attempt yet. The output should be a short list of use cases with evidence, a score you can defend, and a 90-day plan with an owner.
This step-by-step guide is designed for an SME that has not built an AI capability yet. It takes one or two workshops plus some document review. It does not require a data science team.
Prerequisites
- An executive sponsor who can allocate time and approve a pilot.
- Two or three candidate use cases from real work, not a technology demo.
- Access to process owners and a sample of the data each use case needs.
- A system list, an incident log and the current data protection policy.
- A scoring sheet and a facilitator who can keep the discussion honest.
Step 1: Decide what the assessment is for
Choose the decision: whether to invest in AI this year, which use case to pilot first, or whether to buy a tool a vendor is pitching. A clear decision keeps the assessment from becoming a general audit.
Step 2: Build a use-case inventory
List candidate use cases in plain language: "draft replies to customer complaints", "extract invoice line items", "predict stock-outs", "summarise call notes". For each, record the user, the current process, the volume, the cost of the current process and the consequence of an error.
Step 3: Score value and feasibility
Use a simple one-to-five scale and write down the evidence for each score. A weighted model helps, but the weights are a business choice. For example: value 30 percent, data readiness 25 percent, feasibility 20 percent, risk 15 percent and cost 10 percent. The weights are illustrative, not a standard.
A use case that scores high on value but low on data should not be rejected. It may become a data project first. A use case that scores high on feasibility but low on value is a distraction.
Step 4: Assess data readiness
- Availability: does the data exist in a system, a spreadsheet or only in email?
- Quality: are records complete, deduplicated and labelled consistently?
- Rights: do you have a lawful basis to use the data for this purpose?
- Access: can the pilot team get the data without copying it into personal drives?
- Retention: how long is it kept, and how are deletion requests handled?
Step 5: Assess technology and integration
Check whether the systems expose APIs, whether identity is centralised, whether logs exist and whether data can move without manual exports. For a language model use case, also check how prompts and outputs will be logged, how secrets are stored, and whether a vendor trains on your data. The answer belongs in the contract, not the sales deck.
Step 6: Assess people and process
Ask who will own the workflow after the pilot, who reviews exceptions, and what training the team needs. A model that saves the manager time but adds unchecked work for the front desk will be bypassed. Change capacity is a real constraint, especially in a small team.
Step 7: Assess governance and risk
- Is there an AI policy covering acceptable use, data handling and disclosure?
- Who approves a new model or vendor?
- How are errors, bias and harmful outputs reviewed?
- Is there human oversight for decisions that affect customers or staff?
- What is the incident path if data leaks or a model produces a harmful result?
The NIST AI Risk Management Framework is a practical structure for this step: govern, map, measure and manage. ISO/IEC 42001 is useful if you want a management system rather than a one-off review.
Step 8: Model the economics
Estimate the full cost per task: licence or API usage, integration, human review, monitoring and rework. Compare it with the current cost, not with zero. If the benefit depends on removing a role, say so explicitly and test the assumption before the pilot.
Step 9: Score and prioritise
Bring the evidence into one table and rank the use cases. Pick one pilot with a measurable baseline and a clear stop rule. A second use case can be prepared in parallel only if it shares the same data and platform work.
Step 10: Write the 90-day plan
- Days 1-15: confirm the baseline, the owner and the success metric.
- Days 16-45: build a narrow pilot with a human review step and a feedback log.
- Days 46-60: test quality, cost and failure modes against the baseline.
- Days 61-75: decide to scale, change or stop, using evidence.
- Days 76-90: document the operating model, train users and set the next review.
Evidence to keep
- The use-case inventory and the score with its supporting notes.
- A data sample and a short data-quality finding for each shortlisted case.
- The system and integration list, including who owns each API.
- The risk review, the approvals and any open legal questions.
- The baseline metric and the pilot's stop rule.
Common mistakes
- Starting with a tool and looking for a problem.
- No baseline, so the benefit cannot be proved.
- Ignoring data cleaning because the demo looked good.
- No owner for the workflow after the pilot.
- Underestimating human review and exception handling.
- No kill criteria, so a weak pilot continues by inertia.
The Bangladesh context
For a Bangladeshi SME, the binding constraints are often data scattered across WhatsApp, paper records and a local ERP, plus limited experience with model evaluation. Start with a use case where the data already exists in a system and the output can be checked by a person. That is a better first step than a broad chatbot programme.
Metrics
Track task accuracy, review rate, cycle time, cost per task, adoption and incident count. For a first pilot, also track the percentage of outputs a human accepts without editing; that number usually decides whether scale is justified.
Where this fits in our work
If you want a facilitated assessment before committing budget, start with Operations & Business Software. The pricing page explains how scope is built. For the adoption path, read AI adoption for SMEs, year by year.
Sources
- NIST, AI Risk Management Framework, accessed 17 November 2026.
- OECD, AI Principles, accessed 17 November 2026.
- ISO/IEC 42001:2023, AI management systems, accessed 17 November 2026.
- McKinsey & Company, The State of AI 2026, accessed 17 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.