Examine an AI business before adopting its investment story
A regional label does not establish a market opportunity, an advantage in costs or a credible exit. This page provides questions for due diligence, not investment recommendations or forecasts of returns. The Playbook does not publish verified regional deal totals, valuation multiples or guaranteed AI margin gains here.
Start with evidence from the business
Ask what customers pay for, what remains manual and what the model actually contributes. Separate signed contracts from prospects, paid usage from demos and measured results from founder estimates. Verify customer references through authorised channels; do not upload confidential diligence files to a general chatbot.
Use synthetic or approved non-sensitive material when testing a proposed workflow. Keep personal financial, employee, patient and identity records out of the evaluation. Removing names alone does not establish permission or safety.
Questions for four possible business models
Sector software
What task is improved for a particular customer, and how is that improvement measured? Examine total implementation, review and support effort. Local language and regulation create requirements to meet, not automatic competitive protection. Do not infer recurring revenue from a pilot.
Financial-service support
Distinguish administrative assistance from credit, identity, fraud or payment decisions. What permissions, regulatory review, evaluation and appeal processes are required? A model demo or partnership announcement does not establish lawful deployment or financial inclusion.
Language-specific tools
Who evaluates the exact language variety and task? What rights permit use of the training or evaluation material? Ask for failures, reviewer disagreements and comparison with existing alternatives. Do not assume global vendors cannot serve the market or that a dataset partnership automatically grants all reuse rights.
AI-assisted services
How much work remains with people, and what happens to quality when volume grows? Ask for measured labour, correction, customer-support and infrastructure costs. Compare periods and comparable jobs; a faster draft is not evidence of a particular margin increase or employment outcome.
Build a comparable cost model
Include model usage, retries, retrieval, hosting, licences, monitoring, human review, support, security and compliance work. Record currency, price dates and assumptions. Test scenarios for higher usage, vendor price changes, customer churn and a switch to a different model. Do not treat a scenario as a prediction.
A local model still has hardware and operator costs. A second provider is not a ready fallback until compatibility, quality, data handling and contractual terms have been tested.
Inspect the evaluation record
Request the task definition, dataset provenance, model/version, prompts, tools, run date, raw outcomes, scoring method and reviewer qualifications. Separate self-reported from independently reproduced results. Include error consequences and refusal/escalation behaviour.
The proposed LAC evaluation method can help frame questions. It does not provide validated numerical rankings or certify a startup. Country guides and sector guides suggest local tasks, not current market-sizing evidence.
Check rights, controls and accountability
- Who may use the data, outputs, code and brand material, and for which purposes?
- Can an agent send, spend, change accounts or access sensitive systems? What enforces those limits outside its prompt?
- Which decisions require qualified human review, and is that review economically and operationally feasible?
- What incidents occurred, what changed, and how were fixes tested?
- Can customers export records or leave the service?
- Which legal or sector questions remain unresolved, and who is responsible for resolving them?
Evaluate the team’s actual customer knowledge and execution, not stereotypes about where a founder lives. Missing evidence is a diligence question, not proof of fraud or proof that a thesis is attractive.
Keep the conclusion bounded
Record what was checked, what was not and which assumptions would change the decision. Qualified financial and legal advisers should address the actual transaction. Do not use this page or a chatbot score as a basis for buying, selling or allocating capital.
Created by Adrian Dunkley | MaestrosAI | maestroailabs.com | ceo@maestrosai.com. Editorial review: 11 November 2026.