Evaluate an AI company as a business first, then test whether artificial intelligence creates a measurable, defensible advantage and whether management controls the risks that advantage introduces. A convincing model demo, a large AI market, or an “AI-powered” label is not evidence of attractive returns.
Your decision should rest on verifiable customer demand, a credible path to cash generation, durable product differentiation, realistic dependence on models and infrastructure, competent governance, and investment terms that compensate you for the risks. The framework below is for diligence, not a buy-or-sell recommendation or individualized financial advice.
As an Amazon Associate I earn from qualifying purchases.
1. Write down the investment case before reading the pitch deck
Start with a one-page thesis that can be tested. Name the customer, the costly or important problem, the alternatives customers use today, and the specific way AI is supposed to improve the product or economics. State your expected holding period and, for a private offering, the security you would receive and the rights attached to it. Then write what evidence would disprove the thesis.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →This discipline prevents a common error: changing the reason for investing every time management presents a new feature. For example, a company may begin by claiming that automation will reduce customers’ labor costs, then rely on a different claim about proprietary data when automation savings are hard to demonstrate. Those are separate hypotheses and require separate evidence.
#1 Best Overall
- Customer problem: Who pays, who uses the product, and what measurable pain causes a purchase?
- AI contribution: Which prediction, recommendation, generation, or decision changes the customer outcome?
- Economic mechanism: Does AI raise price, increase conversion or retention, lower delivery cost, shorten implementation, or enable a product that was previously impractical?
- Disproof conditions: What result, customer behavior, cost level, or competitor response would make you walk away?
- Investment terms: What valuation, liquidation preference, dilution exposure, lock-up, voting rights, fees, and liquidity constraints apply?
For private-company investments, confirm that the offering is available to you and that you can obtain enough underlying information. Securities rules, disclosure practices, and investor eligibility vary by offering and jurisdiction; the sources cited here do not provide a complete private-placement legal analysis.
2. Verify what the company means by “AI”
Ask management to define AI in the company’s own system, not in marketing language. The National Institute of Standards and Technology (NIST) defines an AI system as “a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments.” Read the definition in the NIST AI Risk Management Framework FAQs and map it to the product.
Obtain an architecture-level explanation of what is built internally and what is bought. Identify foundation models, application programming interfaces, cloud services, data vendors, hardware, fine-tuning, retrieval components, rules engines, and human review. A company can create value by integrating outside models exceptionally well, but that is a different investment proposition from owning a scarce model or dataset. The distinction affects margins, bargaining power, switching costs, and exposure to a supplier’s pricing or policy changes.
Evidence that the claim is real
- A plain-language description of the AI task and where it appears in the customer workflow.
- Pre-deployment evaluation results using representative data and the conditions customers actually face.
- Post-launch performance by customer, use case, geography, or other material segment, with the measurement period stated.
- Human-review, fallback, and escalation behavior when confidence is low or the system fails.
- Monitoring for drift, changing inputs, model updates, latency, error rates, abuse, and unusual output.
- Unit economics showing how inference, storage, labeling, support, and human oversight affect gross profit.
Do not treat a live demonstration as independent validation. Request reproducible evaluation methods, test-set construction, error definitions, and the gap between laboratory results and production outcomes. Ask which results are management-reported and which have been checked by customers, auditors, or another independent party.
Watch for securities-law and credibility risk
In prepared remarks on February 13, 2024, SEC Chair Gary Gensler warned that misleading AI claims can create securities-law concerns. He said: “When disclosing material risks about AI—and a company may face multiple risks, including operational, legal, and competitive—investors benefit from disclosures particularized to the company, not from boilerplate language.” Read the full remarks at the SEC. Compare the company’s specific descriptions with its broad slogans. If management cannot explain the AI component, the measurable outcome, or the limitations, treat the label as an unverified claim rather than an asset.
Rank #2
- Ideal for Gifting
- Ideal for a bookworm
- Compact for travelling
3. Test the customer problem and market opportunity
A large market forecast does not show that this company can win customers profitably. Rebuild the market argument from the bottom up. Separate the number of potential users from the number with a budget, an urgent need, and authority to buy. Check whether the estimate counts the same customer more than once, assumes a new behavior without evidence, or treats total industry spending as revenue available to the company.
Customer evidence
- Interview paying customers and, where possible, former customers. Ask what they replaced, how long deployment took, and whether the product is still used.
- Separate paid production contracts from pilots, trials, proof-of-concept projects, channel announcements, and letters of intent.
- Check renewal, expansion, usage, and implementation data. A high logo count with low usage or no renewals is weak evidence of durable demand.
- Measure the customer’s realized outcome: hours saved, revenue gained, loss avoided, error reduction, cycle-time improvement, or another metric tied to the purchase decision.
- Map buying friction, including security review, privacy requirements, regulatory approvals, integration work, and the need for human sign-off.
Ask for customer concentration by revenue and usage. Dependence on one or two customers can make reported growth fragile, especially when a large contract is a pilot or can be canceled without penalty. Also examine concentration among suppliers: one model provider, cloud platform, data licensor, chip source, or specialist contractor may have leverage over availability and price.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →4. Examine the team, traction, and sales pipeline
Investor-readiness guidance from the UK government identifies market opportunity, team, traction, pipeline evidence, financial realities, and governance as core diligence areas. Its Investor Readiness Essentials is written for a specific program, but these categories are useful for AI companies as well.
Team capability
Assess whether the team has delivered products in the target industry, operated reliable data systems, sold to the stated buyers, and managed security, privacy, and compliance. Technical credentials alone do not demonstrate an ability to price, distribute, support, and renew a product. Look for clear accountability: who owns model quality, data rights, cybersecurity, incident response, finance, and customer commitments?
Traction and pipeline quality
Request a time series rather than a single headline number. Review bookings, recognized revenue, active customers, usage, renewals, expansion, cancellations, sales-cycle length, and the proportion of pipeline that has a defined budget and decision date. Reconcile pipeline claims with staffing, implementation capacity, and historical conversion. A rapidly growing pipeline with no corresponding contracts or delivery resources may represent interest, not demand.
5. Analyze financial resilience without relying on invented benchmarks
AI businesses can show fast top-line growth while consuming substantial cash. Review the income statement, balance sheet, cash-flow statement, cap table, debt, commitments, and management’s financial plan. The SEC’s Glossary explains filing terminology for public-company and small-business disclosures; use the company’s actual filings and records rather than a pitch-deck summary.
Revenue quality
- Separate recurring subscription or usage revenue from one-time services, implementation, hardware, and data-labeling work.
- Reconcile contracted annual value with revenue recognized under the company’s accounting policy.
- Check customer retention, gross retention, net expansion, churn reasons, and the age of receivables where those records are available.
- Identify discounts, credits, minimum commitments, usage caps, and termination rights that could make reported revenue less durable.
Cost structure and AI-specific expenses
- Calculate gross margin after model inference, cloud compute, storage, data licenses, human review, support, and any usage-based supplier charges.
- Identify reserved or minimum cloud and compute commitments, model-provider price changes, and contract termination costs.
- Determine whether a model update, larger context window, higher traffic, or a shift from batch to real-time processing materially changes cost per transaction.
- Separate research and development from customer delivery labor. Services-heavy revenue can conceal the cost of making each deployment work.
Cash needs and financing risk
Compare cash burn with unrestricted cash, committed financing, debt maturities, and the time required to reach the next funding milestone. Stress the plan for slower sales, lower gross margin, a major customer loss, a supplier outage, or a security incident. Examine dilution from employee options, convertible instruments, warrants, and future capital raises. There is no universal revenue, retention, margin, inference-cost, burn, or runway threshold that establishes investment merit for every AI company; the available guidance supports asking these questions, not applying a single target.
6. Demand lifecycle evidence for the AI system
AI performance is a lifecycle issue, not a launch-day score. NIST’s voluntary AI Risk Management Framework organizes risk through design, development, deployment, use, and testing or evaluation. Its trustworthiness considerations include validity and reliability, safety, security and resilience, accountability and transparency, explainability and interpretability, privacy, and fairness with harmful bias managed. The framework is not a certification or a guarantee that a system is safe.
Before deployment
- Define the intended use, prohibited use, users, decision authority, and acceptable failure modes.
- Document data provenance, consent or licensing, labeling quality, representativeness, retention, and access controls.
- Test accuracy and harmful errors on data that reflects production conditions, including edge cases and adversarial inputs.
- Set release criteria, rollback procedures, human-approval requirements, and ownership for unresolved risks.
In production
- Monitor quality, latency, availability, cost, drift, abuse, privacy events, and disparities across relevant groups.
- Keep versioned records of models, prompts, retrieval sources, data changes, evaluations, incidents, and customer-impacting releases.
- Provide users with a way to challenge or correct an output when a person or business decision is affected.
- Test fallback behavior when a model, data source, cloud region, or external API is unavailable.
When it fails
Ask for the incident taxonomy, severity levels, notification rules, root-cause process, customer remediation, and board reporting. A company that reports only average accuracy may not know how often the system fails in the exact cases that matter financially or legally. Look for evidence that lessons from incidents change training data, prompts, controls, contracts, or product scope.
7. Examine dependence on models, data, infrastructure, and people
Determine what the company truly controls. A product may be differentiated by workflow integration, proprietary data rights, distribution, domain expertise, feedback loops, or service quality even when its underlying model is rented. Conversely, a claimed proprietary model may be reproducible by competitors with similar public models and data.
Rank #4
- Model suppliers: Review pricing, rate limits, service levels, indemnities, data-use terms, model-change notices, and termination rights.
- Data: Verify ownership or license scope, geographic restrictions, consent, copyright exposure, freshness, exclusivity, and the ability to continue using the data after a contract ends.
- Compute and hardware: Check capacity reservations, geographic concentration, outage history, replacement lead times, and capital requirements.
- Human operations: Quantify reviewers, domain specialists, implementation staff, and support labor needed to maintain quality at scale.
- Switching costs: Test whether customers can export data, reproduce workflows, or move to another model or vendor without major disruption.
Run a downside scenario in which the primary model supplier doubles prices, restricts a feature, suffers an outage, or becomes a competitor. The company should have a documented migration or fallback plan and enough engineering capacity to execute it.
8. Treat governance, cybersecurity, privacy, and resilience as investment variables
Governance determines whether management notices a problem early and allocates money to fix it. Ask which board or committee oversees AI deployment, who receives risk reports, and how conflicts between growth and safety are resolved.
On December 4, 2025, the SEC Investor Advisory Committee approved a recommendation concerning disclosure of AI’s impact on operations. It is a committee recommendation, not a binding SEC rule. The document is available as a PDF. Use it as a prompt to ask whether oversight is specific, documented, and connected to material risks; do not present it as a new mandatory reporting requirement.
Governance questions
- Which director, committee, or executive owns AI risk, and what expertise do they have?
- How are model, data, privacy, safety, and security changes approved and recorded?
- What metrics reach the board, how often, and what thresholds trigger intervention?
- Who can pause a deployment, and can that person act without a sales or revenue approval?
- How are conflicts involving customer commitments, proprietary data, and model-provider terms escalated?
Cybersecurity and privacy
AI systems expand the attack surface through prompts, model endpoints, training data, retrieval stores, plugins, employee accounts, and third-party suppliers. Review identity and access controls, secrets management, encryption, logging, vulnerability management, penetration testing, backup restoration, and segregation of customer data. Check whether the company can detect prompt injection, data exfiltration, model abuse, and unauthorized training use.
Free tools Windows power users keep installed
One-click scans. No signup required.
The SEC’s CF Disclosure Guidance: Topic No. 2 — Cybersecurity says issuers should consider the likelihood and magnitude of incidents, effects such as data loss or operational disruption, and the adequacy of preventive actions. For a private company, request equivalent evidence even when no public disclosure is required.
Best Value
- It can be a gift option
- Comes with secure packaging
- Helpful in various ways
9. Read public disclosures and private-offering documents critically
Public issuers
Read the latest annual and quarterly filings, material-event reports, earnings materials, risk factors, related-party disclosures, debt documents, and cybersecurity disclosures. Track changes over time: a new AI description, a shift in supplier dependence, a rising concentration, or a newly disclosed limitation can be more informative than a polished product page. Reconcile non-GAAP measures with audited figures and identify adjustments that recur.
Private companies
Request financial statements, cap-table detail, customer and supplier contracts, model and data licenses, security assessments, incident logs, board materials relevant to risk, and the exact subscription or purchase agreement. Confirm liquidation preferences, conversion terms, pro-rata rights, information rights, transfer restrictions, valuation basis, fees, and dilution assumptions. The availability and reliability of these records vary widely; absence of a public filing does not remove the need for evidence.
10. Compare AI companies on the same axes
Build a side-by-side memo for every candidate. Use the same definitions and periods so that one company’s bookings are not compared with another’s recognized revenue or one firm’s pilot count with another’s paying customers.
- Problem and market: urgency, buyer, budget, alternatives, and realistic serviceable market.
- Product and AI differentiation: task, measurable outcome, proprietary assets, evaluation quality, and customer switching cost.
- Traction: paid adoption, usage, retention, expansion, pipeline conversion, and customer concentration.
- Financial resilience: revenue quality, gross-margin drivers, cash burn, runway, debt, commitments, and likely dilution.
- Dependencies: models, cloud, data, hardware, specialist labor, and fallback options.
- Governance and risk: ownership, monitoring, incident response, cybersecurity, privacy, and resilience.
- Terms and liquidity: valuation or price, security rights, preferences, lock-ups, transfer limits, fees, and exit paths.
Mark every item as documented, management-reported, independently corroborated, or unverified. Give greater weight to evidence that would be expensive or embarrassing to fake, such as customer renewals, bank statements, signed contracts, supplier terms, incident records, and reproducible evaluations.
11. A practical diligence sequence
- Define the thesis. Write the customer problem, AI mechanism, economic benefit, horizon, terms, and disproof conditions before contacting management.
- Request a data room. Ask for financial statements, cap table, contracts, customer metrics, supplier agreements, model and data documentation, security material, governance records, and the offering documents that apply to your investment.
- Validate customers. Speak with paying users and former users. Confirm purchase authority, implementation effort, realized outcomes, renewal intent, and alternatives considered.
- Reconstruct the economics. Trace revenue to contracts and cash receipts where possible. Model cost per transaction, human review, support, infrastructure commitments, and the effect of price or usage changes.
- Map the AI stack. Draw the data flow from input to output. Label each component as owned, licensed, or supplied; record rights, limits, fallback options, and change-control obligations.
- Review lifecycle controls. Inspect pre-deployment tests, production monitoring, version history, incident response, rollback, human review, and user appeal mechanisms.
- Test governance. Identify board or committee oversight, decision owners, reporting cadence, privacy and security accountability, and authority to pause a risky release.
- Read disclosures against reality. Compare filings or private documents with customer references, contracts, invoices, technical records, and observed product behavior. Investigate contradictions rather than averaging them away.
- Stress the downside. Model slower sales, customer loss, supplier price increases, model outage, data-license termination, regulatory restrictions, cyber incident, and an emergency financing round.
- Compare and decide. Put candidates on the same scorecard, identify unresolved items, set a maximum price or exposure consistent with your risk tolerance, and document what would trigger a sale or refusal to invest.
12. Red flags that should slow or stop the investment
- The company cannot identify where AI affects the product or cannot distinguish internal technology from third-party services.
- Marketing claims promise autonomy, accuracy, or savings without task-specific tests, error definitions, or production evidence.
- Most traction consists of pilots, unpaid trials, announcements, or one customer whose contract is not durable.
- Gross margin excludes material inference, cloud, data, review, or implementation costs.
- A single model, cloud, data, hardware, or customer relationship is essential and there is no credible fallback.
- Management cannot produce a current cash forecast, cap table, debt schedule, or explanation of dilution.
- Board oversight is informal, incidents are not logged, or nobody has authority to pause deployment.
- Security, privacy, licensing, or data-provenance answers depend on verbal assurances rather than contracts and controls.
- Private-offering terms are unclear, access to information is restricted, or pressure to invest replaces answers to basic questions.
13. Put the AI market in context, not in the valuation
OECD guidance on responsible AI reports global annual AI venture-capital value of about USD 6.4 billion in 2012 and USD 147 billion in 2024, with AI representing 56% of all venture-capital investment value by the third quarter of 2025. The figures appear in the OECD Due Diligence Guidance for Responsible AI.
| Measure | Reported value | Time boundary | What it means for an investor |
|---|---|---|---|
| Global annual AI venture-capital value | About USD 6.4 billion | 2012 | Historical capital-flow context, not a company forecast |
| Global annual AI venture-capital value | USD 147 billion | 2024 | Shows increased funding activity, not expected returns |
| Share of all venture-capital investment value | 56% | By Q3 2025 | Describes market allocation, not valuation or business quality |
Capital flowing into a sector can raise competition, supplier prices, and investor expectations as well as create opportunities. Use these numbers to understand the funding environment, never as a substitute for company-level evidence.
14. Make the decision—and keep testing it
Invest only when you can explain, in plain language, why customers pay, why the company can deliver the outcome at an acceptable cost, what it controls that competitors cannot easily copy, and how it survives a realistic downside scenario. The AI component should improve a measured business result, not merely provide a fashionable description.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteBefore committing money, list every material unknown, assign an owner and deadline for resolving it, and decide which unresolved risks are unacceptable. After investing, monitor the same operating evidence: customer retention and usage, gross-margin drivers, cash and dilution, supplier dependence, model performance, incidents, and governance changes. A thesis that cannot be updated from observable evidence is not a diligence framework; it is a story.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
