Durable AI Value vs Speculation: What Survives After the Hype
Hype cycles reward stories. Customers reward results. The products that survive a correction are the ones whose value does not depend on enthusiasm for AI itself. This framework gives you five tests to separate the two, whether you are building, buying or investing.
TL;DR
- Durable value shows up as retention, margin and workflow ownership. Speculation shows up as demos, valuations and total-addressable-market slides.
- Five tests: retention, margin, ownership of the workflow, data advantage and switching cost.
- Ask whether customers would pay for the outcome if you removed the word AI.
- Score each project honestly, and invest in the ones that pass at least three tests.
- Revisit the scorecard each quarter, because both competitors and costs move.
Why this question matters now
In October 2026, a restatement of one lab’s run-rate moved chip and cloud stocks by several percentage points in a day. The same month, a leading lab filed a prospectus showing very fast growth alongside enormous commitments. Reasonable people read the same facts as a boom and as a bubble. Because you cannot know which, build things that work in both.
Durable value is a property of your product, not of the market. It is what remains if enthusiasm fades and token prices change.
The five tests
| Test | Durable signal | Speculative signal |
|---|---|---|
| 1. Retention | Cohorts flatten, usage is habitual | Strong signup, rapid drop-off |
| 2. Margin | Positive gross margin per task, improving | Costs ignored or subsidised by credits |
| 3. Workflow ownership | Product is where the job gets done | A side feature people try once |
| 4. Data advantage | Your data or feedback improves results over time | Same model, same prompts as everyone |
| 5. Switching cost | Integrations, history and habits are hard to move | Customers could replace you in an afternoon |
1. Retention
Look at cohorts, not totals. A flattening retention curve suggests a core of users found lasting value. If every cohort decays to zero, the product is entertainment, not a tool.
2. Margin
Calculate the gross margin per task after model, retrieval and hosting costs, and stress-test at higher token prices. A feature that is profitable only while credits last is a subsidy. See AI-native survival.
3. Workflow ownership
Ask where the customer does the work. If your product is the system of record or the place where decisions get made, you are hard to remove. If it is a separate tab used occasionally, you are optional.
4. Data advantage
Does using the product make it better for that customer or for all customers? Feedback loops, domain-specific corrections and proprietary datasets can create advantage a rented model cannot copy. A prompt is not data.
5. Switching cost
Integrations, saved configurations, history and team habits keep customers from leaving. They are not a trap if the product keeps earning its place, and they are a real moat if it does.
Scoring a project
Rate each test from 0 to 2: 0 means absent, 1 partial, 2 clear. A maximum score is 10.
| Project | Retention | Margin | Workflow | Data | Switching | Total |
|---|---|---|---|---|---|---|
| Contract-review assistant inside a legal team’s daily tool | 2 | 1 | 2 | 1 | 2 | 8 |
| Generic chatbot on the marketing site | 0 | 1 | 0 | 0 | 0 | 1 |
| Support triage trained on the company’s own tickets | 2 | 2 | 1 | 2 | 1 | 8 |
The numbers are illustrative, but the exercise forces specifics. Projects scoring below five are experiments: fund them small, set a stop rule and learn.
Ask the label question
A good test: would customers pay for the outcome if you removed the word AI from the page? “Reviews contracts in half the time” is a result. “AI-powered legal platform” is a label. If customers care only about the label, you are selling novelty, and novelty fades.
Use evidence customers can feel
- Time or money saved, measured with customers, not estimated.
- Retention after 8 and 12 weeks.
- Willingness to pay at the real price.
- Renewal reasons in customer words.
Where speculation hides
- Valuations and slides. TAM charts are not evidence of demand.
- Run-rate numbers. They multiply a short period by twelve. See the OpenAI revenue gap.
- Pilots. Most do not reach production. MIT NANDA’s 2025 report found about 95% of organisations saw no measurable return.
- Vendor goodwill. Favourable pricing may not last.
What to do with the scorecard
- Score every AI initiative quarterly.
- Stop or shrink anything scoring under five.
- Double down on one or two projects scoring eight or more.
- Share the scores with your team and investors, since honest scoring builds trust.
Also review pricing, because durable products capture value in how they are packaged, and the sceptic’s case for the risks that apply across the sector.
Using the scorecard with investors
Investors respond well to honest scoring. Show the five tests for your main product with evidence for each, such as cohort curves and margin per task. Point out the weakest test and explain what you are doing about it. A founder who knows where the product is fragile inspires more confidence than one who claims there is no risk.
Your next step
Score your three main AI projects with the five tests in a thirty-minute meeting with product and finance. Share the scores openly and stop or shrink anything below five. One honest hour now saves months of funding a demo.
Bottom line
Durable AI value is the combination of retention, margin, workflow ownership, data advantage and switching cost. Score your projects honestly, strip away the label, and treat anything that fails the tests as a cheap experiment. The products that survive hype cycles are the ones customers still pay for when nobody is talking about AI.
Frequently Asked Questions
What is durable AI value?
An improvement to a customer's job that persists after the novelty fades, at a margin that makes the business sustainable. You can see it in retention, repeat usage and willingness to pay.
How do I tell if a feature is just a demo?
Look at usage after the first week. If people try it once and do not return, or it does not appear in renewal conversations, it is a demo. Durable features tie into daily work.
Is a wrapper around a model always weak?
Not always. A thin wrapper is easy to copy, but a product that owns a workflow, integrates deeply and improves with proprietary data can be defensible even if the model underneath is rented.
How should an investor or founder use this?
As a checklist for diligence and roadmap decisions. A project that fails most tests may still be a good experiment, but it should be funded like one, with small budgets and a clear stop rule.
