In short: buyers do not pay for a model. They pay for a finished job. Position on the job, prove every AI claim, and tell investors a different story from customers.
Key points
- Write your product truth without naming a model, a vendor or the word AI.
- Most AI product value lives in intake, retrieval, permissions, review, exceptions and hand-off to the system of record.
- Answer 'isn't this just a wrapper?' by showing you own data, distribution or workflow depth, with proof.
- If a human approves every action, call it an assistant, not an agent.
- Investors buy the category; customers buy this task, team, cost and risk this quarter.
Why every pitch sounds the same
The model is shared infrastructure. Whatever it can do today, your competitor's version can do by Tuesday. So pitches that lead with the model converge. "AI-powered." "Built on frontier models." "Intelligent assistant for X." The buyer cannot tell you apart and stops trying.
Positioning starts where the model ends. What does the buyer have at the end that they did not have before? A closed ticket. A reconciled ledger. A contract redlined to their playbook. Name that, in the buyer's words, and the model becomes a detail.
Workflow and outcome over model
Write your one product truth (Part 2) without naming a model, a vendor or the word AI. If you cannot, you have described a feature, not a product.
Then map the workflow. Where does the work start? Who touches it? What breaks today? Where does your product sit, and what does it hand back?
Most of the value in AI products lives in the unglamorous parts. Intake. Retrieval. Permissions. Review. Exceptions. The hand-off to the system of record. That is also where it gets hard to copy. A competitor can match your prompt in an afternoon. Matching your exception handling takes a year of customer calls.
The wrapper objection
You will hear "isn't this just a wrapper?" from investors, buyers and engineers. Do not argue. Agree that the model is rented, then show which of three things you own.
| What you own | What it looks like | How to prove it |
|---|---|---|
| Data | Proprietary, labelled or permissioned data the model cannot see | Show results with and without it |
| Distribution | A channel, install base or integration the model provider does not have | Show where the users come from |
| Workflow depth | The steps around the prompt that make it fit a real job | Show a user doing the job end to end |
One is enough to start. Two is a business. If you own none, the objection stands, and polishing the deck will not fix it.
AI claims in the claims register
AI products attract claims that cannot survive a demo. Put every one in the register from Part 2, with evidence, owner and date.
Accuracy: on what dataset, measured how, by whom, when? Production accuracy is lower than demo accuracy. Say which one you are quoting. If you only have the demo number, say that too.
"Agent": does it act without a human in the loop? For which steps? What happens when it is wrong? If a human approves every action, call it an assistant. Buyers trust you more for it, and the security review goes faster.
"Learns from your data": does the model update, or does retrieval? Say which. They are different products with different risks.
Anything without a test behind it comes out of the copy. Not softened. Out.
Investor story and customer story
Investors buy the category. Why this workflow, why now, why you will own the data or the distribution in three years.
Customers buy Tuesday. This task, this team, this cost, this risk, this quarter.
Mixing them hurts both. Customers hear "platform" and wait for the roadmap. Investors hear a feature list and see a tool. Write both stories, keep them in separate documents, and make sure the same claims register underwrites each one. If a claim is only in the investor deck, ask why.
The checklist
- [ ] Product truth written without "AI", a model name or a vendor name
- [ ] Workflow mapped end to end, with your product's entry and exit points marked
- [ ] Wrapper answer ready: data, distribution or workflow depth, with proof
- [ ] Every accuracy and "agent" claim in the register with a test behind it
- [ ] Investor story and customer story written as separate documents
- [ ] Homepage still makes sense with "AI" deleted
Common questions
How should an AI startup position its product?
Position on the finished job, in the buyer's words, not on the model. Map the workflow end to end and show where your product sits and what it hands back.
How do I answer the 'it's just a wrapper' objection?
Agree the model is rented, then show which of three things you own: proprietary data, distribution, or workflow depth, with proof. One is enough to start; two is a business.
When should an AI product be called an agent?
Only if it acts without a human in the loop, and you should say for which steps and what happens when it is wrong. If a human approves every action, call it an assistant.
Should the investor pitch and customer pitch be the same?
No. Write them as separate documents, since investors buy the category and customers buy a specific task this quarter, but make sure the same claims register underwrites both.