ChatGPT Intelligent UI for founders: make the assumptions visible before trusting the calculator

A slider makes an assumption easier to change. It does not make the assumption true.

ChatGPT’s new Intelligent UI can put interactive tools inside an answer rather than giving you another page of instructions. For a founder comparing offers or planning a small campaign, the useful part is being able to see and change the inputs. The danger is treating a polished calculator as a checked forecast.

We tried a fictional offer calculator on October 8. The response promised editable controls, but no usable calculator appeared. That is part of the result, not a reason to invent the missing output.

What Intelligent UI actually means

OpenAI announced GPT-6 and Intelligent UI on October 7. Its public announcement discussion describes answers combining text, graphics, buttons, forms and tools you can use directly in the conversation. It also says simple questions may still get plain text.

The rollout began October 7 for Plus, Pro, Business and Enterprise, with Free and Go starting October 8, according to that announcement and The Verge’s report. Starting a rollout is not confirmation that every account has the feature. Our signed-out session did not identify a model version.

This is not the same claim as asking a chatbot to write code for a separate app. The proposed tool lives in the response, with controls you can use there.

A reviewer really used the controls

In WIRED’s October 7 hands-on account, the reviewer tried the interface before launch. A San Francisco apartment-affordability calculator had movable sliders for income and expenses, showing rent as a share of take-home pay. An airplane-seat explorer let the reviewer tap seats across four Alaska and Delta aircraft layouts. A slug diagram had buttons for different body parts.

Those are reported interactions, not merely product screenshots. They establish that the reviewer could use controls in those examples. The article does not give an arithmetic audit of the affordability model or verify every aircraft layout.

TechCrunch’s launch report describes an OpenAI demonstration of diagrams, including an airplane wing. That is a witnessed company demo, not another independent calculator test.

Our attempt: a promise, then no calculator

We opened the signed-out ChatGPT web interface on October 8 and submitted this exact prompt. No real people, private data or actual outreach results were used.

Create an inline interactive calculator using ChatGPT Intelligent UI, not code for me to run elsewhere. This is a fictional offer model, not observed outreach data. Inputs: 100 contacts, reply rate 10% of contacts, qualified rate 20% of replies, close rate 25% of qualified replies, $500 per customer. Show all input assumptions and the formulas. Outputs: replies, qualified replies, expected customers (keep fractional values), and gross revenue before costs. Provide working controls for contacts and each assumption. Do not browse or infer realistic rates. I will test changing contacts to 200 and then 0. If inline interactive controls are unavailable in this session, state that and do not pretend a text table is interactive.

The actual response said:

I'll build this as an inline calculator with editable controls for every assumption, including zero-contact handling.

The page then displayed:

Content diagnostic: content_references

After the response finished, the embedded area was empty. We inspected the page and found no calculator inputs or numeric outputs. We therefore could not change the contacts control to 200 or zero. Those interaction tests were not completed.

We do not know whether this was a rendering problem, a session limitation or another fault. It does not prove Intelligent UI fails in all accounts. It does mean this attempt did not deliver the tool we asked for. A sentence promising controls is not evidence that working controls exist.

Check the arithmetic outside the generated tool

Before trying the interface, we calculated the expected values separately with Python arithmetic. The rates below are deliberately fictional assumptions, not LinkedIn benchmarks.

Replies = contacts × 0.10
Qualified replies = replies × 0.20
Expected customers = qualified replies × 0.25
Gross revenue = expected customers × $500

• 100 contacts: 10 replies, 2 qualified replies, 0.5 expected customers, $250 gross revenue.
• 200 contacts: 20 replies, 4 qualified replies, 1 expected customer, $500 gross revenue.
• 0 contacts: 0 replies, 0 qualified replies, 0 expected customers, $0.

These numbers are our independently calculated reference, not ChatGPT output. There is no successful UI result to compare against it from this run.

Half an expected customer is an average implied by the assumptions, not half a person and not a promised sale. Gross revenue excludes delivery costs, software, refunds, taxes and the sender’s time. Neither a slider nor correct multiplication tells you whether anyone will buy.

What to inspect when the tool does render

Start with the displayed defaults. Are the rates percentages or decimals? Does qualification apply to replies, or accidentally to all contacts? Does the tool preserve fractional expected customers, or round before calculating revenue?

Then change only the contacts input. With the fictional model above, doubling it should double every output. Setting it to zero should produce zero throughout, without a stale chart or an undefined result. Finally, change one rate and check its downstream effect against the formula.

That is a proposed check procedure, not a description of controls we used. If the interface gives you a chart but hides the formula, ask for the formula. If it cannot show the inputs or update them correctly, use a spreadsheet you can inspect instead.

For an actual campaign, keep measured counts separate from assumed future rates. Our outreach-log audit works with recorded fictional rows and denominator checks; this exercise checks a hypothetical model. Before choosing people for either, use the buyer-situation worksheet to establish a reason to write.

Checked October 8, 2026: the OpenAI summary and public announcement, WIRED’s hands-on account, TechCrunch and The Verge, plus one signed-out ChatGPT attempt and separate arithmetic. No completed calculator interaction, reply-rate experiment or sales forecast is claimed.