Two AIs. Two very different challenges. Thirty days. Measurable outcomes. Alex is being tested on real sales and marketing; Morgan is being tested on recorded market forecasts and virtual trading decisions.
The rule is simple: we record the decisions before the outcome is known, and we publish the failures as well as the successes.
See the live results Experiment journal
AI demonstrations are easy to make impressive. These experiments are deliberately harder. Each AI has a defined objective, a scoreboard and an outcome that cannot be explained away afterwards.
Alex is the AI Marketing & Sales Manager for Innocent Soap. The challenge is to sell 100 paid dog-shampoo bars in 30 days using real marketing, prospecting and sales activity.
Current: Day 11/30 · 0/100 bars · £0.00 verified revenue.
Alex resultsMorgan starts with a £10,000 virtual portfolio. Forecasts and trading decisions are recorded in advance so they can later be scored against what actually happened.
Current: Day 4/30 · £10,000 · 0.00% return.
Morgan resultsThe objective: sell 100 paid Innocent Soap bars in 30 days and separate genuine commercial results from AI activity that merely looks productive.
Alex has completed 61 initial trade approaches and processed 29 follow-ups. One retailer has requested wholesale information, but no paid order has followed. A separate retailer declined, citing its previous experience selling solid shampoo. The score remains 0/100 paid bars, £0.00 verified revenue and 0 AI-attributed bars.
Two trade email routes and one media route have confirmed delivery failures. Acknowledgements, enquiries, unsuccessful deliveries and Robin’s test purchase are excluded from the sales total.
Emerging finding: AI can operate a traceable prospecting pipeline, but activity and buying interest have not yet translated into paid demand. The next stage will examine targeting and conversion rather than email volume alone.
The objective is not to force a trade every day. Morgan must analyse the selected markets, record forward-looking calls and protect the integrity of the £10,000 virtual portfolio.
Morgan analysed eight commodity and currency markets and froze eight predictions for later scoring. Seven calls were bearish; USD/JPY was the sole bullish call.
| Market | Call | Confidence |
|---|---|---|
| Gold — XAU/USD | Bearish ↓ | 62% |
| Silver — XAG/USD | Bearish ↓ | 60% |
| Brent Crude | Bearish ↓ | 58% |
| WTI Crude | Bearish ↓ | 59% |
| Copper | Bearish ↓ | 57% |
| GBP/USD | Bearish ↓ | 64% |
| EUR/USD | Bearish ↓ | 63% |
| USD/JPY | Bullish ↑ | 61% |
Why no trade? The forecasts were valid for the prediction test, but the experiment's frozen cost and position-sizing rules were not yet complete. Rather than improvise those rules after seeing the market, Morgan withheld the trade and preserved the integrity of the experiment.
Portfolio at close: £10,000 cash. Open positions: 0. Realised P/L: £0. Unrealised P/L: £0. Trading costs: £0. Maximum drawdown: 0%.
MORGAN — AI MARKET FORECASTER
29 September 2026
Morgan completed the third trading session of the 30-day AI Market Forecaster experiment.
The AI analysed eight global commodity and currency markets: gold, silver, Brent crude, WTI crude, copper, GBP/USD, EUR/USD and USD/JPY.
Directional forecasts were issued before subsequent market movements were known and have now been frozen as part of the experiment's permanent record.
| Market | Forecast | Probability |
|---|---|---|
| Gold (XAU/USD) | Bearish | 64% |
| Silver (XAG/USD) | Bearish | 62% |
| Brent Crude | Bullish | 63% |
| WTI Crude | Bullish | 60% |
| Copper | Neutral / Bullish | 56% upside |
| GBP/USD | Bearish | 61% |
| EUR/USD | Bearish | 63% |
| USD/JPY | Neutral / Bearish | 57% downside |
For the third consecutive session, Morgan made forecasts but did not place a virtual trade.
An audit of the original Version 1.0 rules identified an important flaw in the experimental design.
Although the experiment established a £10,000 starting portfolio, maximum risk of 1% per trade, maximum aggregate risk of 3% and a maximum of four simultaneous positions, the detailed transaction-cost schedule, exact position-sizing convention and mechanical trend-following benchmark were not completely specified before the experiment began.
Introducing those rules now, after observing several days of market behaviour, could introduce hindsight bias.
Morgan therefore declined to retrospectively create the missing rules and continued to record forecasts without risking virtual capital.
Subsequent market movements again demonstrated why a probability forecast should not be confused with certainty.
Gold recovered during Tuesday's session despite Morgan's bearish forecast, while Brent crude moved lower despite Morgan's bullish forecast.
These outcomes remain part of the permanent record. Forecasts will not be rewritten or reinterpreted because the subsequent result is now known.
Day 3 is exposing an important distinction that was not obvious when the experiment began:
Can AI forecast market direction?
And can AI turn those forecasts into a disciplined, executable trading system?
They are not necessarily the same problem.
Morgan currently has a portfolio return of 0.00%. This is neither a profit nor a loss and should not be interpreted as evidence of successful trading.
Morgan's forecasting performance will therefore continue to be recorded independently from portfolio performance.
Nothing in this historical log will be rewritten because we subsequently know what happened.
Mistakes, incorrect forecasts, methodological problems and human intervention are all retained as part of the experiment.
This is an experiment using virtual funds, not investment advice. No real capital is being traded by Morgan.
Not whether AI can write convincing copy about itself. We are testing whether currently available AI can perform useful work when success and failure can be counted.
Paid bars, verified revenue, portfolio value, forecast accuracy, wins, losses and drawdown.
Where does the AI work independently, and where does it still need a human to approve, correct or complete an action?
Bad forecasts, bounced emails, zero-sales days and mistakes stay in the record. Nothing gets quietly deleted because it did not work.
Dated updates from both experiments. This is the audit trail behind the headline numbers.
Morgan recorded directional calls across gold, silver, Brent, WTI, copper, GBP/USD, EUR/USD and USD/JPY. The portfolio remained entirely in cash because the frozen trading-cost and position-sizing rules were not yet complete. The predictions remain recorded and will still be scored.
The campaign produced genuine B2B interest and questions about the range and minimum order quantity. The scoreboard did not move because enquiries are not counted as sales. Only verified paid bars count towards the 100-bar target.
Alex began researching UK groomers and independent pet retailers, preparing outreach and establishing the campaign scoreboard. The starting position was 0/100 paid bars and £0.00 verified revenue.
The final report will compare results, accuracy, costs, human intervention, mistakes, time saved and the practical limitations of using AI for real commercial work.
The useful question is not “Can AI look impressive?” It is “What happens when we actually try it?”
For a fuller record of both live experiments, download the Day 11 interim report. It brings together Alex’s verified sales activity and results, Morgan’s market forecasts and virtual portfolio decisions, the controls used in each experiment, and the findings emerging so far.
The report is deliberately an interim record rather than a success story: useful capabilities, mistakes, resistance, methodological limitations and measurable outcomes are all retained.
Download Day 11 Interim Report (PDF)