How Outcome works.
The data, methodology and reporting logic behind the working report.
The plan goes in at intake. The record follows the work.
The report is one stage of four. People set the plan at intake. Systems record the work while it happens. The record gets reconciled into one file per project. The report reads that file and nothing else, and whatever it learns goes back to intake as a proposed rule.
The loop
Intake
Planned assets, budget, hours, and the first-pass target. The approver list. Which assets are AI-assisted.
No AICapture
Project tool, finance, review tool, calendar, AI tool logs. Recorded as the work happens, not interpreted.
No AIReconcile
Match each asset across tools, compare it to the plan, and score how complete the data is.
AI matches names, flags gapsReport
Numbers render with their source. Diagnoses cite evidence. A producer confirms each one.
AI drafts diagnosesThe plan gets set at intake. Tools record timestamps, invoices, calendar entries, and version histories as the work happens. At close, the system reconciles them; a person still confirms diagnoses and approves any rule proposed for the next project. Less glamorous than a button labeled “AI insights,” which is probably why I trust it more.
Seven sources. Every number traces to one.
Each figure on the report carries its source, and each source carries a coverage score. If the calendar only sees 78% of the approvers, the report says so before anyone can quote the number.
Where each number comes from
| Source | Feeds | Coverage | Trust |
|---|---|---|---|
| Project tool | Planned and actual dates, task hours | 96% of tasks | Counted |
| Finance system | Budget and actual spend | 100% | Counted |
| Review tool | Assets, versions, approvals, wait times | 68 of 68 assets | Counted |
| Calendar | Meetings, context switches, review time | 78% of approvers | Calculated |
| AI logs and asset tags | AI-assisted share, generation counts | 85% tagged | Calculated |
| Marketing analytics | Demo requests against baseline | 4 weeks after launch | Calculated |
| Creative director | Craft, fit to brief, brand consistency | 20 of 68 assets | Judged |
AI helps diagnose. It never writes a number.
I gave AI the messy work it is good at: matching assets with inconsistent names, spotting gaps, and proposing explanations. I did not give it permission to invent a missing hour or a convenient number.
The math comes from recorded sources and rules someone can inspect. AI works in three places around them.
Reconcile
Match messy asset names across tools. Flag gaps and missing data. It never fills a gap with a guess.
Diagnose
Read the reconciled record and propose why. Each hypothesis cites its evidence and states its confidence. A producer confirms or rejects it before it reaches a client.
Propose
Draft the rule that would prevent the problem next time. A manager approves it. It goes to intake as a proposal, never as a silent change.
Three paths through the report
The third path is easy to skip. AI shows up on the report as a cost as well as a helper. Someone tags which assets AI touched before work starts. The review tool times how long each one took to check. The report then estimates what the same work would have cost without AI, and subtracts. That last step is a guess, which is why the result is marked Estimated. In the simulated project AI saved about 96 hours of making and added 112 hours of checking. The 96 is an estimate. The 112 is not. The report says so.
Every number says how far to trust it.
The report holds one figure that everyone will quote and no tool can count: what AI saved. Every number gets a mark, and the same encoding runs through the charts. Solid means counted. Hatched means estimated. A reader learns it once and never opens a legend again.
The four marks
Pulled straight from a tool. Solid mark, solid bar.
Math on counted numbers. Half-filled mark.
A comparison against work that never happened. Dashed mark, hatched bar.
A person scored it against the brief. Diamond, because nothing counted it.
The marks in a chart
Business impact shows the split. The demo count is recorded within this simulated example, and it ran 18% over baseline. What no tool can count is why. There was no control audience and an email push landed in the same window. The figure carries a Calculated mark and the verdict says Unproven, in gray. The count is solid. The claim about it is not.
Verdict first. Number second. Chart only on request.
The summary has one job. Someone glances at it and knows what to argue about. Five decisions get it there.
The verdict is a word.
Yes, and more. Late. Missed. Unproven. A word commits. A gauge can be argued with.
Seven questions, five groups.
Delivery, cost, review, AI, business. The groups are the report's chapters, so the summary doubles as the table of contents.
Red is reserved for misses.
A page with a lot of red reads as a warning, and it should. Wins are green, mixed results stay ink, unproven results go gray.
Percentages sit where the eye ends.
Read the verdict, land on the number. The mark under it says how far to believe it.
The summary fits one screen.
It scales to the window with nothing to scroll. Depth lives one click down, in chapters that scroll.
What the report said about a campaign that never ran.
The simulated campaign shipped 70% more assets for 14% more money. That sounds like a win until you look at the approval trail. The message changed after adaptations began, the approver list doubled, and AI-assisted assets took more checking. The bottleneck moved. It did not disappear.
The message moved after work started.
The framework changed twice in week 3, after social adaptations had begun. It caused 41 of 70 revision rounds.
The approver list doubled.
Three approvers became six, with no turnaround time. Median wait went from 18 hours to 31.
AI outran the people checking it.
AI-assisted assets passed first try 45% of the time against 68% for hand-built. They took about 1.5 times the review hours.
The report ends on a proposed rule, not a conclusion: lock the message before any adaptation starts. A manager approves it, and it goes back to intake.
The numbers are invented. Four things would have to be true.
The campaign never ran. The structure is the work: the questions, the groups, the trust logic, the rule about AI. Real data needs four things from the team using it.
Review timestamps
Approval wait and revision rounds come from the review tool. If the team reviews in email, the report has no wait time to show.
AI tags at intake
Someone marks which assets are AI-assisted before work starts. 85% tagged is a number this system can live with. Below half, the comparison means nothing.
A control audience
The demo count is a recorded figure in this simulated example. The verdict on what caused it stays Unproven until a real campaign holds an audience out.
A person who scores the work
Creative quality is the one line no system counts. A creative director scores a sample against the brief. That stays human on purpose.
Tagging would probably break first. It is never the clever chart. It is the checkbox nobody filled in on Tuesday. If people skip it, the AI comparison falls back to an estimate, and the report says so.