ChatGPT
Claude
Google AI
Copilot
A revenue lead pins a chart next to the opportunity pipeline and draws an arrow between them. The arrow has no field name. The chart has a figure. Nobody has said which opportunity receives it. The room treats the arrow as an integration. It is a poster. An integration names the system, the object, and the failure when the call does not land. The system is the analysis. The object is the field someone named. The failure is a number written onto the wrong opportunity.
Spindle AI is the analysis work this acquisition brought in. Salesforce completed the acquisition on 21 November 2025. An analysis can be useful while it stays a reading. It becomes a write-back only when a person names the field, the opportunity, and the rule that decides the number is allowed to replace what is already there. This page will not publish a forecast number. A figure on this page would be an invention. The only sentence this page quotes from the announcement is the completion line. "Salesforce completed its acquisition of Spindle AI on November 21, 2025."
The system is the analysis
The system reads data you already trust and produces a number or a comment a person can argue with. It might sit beside a forecast process you already run. It does not replace that process by being present. Write the question the analysis answers. "What changed in this opportunity stage history" is a question. "Give us the number" is not. A question without a source will be answered from whatever was nearest.
Name the inputs. Opportunity history. A finance extract you already file. A period. Leave out any source you would not show an auditor. The analysis is only as shareable as its inputs. If the inputs include a restricted sheet, the output is restricted too. Do not paste the output onto a record a wider audience can open.
This page does not describe a current screen of an analytics product. Screens move. The contract is the question, the inputs, and the field you may write. Spindle AI and the forecast method is where you keep the figure blank and scope the reading. This page is the moment someone asks to write the reading back. What happened to Spindle AI dates the close. It does not give you a number to type.
The object is the field you named
An analysis is not a write-back until someone names the field. The sentence is the whole gate. Until the field has a name your sellers already see, the analysis stays in the reading. A comment in a chat, a slide, a file. Those are not the opportunity.
The field has to exist before the project. Do not create a field in the same afternoon as the first write so the demo has somewhere to go. A new field needs a label sellers understand, a help text, and a decision about who may edit it by hand. If sellers can edit it, the write-back will fight them. If they cannot, you have taken a field away. Write that decision. It is part of the object.
Name the opportunity by id, not by a fuzzy match on name. Account names repeat. Opportunity names repeat. A match on name is how the number lands on the wrong record, which is the failure. The id is the contract. If the analysis does not know the id, it is not ready to write. Show the reading to a person and let them choose the id. Choosing is a control, not a delay.
Also name what must not be written. Amount, close date, and stage are fields with owners in most orgs. An analysis that overwrites amount because the field was nearby has changed the forecast a manager will submit. If those fields are out of bounds, say so in the diagram. Out of bounds is a feature of the integration.
The failure is the wrong opportunity
The number is plausible. The write succeeds. The opportunity that changed is not the one the seller is working. A sibling deal, an old renewal, a test record with a similar name. The seller's real opportunity is untouched. The wrong one now carries a figure someone will report. That is the failure: a number written onto the wrong opportunity.
The twin is the right opportunity and the wrong field. The reading was a risk note. The write put it in amount. Or the reading was a range and the write stored one end of the range as fact. A range is not a field until you say which end, or until you store the range in a text field labelled as a range. Do not collapse it.
The third shape is a correct write the seller then cannot explain. They did not ask for it. Their manager sees a change in history with an integration user as the editor. Trust drops, and the next reading is ignored. A write-back without a named human sponsor on that opportunity is how you get this shape. The sponsor does not have to click every row. They have to own the rule.
No forecast number belongs in the test data except fixtures your team invented and labelled as fixtures. Do not copy a figure from a keynote. Do not invent a target to make the screen look serious. The test checks identity and field, not whether the number is impressive.
What the diagram has to show
System: the analysis, with its question and its inputs. Object: the field on the opportunity, addressed by id. Failure: a number on the wrong opportunity, or on the wrong field. You may show the human sponsor. You may not show a forecast figure, a connector name, a price, a cap, or a screen. The source does not give them.
| Name on the diagram | What you write | A poster instead |
|---|---|---|
| System | The question and the inputs | A chart with no source |
| Object | The field name and the opportunity id | An arrow into pipeline |
| Failure | The number lands on the wrong opportunity | A green write receipt |
| Gate | No write until the field is named. Amount and stage called out if forbidden | A nightly overwrite |
| Out of diagram | A forecast number, a connector, a screen, a price | A logo on the chart |
Say the field name out loud. If the room cannot, you are still looking at an analysis. Keep it that way. An early write is harder to undo than a late one, because history will contain the wrong figure.
What you refuse to invent
A forecast number is the thing a steering group will request. Refuse. The announcement this page cites does not contain one. Your finance lead may have a method. That method can live on the implementation page as a blank to fill from your books. It cannot be filled here.
A connector name would make the arrow look like a product. The arrow is a field update you have permitted. How the update is performed is a build choice after the field exists. Do not coin an API name. Your platform owner can point at the update path they already audit.
A current screen of a model, a notebook, or a dashboard is a poster. The seller never has to see it. They have to see the field, or they have to see nothing, if you have not named the field yet. Nothing is an acceptable landing. It means you kept the gate.
No price and no cap. How many opportunities a pilot may touch is a count you set, small on purpose. Set it in the brief as these ids. A cap attributed to the product would be fiction.
How you test that the call landed
Pick three opportunity ids. One is the intended target. One has a similar name and must not change. One is closed and must not change. Run the analysis. If the field is not named, confirm no opportunity changed. That is a pass. If the field is named, confirm only the intended id changed, and only that field. Read the history. The editor should be the user you chose, not a surprise administrator.
Then run it twice. The second run should be boring. If the number is the same, the field should not churn history without a reason you wrote down. Needless history is how sellers lose the plot of who changed amount. If amount is forbidden, assert it is unchanged.
Ask the seller who owns the intended opportunity whether the field means what they think. If they read it as a commitment and you stored a suggestion, you named the wrong field. Clear it. Relabel. Do not explain over the top of a bad label.
Widen the id list only after three runs stay on the right records. A model that is usually right about identity is not ready. Identity is exact or it is the failure. Keep readings that are not ready in the file the analyst already uses. Copying them onto opportunities because the quarter is loud is how the wrong deal gets the number.
A partner who arrives with a sample forecast has invented the figure the source withheld. Ask for the field name and the three ids instead. The implementation page can hold the method. This page holds the gate.
A reading that stays a reading
Some analyses should never write. A weekly note for a manager, a comparison across regions, a question you have not finished defining. Mark those as readings in the diagram, with no field. The absence of a field is the control. A later reader who wants a write-back has to add the field in a change you can see. They do not get to point at an old chart and claim the integration already included it.
Store the reading where the audience is the same as the audience for the inputs. If finance can see the extract and sellers cannot, the reading stays with finance. Putting it on the opportunity publishes it to everyone who can open the deal. That widening is a data decision. Make it on purpose. An accidental widening is how a restricted input becomes a number on a record the whole forecast call can see.
When you do write, write one field for one purpose. A second field in the same change doubles the ways to hit the wrong opportunity. Finish the first field. Watch a forecast cycle. Then talk about the second. The cycle will show you whether sellers trusted the field or overwrote it. Their behaviour is the landing report. A dashboard of successful updates will not show the overwrite. The history will.
