Spindle AI: observability, and a forecast method

A steering slide has a blank cell labelled ROI forecast. The brief says Spindle AI, and someone asks for a number to type before Thursday. Leave the cell blank. Spindle AI is agentic analytics. The announcement said it would join Agentforce work on observability and self-improvement, including ROI forecasting. Forecasting is a method. A method without a baseline, a trace, and an outcome you defined in advance is a costume for a guess. This page will not publish a forecast figure. Anyone who prints one on your behalf is ahead of the evidence.

The line you can date is this one. "Salesforce completed its acquisition of Spindle AI on November 21, 2025." The close is not a result for your pipeline, your queue, or your margin. Those results, if they arrive, will be computed from your traces after you have agreed what counts.

#Observability is a trace you can open

Start with one agent or one process you already run. Write down what you can already see. The request. The actions that ran. The record that changed. The handoff. Whether the work stayed done, using a definition you already use, such as a case that remained closed for the window your service team already argues about. If you cannot see those things, the first quote is to capture them. Analytics on an empty trace is a chart of nothing.

Name the outcome before you look at the chart. Containment that stayed resolved, for a topic you list. A stage change a seller actually wanted. A handoff a colleague marked useful. Pick one. A dashboard of many rates will move, and you will not know which move was the point. One outcome, defined in a sentence, can be wrong in a way you can see. Twelve outcomes can hide.

The baseline is the period you already report, for the same outcome, before the change you are studying. If you have no before, say so. A pilot with no before can still teach you how the trace looks. It cannot support a forecast of return. Write "no baseline" in the cell rather than a borrowed rate. Borrowed rates describe someone else's mix of topics and their willingness to hand work to a person.

Self-improvement needs a fence. The announcement includes it. Your quote decides what may change without a meeting, and what must not. A threshold on a review queue might be adjustable. The words of a policy, a sharing rule, a price, and a legal text must not move because a loop found a pattern. Write the fence. A partner who treats self-improvement as permission to edit production has skipped the change control you already require of people.

#What the quote has to name

The trace source. Where the actions are logged, and who can read the log. The outcome sentence. The baseline period, or an explicit statement that you have none. The fence around self-improvement. The person who reviews a sample of traces each week. The decision the forecast is allowed to inform, if you ever run one: continue a pilot, stop a topic, add a reviewer. A forecast that informs nothing is a slide.

What you will not forecast. Revenue for the company. Headcount removed. A payback month. This page refuses those numbers even as illustrations, because illustrations become the number people remember. If your finance team later builds a model from your baseline, that model is theirs, dated, and sourced. It is not a figure to copy from a proposal.

Licences you already hold for Agentforce and for any analytics product on the contract. The acquisition does not tell this page that a new SKU is included. Check. Leave the cell blank when you are unsure. A blank licence line is a question. A guessed product is a cost you did not choose.

Volume is a band you fill. Traces a day. Topics in the pilot. How large a sample a person can read. The sample size is a staffing fact, not a statistical ornament. If nobody can read twenty traces a week, do not design a review that needs two hundred. Shrink the pilot.

#What cost means before anyone prices it

Cost is capture of the trace, the outcome definition, the baseline if you have to reconstruct it from records you already store, the review hours, the fence, and any later model your own people run. Partner time connects the log and helps you read the first sample. It does not include a right to publish a return. Internal time is the reviewer and the owner of the outcome. Those hours continue after go-live. A quote that ends at the dashboard has priced the picture and left you the judgement.

Out of scope: a number for the steering cell, a change to sharing or pricing driven by the model, a rebuild of the agent, a process-mining programme for the whole company, a customer story with a rate. Each of those can be proposed later, with evidence. They are not the observability quote.

The acceptance test is a week of traces a reviewer can open, each tied to the outcome, with the fence written down. A pass is that the reviewer can say, for a sample, whether the outcome happened. A fail is a chart with no path back to a record. If you cannot open the record from the point on the chart, you cannot audit the forecast that will eventually sit on top of it.

#Forecasting without a published figure

When you do have a baseline and a trace, a forecast is a statement of the form: if this outcome continues at the rate we measured, the next period looks like this, and we will stop believing it if the trace breaks. You still do not copy the sentence onto this page. Your rate is yours. The quote should buy the ability to write that sentence, not the sentence itself.

State the disqualifiers in the method. A window you shortened so the chart improved. A topic you removed from the denominator. A blended average that hides the hard cases. Supervisor hours left in another budget. Any of those makes the figure dishonest even if the arithmetic is neat. Put the disqualifiers next to the blank cell so Thursday's meeting has something better than a guess to discuss.

Self-improvement, if you allow it inside the fence, should be logged like any other change. What moved. Who allowed that class of move. How you reverse it. An improvement you cannot reverse is a release. Treat it as a release. The analytics layer does not get a private path into production.

Compare like with like. A topic you automated is not the baseline for a topic you still refuse. Split them. Otherwise the forecast will congratulate you for removing the easy work from the average. The operations lead already knows this trick. The quote should not reintroduce it in a new tool.

#The hiring test

People search for the best spindle ai Salesforce integration agency. Give the partner one outcome sentence and a sample trace, or the honest admission that you cannot yet export one. Ask where the trace will live. Ask what the model must not change. Ask them to leave the ROI cell blank. Ask what would make a later figure dishonest. Ask which licence they need to see.

A pass describes capture, review, and a fence. They ask for your existing reporting window. They refuse a revenue number. They can say that no baseline means no forecast. They do not invent a customer result to make the method feel proven.

A fail types a figure into the cell. They promise self-improvement will tune the agent without a change log. They blend topics. They call themselves the best agency. The blank cell was the test. They filled it.

Second test: a trace that shows a handoff with an empty note. The partner should count that as a failed outcome if your sentence says the colleague must be able to start work. If they hide it inside a success rate, they are optimising the slide.

#The review that continues

Keep the sample going after the pilot excitement fades. The reviewer reads traces on a cadence you will keep. They look for the fence being nudged. A threshold that moved without a note is an incident. A policy sentence that changed is a bigger one. The hours for this reading were on the cost list. If the list skipped them, add them before you rely on the chart.

When a real baseline exists, let your own analysts write the forecast in the unit you already report. Date it. Attach the outcome sentence. Attach the disqualifiers. That document can go to a steering group. This guide still will not supply the number, and a partner should not paste one into your pack as if it had come from here.

Keep the raw trace for the period the forecast covers. A steering group that can see the chart and cannot open ten underlying records has been given a slogan. Ten is a sample you choose because a person can read it in a sitting. If your volume is large, say how you picked the ten, and pick them the same way next time. A sample that changes its rule each month will manufacture a story. The rule of selection is part of the method, and it belongs next to the blank cell until your analysts are ready to write.

A picture of how a human process actually runs is a different instrument from a trace of an agent. Apromore and the event log is that quote. Use it when the stall is in the work. Use this page when the question is what the agent did, and what it must not change by itself.

Send the outcome sentence and either a sample trace or a clear gap. Ask for the fence before the fee. A partner who can keep the cell blank, and who can point from a chart back to a record, is fit to price the capture. A partner who needs a number to look serious has failed the only test that matters on Thursday.

#The other searches

The project quote is on spindle ai implementation cost. That page does not print a figure. Who bought the company is on what happened to spindle ai. The connector is on integrating spindle ai with Salesforce. Whether to move is on should I migrate from spindle ai. The customer fee is on spindle ai licence cost. Who to hire is on hiring for spindle ai. This page does not rank an agency.