AI Agents
AI agents for sales commissions: what Spiff's AI can do, and what it can't
A rep opens their commission statement, doesn't trust it, and opens the spreadsheet they keep on the side. An AI agent is the right tool for 'why was I paid this' and the wrong tool for computing the payout. That split is the whole design.
Every payday, a rep opens their commission statement, doesn’t trust it, and opens the spreadsheet they keep on the side.
That private spreadsheet has a name in comp circles: shadow accounting. It exists because the statement doesn’t show its work, and it signals a trust problem, not a calculation problem.
Salesforce’s own research puts numbers on it. In the migration story for its internal rollout, it cites that 74% of reps want more transparency in how their pay is calculated and 47% say it doesn’t always seem fair.
An AI agent is the right tool for the first number and the wrong tool for the payout itself, and that split is the entire design. A deterministic engine calculates the commission and owns it as the system of record. The agent reads that result, explains it, and triages the dispute. It never recomputes the number.
Where Spiff sits in the stack now
Salesforce Spiff is the company’s incentive compensation management product, acquired in 2024 and relaunched as a native Sales Cloud add-on rather than a separate system you sync to. That native posture matters more than it sounds.
Opportunity, product, quota, and user data feed the calculation from the same org they live in. There’s no nightly connector to drift out of sync, and no mapping layer to drop a field that changes a payout.
It lists at $75 per user per month as an add-on, and it’s bundled into the top Sales edition, the tier the new Core, Advanced, and Max lineup calls Max. What you get is a small set of parts worth naming because each one maps to a specific trust failure:
- Commission Designer, a low-code plan builder that lets comp admins model the impact of a plan change before they ship it, instead of discovering it on the next statement.
- Commission Estimator, which shows a rep their projected commission on an opportunity or quote while they’re still working it, not a month later.
- Real-time rep statements on web and mobile.
- Tracing and audit trails, the drill-down from a payout to the deals, rates, and accelerators that produced it.
The tracing is the part that ends shadow accounting. When a rep can click a number and see the deals and rates behind it, most disputes resolve before they reach anyone’s desk.
The agent’s real job is the “why,” not the “how much”
The conversational layer Salesforce ships today is the Spiff Assistant, described as a conversational AI assistant that answers questions about plans, rules, and calculations, grounded in your plan logic. Put plainly, it’s the natural-language front door to the tracing data a rep would otherwise have to dig for.
An agent should have exactly that job here. “Why was my Q3 commission lower than Q2?” is a question with a deterministic answer sitting in the engine: a deal slipped a quarter, an accelerator didn’t trigger, a split changed.
The agent’s work is to find that answer, read it back in plain language, and link to the statement line that proves it. The calculation already happened; the agent explains a fact, it doesn’t produce one.
Salesforce has said richer questions are coming, the kind that move from “why was I paid X” to “where should I focus to maximize my payout.” Treat that as roadmap, not as something to architect around today. The capability that’s live and worth deploying now is the explanation of a payout that already exists, and it’s enough to retire a lot of spreadsheets on its own.
Why the agent must never compute the payout
The temptation, once you have a capable language model in the conversation, is to let it do the arithmetic. Resist it, and the reason is not caution for its own sake.
A commission calculation has to be reproducible and auditable. The same inputs must produce the same payout every time, and every payout has to trace to its source data and logic for revenue-recognition and audit purposes under rules like ASC 606. A language model is probabilistic by construction. It can return a slightly different answer to the same prompt, and it can’t hand an auditor a deterministic lineage. Those two properties are disqualifying for a system of record.
Salesforce makes this argument itself in its guidance against building comp tooling on general-purpose AI: general-purpose AI tools are not designed to serve as compensation systems of record, and putting the core calculation in them creates liability, governance, and auditability problems. The principle it lands on is the one to carry into any design. Use AI within a governed, auditable compensation system, on top of the calculation, never as the calculation.
The same discipline keeps any customer-facing agent trustworthy. An agent that grounds on a semantic layer instead of inventing your metrics runs the commission-engine pattern in another domain: the number comes from the authoritative source, and the model fetches and phrases it rather than estimating it. A commission agent that generates a figure is the payroll version of an agent that quotes a price it made up.
The number is only as good as the data under it
Because the calculation runs natively on your Salesforce objects, the quality of those objects is the quality of the payout. An opportunity with the wrong close date, a missing product line, a stale territory assignment on a user record: each one produces a wrong commission that the engine will calculate perfectly and the agent will explain confidently.
Every comp automation project rests on that unglamorous prerequisite. Before an agent can credibly answer “why was I paid this,” the data feeding the calculation has to be clean enough that the true answer isn’t “because the opportunity was mis-staged.” Plan to spend real time on the fields that feed the calc, and treat the agent’s first weeks as a data-quality audit that happens to speak English.
Where a human still decides
Drawing the automation line here is the same exercise as in any other human-in-the-loop agent design: the agent handles the high-volume, rule-bound, explainable work, and the consequential judgments stay with people.
Plan design stays human. The Commission Designer models the impact of a change, and a comp leader decides whether to make it. Dispute adjudication stays human when the rep and the statement disagree about what should have happened, as opposed to the rep not understanding what did. Clawbacks, exceptions, and the sign-off an auditor relies on are decisions, not lookups, and an agent that drafts the explanation for them is useful precisely because it doesn’t make the call.
The agent’s territory is the inquiry that has a correct answer in the system and just needs retrieving and explaining. It covers most of the volume, which is the point, and none of the liability, which is the design.
Where this is going, and how to get ready
Salesforce has signaled that commission data will become reachable wherever sellers already work, in Slack, in the Lightning UI, and through Model Context Protocol and a headless approach, so the answer can come to the rep in the tool they’re in rather than a portal they have to open. That direction is credible and mostly still ahead of us, so build for it without betting on dates.
The way to get ready costs nothing speculative. Get the comp plans into the engine so the calculation is deterministic and traceable. Clean the opportunity, product, and user data that feeds it. Turn on the assistant over that foundation so reps can ask why, in plain language, and get an answer that points at the statement. When the headless and Slack surfaces land, an agent built on a clean engine inherits them. An agent built on a spreadsheet inherits nothing.
The measure of success isn’t how much the agent automates. It’s how few shadow spreadsheets survive the quarter.
Understanding the basics
What is Salesforce Spiff?
Salesforce Spiff is Salesforce’s incentive compensation management (ICM) product, acquired in 2024 and offered as a native Sales Cloud add-on. It automates commission calculation using the opportunity, product, quota, and user data already in the org, and includes a plan builder (Commission Designer), a rep-facing estimator, real-time statements, and calculation tracing. It lists at $75 per user per month and is bundled into the top Sales edition.
Can an AI agent calculate sales commissions?
It shouldn’t. Commission calculations must be reproducible and auditable, and a probabilistic language model can’t guarantee the same inputs always yield the same payout or produce the deterministic audit trail that revenue recognition requires. The sound pattern is a deterministic engine that computes and owns the payout, with an AI agent layered on top to explain and answer questions about it.
What does the Spiff Assistant do?
The Spiff Assistant is a conversational AI assistant that answers questions about comp plans, rules, and calculations, grounded in your plan logic. Its practical value is turning the tracing data behind a payout into a plain-language answer, so a rep can ask why a commission came out the way it did instead of digging through a statement or filing a dispute.
How do you stop a commission agent from giving a wrong number?
Keep the calculation in a deterministic engine and let the agent only read and explain it, never generate it. Then make sure the data feeding the engine is clean, because the agent will confidently explain a wrong payout that a mis-staged opportunity or stale territory produced. The agent’s honesty depends on the engine’s determinism and the data’s quality, not on the model.
Deciding where an agent belongs in your comp process, and drawing the line between the deterministic engine that owns the payout and the agent that explains it? Talk to us. Getting the calculation trustworthy first, and the agent grounded on it second, is exactly the work we do.
