Steve Pritchard recently shared something that caught my attention: "Started scoping a commission calculation AI agent for a mid-size sales org this week. 47 reps. 12 different comp plans." Then he added the detail that should make any sales leader uneasy: one person calculates commissions manually, it takes three full days every month, and the accuracy check is basically, "If nobody complains, we assume it's right."

That last line is the point. As Steve put it, "That's not a system. That's a prayer." I want to expand on why this situation is more common than we admit, what it really costs (beyond three days of work), and what a practical fix looks like without turning comp into a science project.

The real problem is not math, it is ambiguity

Commission calculation is rarely hard because of arithmetic. It is hard because comp plans are living documents with exceptions.

Steve’s example hits all the classic complexity traps:

Accelerators after quota
Tiered rates that change quarterly
SPIFs that shift monthly
Legacy plans that "nobody touches" because they were agreed years ago

Individually, each rule seems reasonable. Collectively, they produce a system where two people can follow the same plan and still disagree on the answer because the definitions are unclear.

"Forty-seven people's pay... And the quality check is silence." - Steve Pritchard

When the verification method is "no news is good news," what you really have is unmeasured error. Some reps will not notice. Some will notice and not complain. Some will complain only when it is materially wrong. Silence does not equal accuracy.

Why manual commission processing becomes a business risk

Three days of manual work sounds like an efficiency problem. It is also a control problem.

1) Trust and retention

Salespeople can handle aggressive targets. What they cannot handle is uncertainty about pay. If commission statements feel arbitrary, reps start discounting your plan psychologically. They assume they will be shorted, so they push for higher base, bigger guarantees, or they leave.

2) Shadow accounting and manager overhead

When reps do not trust statements, they build their own trackers. Managers spend time adjudicating disputes, pulling reports, and mediating conversations that should not exist. Even if disputes are rare, the background noise erodes time and focus.

3) Financial reporting and audit exposure

If commissions are material, errors affect accruals and forecasting. A manual process that depends on a single person becomes a key person risk, and it is difficult to defend in an audit because you cannot prove consistent application of rules.

4) Incentives drift from intent

Comp is supposed to shape behavior. If the calculation is inconsistent, the plan does not reliably reward the actions you care about. That means you are paying for outcomes you did not intend, and underpaying the behaviors you do want.

The hidden cost: comp plan sprawl

Steve’s "12 different comp plans" detail is the other big signal. Comp plan sprawl happens for understandable reasons:

You create a new plan for a new role, segment, or product
You grandfather exceptions for senior reps or managers
You add SPIFs to fix short-term issues
You patch edge cases instead of simplifying

Over time, the comp plan becomes a set of overlapping policies rather than a coherent system. The manual calculator becomes the only person who can translate policy into payouts. That is not operations, it is institutional memory.

What a good commission system must do (before you add AI)

Steve mentioned scoping an AI agent. That is exciting, but the best results come when you treat automation as the final layer, not the first.

Here is the minimum viable set of capabilities any commission process should have.

1) A single source of truth for inputs

Define where each input comes from and lock it down:

Bookings, revenue, or collections
Product and SKU mapping
Territory and account assignment
Quota and ramp schedules
Eligibility rules (start date, role, plan assignment)

If inputs are coming from spreadsheets, emails, and CRM exports, the calculation will always be fragile.

2) A rules engine that is readable by humans

You need explicit definitions for each rule:

When does an accelerator start? At 100% of quota? At 100% of quarterly quota or monthly? Are returns netted?
How do tier boundaries work when the quarter changes?
How are SPIFs applied when a deal spans periods?

If the "rule" is a person’s interpretation, it will vary.

3) An audit trail and version control

Comp plans change. Your system must record:

Which version of the plan was applied
When it changed
Who approved it
Why it changed

Without that, you cannot answer simple questions like "Why was my rate different last month?" or "Why did two reps with similar performance get different payouts?"

4) Verification that does not depend on complaints

Steve’s story highlights a missing control. Verification can be lightweight and still effective:

Automated reconciliation checks (totals by rep, by plan, by product)
Variance flags (payout changes over a threshold month-to-month)
Spot checks with a sampling plan
Parallel runs after plan changes

A quality check based on silence is a liability, not a process.

Where an AI agent actually helps (and where it does not)

An AI agent can be valuable, but it should be deployed in places where it improves speed and clarity, not in places where it invents logic.

Strong use cases

Parsing plan documents into structured rules (with human approval)
Detecting anomalies ("this payout is 3.2x higher than typical for this rep")
Answering rep questions with citations ("your accelerator applied because you crossed 100% of quota on Feb 18")
Drafting clear commission statements and explanations

Risky use cases

Letting the agent "decide" how to interpret ambiguous plan language
Auto-correcting payouts without approvals
Producing numbers without showing inputs, formulas, and traceability

In other words, use AI for interpretation support and communication, but keep calculation deterministic and auditable.

A practical path from prayer to process

If you recognize Steve’s scenario in your own org, a full rebuild is not the only option. A phased approach works.

Phase 1: Reduce ambiguity

Inventory every plan and exception
Kill or sunset SPIFs that no longer serve a purpose
Rewrite rules into a standard template with examples

Phase 2: Standardize data

Define the canonical fields and their owners
Fix mappings (products, segments, crediting rules)
Create a stable commission-ready dataset each period

Phase 3: Automate calculation with controls

Implement a rules-based model (commercial ICM tool or well-governed internal system)
Add checks, variance thresholds, and approvals
Produce rep-facing explanations, not just numbers

Phase 4: Add AI where it strengthens the experience

Conversational Q&A with policy citations
Exception intake and routing
Faster plan change impact analysis

The leadership takeaway

Steve Pritchard’s post is a reminder that commissions are not back office admin. They are the contract of trust between the company and the sales team. If that contract is calculated manually, verified by silence, and held together by legacy exceptions, you are not just risking errors. You are teaching your sales org that the system is arbitrary.

Replacing "prayer" with process does not require perfection. It requires clarity, traceability, and verification. Then automation, and only then AI.

This blog post expands on a viral LinkedIn post by Steve Pritchard. View the original LinkedIn post →