Reading a Profit and Loss

The idea in one minute#

A profit and loss statement is a short list of numbers in a fixed order, and once you know what each line means you can read your own company’s in about ninety seconds. The reason it is worth learning is not fluency for its own sake: it is that product decisions land on specific lines, and knowing which one tells you who will care and what argument will work. A feature that moves gross margin is a different conversation from one that moves sales efficiency, even when both are worth the same amount, because they are owned by different people and judged on different timescales.

A picture#

flowchart TB
  REV[":i-banknote: <b>Revenue</b>"] --> COGS[":i-receipt: <b>Cost of revenue</b><br/><small>hosting, support, payments, inference</small>"]
  COGS --> GP[(":i-percent: <b>Gross profit</b><br/><small>and the margin percentage</small>")]
  GP --> SM[":i-megaphone: <b>Sales and marketing</b>"]
  GP --> RD[":i-wrench: <b>Research and development</b>"]
  GP --> GA[":i-building-2: <b>General and administrative</b>"]
  SM --> OP[(":i-coins: <b>Operating profit or loss</b>")]
  RD --> OP
  GA --> OP
  PM[":i-target: <b>Product moves these two most</b>"] -.-> COGS
  PM -.-> REV
  class REV compute
  class COGS,SM,RD,GA queue
  class GP,OP memory
  class PM neutral

Revenue minus cost of revenue is gross profit. Three expense categories come out of that. Product has the most direct effect on the top two boxes.

How it really works#

The lines, and what belongs in each#

The line that causes most confusion is cost of revenue. The test is whether the cost scales with customers served. Hosting does, so it belongs there. An engineer building a new feature does not, so they belong in research and development. Inference almost always does, and is the line most often put in the wrong place.

Why gross margin is the number that gets quoted#

Because it describes the shape of the business rather than its current size. A company at eighty percent gross margin can spend a lot to acquire customers and still work. One at thirty percent cannot, no matter how fast it grows, which is why hardware and services businesses are valued differently from software ones.

When a model goes into the product and inference lands in cost of revenue, gross margin falls. That is the single most consequential financial effect of AI on software companies, and it is treated properly in AI Cost and Margin.

Which line your work lands on#

• A feature that reduces support contacts moves cost of revenue.
• A feature that improves conversion moves revenue and makes sales and marketing more efficient.
• A feature that lets you charge more moves revenue directly, with no cost increase, which is the best shape available.
• Reducing inference cost moves cost of revenue, and is worth more per dollar than reducing an engineering cost, because it scales with customers.
• Engineering productivity work moves research and development, which is real but rarely what the board is asking about.

Naming the line before making the argument is what turns a product proposal into a business one.

Recognition, briefly#

Revenue is recognised as the service is delivered, not when cash arrives. An annual contract paid up front is recognised across twelve months. This matters to product in one practical way: a pricing or packaging change can alter when revenue appears without altering how much there is, and finance will care about that even when the total is identical.

Run the numbers#

A small company’s statement, with a product change applied, to see which lines move.

// pandl.go: where a product change lands.
package main

import "fmt"

type pandl struct {
	revenue, hosting, support, inference, payments float64
	salesMarketing, randD, admin                   float64
}

func (p pandl) cogs() float64  { return p.hosting + p.support + p.inference + p.payments }
func (p pandl) gross() float64 { return p.revenue - p.cogs() }
func (p pandl) opex() float64  { return p.salesMarketing + p.randD + p.admin }
func (p pandl) op() float64    { return p.gross() - p.opex() }

func show(label string, p pandl) {
	fmt.Printf("%-26s %10.0f %10.0f %8.0f%% %10.0f\n",
		label, p.revenue, p.gross(), 100*p.gross()/p.revenue, p.op())
}

func main() {
	base := pandl{
		revenue: 12_000_000, hosting: 900_000, support: 1_100_000,
		inference: 260_000, payments: 340_000,
		salesMarketing: 5_400_000, randD: 3_800_000, admin: 1_300_000,
	}
	deflect := base
	deflect.support -= 420_000 // a feature that removes a class of ticket

	aiFeature := base
	aiFeature.inference += 1_150_000 // a model in the main workflow
	aiFeature.revenue += 900_000     // and the upsell it enables

	pricing := base
	pricing.revenue += 840_000 // a packaging change, no new cost

	fmt.Printf("%-26s %10s %10s %9s %10s\n", "", "revenue", "gross", "margin", "operating")
	show("baseline", base)
	show("support deflection", deflect)
	show("AI feature with upsell", aiFeature)
	show("packaging change", pricing)
	fmt.Println("\nThe packaging change adds less revenue than the AI feature and more operating profit,")
	fmt.Println("because it arrives with no cost of revenue attached.")
}

Read the revenue column and the operating column together. The AI feature adds the most revenue of the three and is the only one that makes the company worse off, because nine hundred thousand of upsell arrives with one and a bit million of inference attached. It also costs seven points of gross margin, which is the number a board will ask about.

Meanwhile the packaging change adds slightly less revenue and nearly a million of operating profit, because it carries no cost of revenue at all. And the support deflection adds no revenue whatsoever, which makes it invisible in the way product work is usually described, while improving margin by four points.

Three proposals, each defensible as “worth about a million”, and they do completely different things to the business. Presenting any of them without naming the line it lands on is how a good proposal loses to a clearer one.

Sector lens#

Remember this#

• Revenue, cost of revenue, gross profit, three expense lines, operating profit. That is the whole structure.
• Cost of revenue is anything that scales with customers served. Inference belongs there.
• Gross margin describes the shape of the business and decides how much you can afford to spend acquiring customers.
• Name the line your work lands on before making the argument.
• Revenue with no cost attached is the best shape available, which is why pricing work is underrated and why an AI upsell can add revenue while losing money.

Try it#

1. Run pandl.go with your company’s real figures. What is your gross margin, and where does inference sit?
2. Find your company’s actual statement, or the closest thing to it, and locate every line in the table above.
3. For your top three roadmap items, name the line each one moves. Then check whether the person who owns that line knows the work exists.

Check yourself#

1. What is the test for whether a cost belongs in cost of revenue?
2. An AI feature adds nine hundred thousand of revenue and makes the company worse off. What does that tell you to check before approving one?
3. Why does putting inference in the wrong line matter, given the total is unchanged?