Agent Companies Sign More and More Contracts—Why Are Profit Margins Getting Thinner?

Deep Thought, Discussing AI and Aspirations —— ByteDance Deep Thought Circle

A wave of Agent companies is experiencing a strange kind of prosperity: contracts keep piling up, revenue curves steep enough to print on fundraising decks, but year-end accounting shows negative profit. What’s worse, most founders can’t pinpoint where the losses are coming from. It’s not that they’re incompetent—this business, from pricing to cost structure, hasn’t fully developed its accounting framework yet.

First, See Which Step You’re Charging At

Agent pricing models lay out as a four-tier staircase.

Most companies start at the first tier. The problem is treating the starting point as the endpoint. Stay at per-action pricing, and your rate card becomes a price list waiting to be copied: next month a competitor arrives with similar functionality at half the price, and customers have no reason not to switch.

The higher up the staircase, the harder it is to price-compare. With outcome-based pricing, customers compare “how much is this thing actually worth”; with headcount-based pricing, customers compare “how much does it cost to employ someone.” Neither of these numbers has publicly available rates to check against, and your pricing latitude suddenly opens up.

Look at a real comparison. A fully-loaded Sales Development Representative costs seventy to ninety thousand dollars a year in salary plus overhead. Some companies simply charge per Agent—twenty thousand dollars each, doing the same work, with bonuses available for hitting appointment targets. Customers do the math and understand immediately: this is a headcount calculation. The headcount budget pool is an order of magnitude larger than the software budget. Many Agent companies are fighting over scraps within the software budget when they’re actually standing at the wrong counter.

My take: what these four tiers really measure is just one thing—customer trust in you. Per-action pricing means customers trust that “this one call is worth this price”; outcome-based pricing means customers trust that “you can deliver consistently.” So pricing upgrades are never as simple as changing a rate card—they’re about monetizing trust you’ve already accumulated. Jump tiers before trust is built, and customers will vote with their feet; accumulate trust but don’t jump, and you’re essentially giving customers free money.

To Charge for Outcomes, First Learn to Charge Bonuses

Many people understand outcome-based pricing as a binary choice: either charge subscription fees as a tool, or charge entirely based on outcomes. This understanding either scares people away or gets them into trouble. There’s a transitional design in between called outcome bonuses: charge the baseline price as usual, but when results meet agreed quality standards, charge an additional fee.

The real value of this design isn’t in the bonus itself—it’s in the negotiation process. To negotiate a bonus, you must align with the customer on three things: what counts as a good outcome, who determines it, and what the criteria are. Once this conversation succeeds, the contract transforms from standard pricing to a custom contract, and your relationship shifts from vendor to something more like a shared stakeholder. Custom contracts are extremely difficult to replace, because for a competitor to steal this customer, they’d have to sit down and walk through these three questions from scratch.

This logic also works as a filter when selecting engagements. In insurance, some companies specialize in policy renewals. In freight, some companies deploy two thousand Agents to call truck drivers, negotiating rates, arranging shipments, and following through until delivery. These types of work share a common trait: outcomes are clearly measurable, the price ceiling is the original labor cost, and customers can calculate effortlessly. Work with countable outcomes deserves outcome-based pricing; work where outcomes can only be determined by customer subjective feeling, like “experience improved,” shouldn’t touch outcome pricing yet—go back and build measurement capability first.

Another phenomenon worth noting: many outcome-based contracts today are first proposed by the customer. Everyone’s uncertain about AI, and customers want to push risk back onto the vendor using outcome clauses—this is purely self-protection. So don’t treat outcome pricing as some offensive weapon; first use it as a defensive tool to catch the ball, then embed upgrade hooks in the contract. Also, be wary of that early per-action revenue—it looks like income, but it’s really more like trial samples. Customers haven’t seriously assessed value; renewal day is when real pricing negotiation happens, and many companies only discover at this moment that the price they charged and the cost they delivered never reconciled.

The Cost Black Box: Money-Losing Customers Mixed with Profitable Ones

Pricing is only one side of the ledger; the other side is more hidden.

Agent companies universally have a murky account. On the surface, cost is just model invocation fees. In reality, for a multimodal Agent that can make calls and execute operations, the bulk of costs often lies in call duration, dial frequency, third-party APIs, voice synthesis—accumulating per minute and per call, snowballing as customers grow.

What’s worse is this account is invisible. Right now everyone uses evaluation frameworks to chain Agent behaviors together and run them, only seeing a total cost at the end. Which customers are profitable, which are losing money, which steps are burning cash—all unknown. Traditional retail at least knows how much inventory was lost during month-end reconciliation; Agent companies’ value leakage can’t even be detected. Situations where customers extract all the value while you retain no profit often aren’t discovered until renewal negotiations, when bargaining chips are already scarce.

Here I need to pour cold water on a popular optimistic assumption: models are getting cheaper, so cost problems will resolve themselves. This is only half right. Token unit prices are indeed falling, but reasoning models cause token consumption per task to multiply, so the total bill per task won’t necessarily decrease. Two things are true simultaneously: tokens are getting cheaper, bills are getting more expensive. Companies betting on “costs will automatically resolve themselves” will likely be resolved by costs first.

Three Moves to Balance the Books

Concretely, this comes down to three actions.

Calculate “human-equivalent cost” for each customer. If humans did this work, how much salary and time would it cost in the customer’s location? The resulting number is your pricing anchor and your justification for future rate increases. Customers won’t compare you against competitors’ rate cards—they’ll compare you against their payroll.

Each quarter, check which tier you’re on, then proactively approach customers about upgrading—don’t wait for them to ask. Renewals and expansions are natural windows for pricing discussions; results are already on the table, making the conversation easiest. Miss the window, and customers will pull you back into the price-comparison game in the next RFP round.

Break costs down to the customer level. Which customers have negative gross margins—either adjust pricing, change delivery methods, or let go. Leave it unaddressed, and one money-losing customer will quietly eat the profits from two profitable customers.

Finally, state the boundaries. This playbook assumes your work has countable outcomes and customers large enough to sit down for custom contract negotiations. If customers only accept standard products, or outcomes can’t yet be objectively determined, then first build measurement capability honestly at the workflow level. Climb the stairs one step at a time, but don’t reverse direction: Agent companies stuck at the bottom tier—the more contracts they sign, the more labor they do for customers.