By Pavan Kumar T V · 22 August 2026

Series: Part 1 defined what an agent actually is. This is Part 2, on money. Part 3 is careers, Part 4 is the org.

Here's a sequence I've watched play out a dozen times.

Someone builds an agent that saves a person four hours a week. They open a competitor's pricing page. They see $19/month. They type $29 into their own page and ship it.

That takes about ninety seconds and it's the most expensive ninety seconds in this category.

Meanwhile, in the same market, vendors are charging $800 to $2,000 per month for a single agent and closing deals. Same underlying technology. Roughly 50x the price.

The difference isn't the software. It's what the buyer is comparing it against.

Nothing Is Priced on What It Costs to Make

Start outside software, because this isn't a software observation.

There's a chocolate bar in Dubai selling for several times what a bar of chocolate costs. The cocoa isn't several times better. There are tomatoes with a sticker on them going for a multiple of the tomatoes sitting beside them. There's a factory in Switzerland that has spent a century convincing people its identical-tasting product belongs in a gold box.

Meanwhile food, the one input where going without will kill you, is the cheapest thing in the shop.

Value has close to zero correlation with cost of production. It's set by scarcity, by framing, by who's buying, and by what they're comparing against.

Software people hate this. We want the good architecture to be the expensive one. It never was.

A Tool Competes With Tools. A Colleague Competes With a Salary.

This is the only pricing idea in this post that matters.

Buyers don't price things. They price the alternative.

Show them a tool, and the alternative is another tool. Your ceiling is whatever the cheapest competitor charges, and the floor falls out the moment somebody ships a free tier.

Show them something that does a job a person currently does, and the alternative is a person.

Same software. Two completely different numbers.

The market has already moved. Seat-based pricing for AI products dropped from 21% to 15% of vendors in a single year. The vendors charging $800 to $2,000 a month per agent aren't anchoring to a $20 software license. They're anchoring to a $60,000 salary, and they're selling into the HR budget instead of the IT budget.

Per-seat isn't just leaving money behind, either. It's structurally broken for agents. Per-seat means you get paid per human using the product. But a good agent means the customer needs fewer humans. You've built a pricing model that pays you more when your product works less well.

We've been salary-pricing the whole time without admitting it, by the way. A seat is a person. We just never said it out loud, because the software sat next to the person instead of doing their work.

Agents drop the disguise.

Whose Salary, and How Many of Them

If the anchor is payroll, the market question stops being what does it do and becomes whose day does it change, and how many of them are there.

A hundred operations people getting meaningfully faster is a bigger business than one manager getting brilliant. The software might be literally identical. Depth of improvement matters far less than the size of the payroll pool it touches.

Before writing anything, answer three questions:

whose        which role's day changes, by name and salary band
how many     ten of them in the company, or four hundred
what kind    is the salary eliminated, amplified, or made unnecessary

The third one decides whether you're in a knife fight.

Eliminating a role means competing against the full cost of that role. Sounds great, until you notice the person whose role it is has to adopt your product. See Part 4 for how that goes.

Amplifying a role means a smaller number and a much easier sale, because everyone in the room wants it to work.

I'd take the smaller number and the friendly room most days. That's a strategy decision, not a pricing one, but it gets made at pricing time whether you meant it or not.

Token Pricing Is a Trapdoor. Outcome Pricing Isn't.

This is where I need to be more precise than "don't price on usage," because the whole market is pricing on usage and it's mostly working.

Traditional software had near-zero marginal cost. One more customer cost you nothing, which is the entire reason SaaS multiples look the way they do. Agents don't have that property. Every unit of work has a real cost underneath it, and that cost moves with somebody else's price list.

So the safe-feeling move is cost-plus. Price per token, per call, per compute minute.

Run it forward. Models get an order of magnitude cheaper, which they keep doing. If your price is pinned to consumption, your revenue collapses in lockstep with your costs. You did the work. Someone else's GPU margin improved.

Price per token and you're a reseller of somebody else's cost curve.

Now look at what the market actually does. Intercom's Fin charges $0.99 per resolution. Zendesk is around $2.00 pay-as-you-go. HubSpot is around $0.50. Gartner expects 40% of enterprise SaaS contracts to carry outcome-based elements by 2026, up from 15%.

That's consumption pricing too. But it's a different animal, and the difference is one word:

• Token pricing meters a cost unit. Your revenue is chained to your supplier's price list.
• Outcome pricing meters a value unit. Your revenue is chained to how much the customer got.

Same mechanism, opposite direction. When the model gets ten times cheaper, the token-priced vendor loses 90% of revenue and the outcome-priced vendor keeps every cent and books the spread.

Price the resolved ticket, the filed return, the reconciled invoice. Not the tokens it took.

The Moat Is the Memory

One more consequence, and it's the most under-discussed thing in the category.

If an agent compounds, the version working with a customer for two years is not the same product as a fresh install. It knows their exceptions, their vocabulary, the six things they always do differently, the answer one person gave in March that nobody wrote down anywhere else.

A competitor can match your features in a weekend. They cannot match two years of your customer's own corrections.

We used to build moats out of integrations and stored data. Both real, both shallow. An export script and a migration weekend beats them. Accumulated judgement is different, because the customer can't export it either. They never wrote it down. It exists only as the thing that stopped being wrong.

I've argued before that UI is not a moat for AI startups. Neither is the model, which you rent from the same three vendors as everyone else.

Here's what would kill this argument, because it's the most falsifiable thing I'm claiming. If context windows get long enough and cheap enough that a fresh install can swallow two years of a customer's documents, tickets and message history in one pass and behave like it lived through them, the moat evaporates overnight. I don't think that happens, because most of what accumulates was never written down anywhere. But that's the bet, and you should know it's a bet.

Which puts a strange obligation on the engineering. The learning path, meaning how a correction gets captured, versioned and made durable, isn't a nice-to-have on the roadmap. It's the asset. Everything else on your pricing page is rentable.

The Bottom Line

Sell a tool, and you're in a race to the bottom against every other tool. The bottom is free.

Sell a job, and you're compared against what that job costs, which is the largest number on the company's income statement.

The engineering that gets you from the first to the second is in Part 1 and Part 4. The decision isn't engineering. It's what you're willing to claim your software actually is.

Charge like a subscription, and someone undercuts you by a dollar next quarter.

Charge like a hire, and the only thing that can undercut you is a person.