How would you price an agent product?

Interviewer's question: "How would you price an agent product?" Quillfeather Legal AI sells Quillfeather Redline, an agent that reads a vendor contract and produces a full first-pass redline on its own. Emrys Buchwald runs pricing at Quillfeather. Vesna Okonkwo is general counsel at Tallis & Vane, evaluating whether to buy it.

How to answer this, stage by stage

Nobody is grading whether you can quote a number. They are grading whether you can show the arithmetic behind it and say which part of it you are least sure of.

Let's learn

Quillfeather Redline reads a vendor contract and produces a full first-pass redline on its own, marking every clause it changed and quoting the exact original text next to each edit.

Before Quillfeather, a contract took a Tallis & Vane paralegal about forty minutes to mark up by hand, at a loaded cost of about twenty eight dollars. Anything unusual went to outside counsel, at a hundred eighty to four hundred fifty dollars a contract.

Now most contracts come back with a full redline in about ninety seconds, each edit sitting next to the exact clause it changes, so a paralegal can check the reasoning instead of redoing the work.

Here is the turn: the ninety seconds is not really what Tallis & Vane is buying. What they are buying is the confidence to keep the legal team at its current size while contract volume keeps growing. That confidence has to be priced against what growing the team the old way would have cost, not against how fast the agent feels in a demo.

Same review, two very different costs. The price has to sit somewhere between them, not equal to either one.

At its worst: price this like ordinary software, one flat monthly fee for unlimited contracts, and a customer running heavy volume costs more in escalations and compute than the fee ever brings in. The vendor loses money scaling with its best, most active customers, which is exactly backwards.

What I would leave alone: a small flat monthly fee is fine for a pilot customer testing under fifty contracts a month. Escalation volume is too low there to matter, and a simple number closes the pilot faster than a metered one would.

The lesson: a price that only works at one guessed volume is not a price, it is a coincidence. Model the range before the first customer signs, not after the tenth invoice looks wrong.

Now here is the same thing as a story

The short version above is what you would say pricing this deal live. Read this one for how the flat fee actually broke.

Emrys Buchwald built Quillfeather's first price sheet the week before the company's biggest pilot went live: one number, one flat monthly fee, unlimited contracts. Simple to say on a sales call, simple to put on a slide.

For the first three months, that fee looked smart. Tallis & Vane ran the pilot lightly, about forty contracts a month, mostly simple vendor NDAs. The agent handled nearly all of them without a hand-off. Emrys watched the cost-to-serve number sit comfortably under the fee every single month.

Then Tallis & Vane's furniture line expanded into three new countries, and Vesna Okonkwo's legal team started sending Quillfeather everything, not just NDAs. Supply agreements. Distribution contracts. A few genuinely complicated ones. Contract volume tripled to nearly three hundred a month, and the mix got harder.

The pilot ran almost entirely in the bottom left corner. The real account did not.

An auditor at Quillfeather doing the monthly finance review found it first, on a spreadsheet, not in a customer complaint. The Tallis & Vane account was now costing more in compute and escalation reviews than the flat fee brought in. Nobody had done anything wrong. The fee had simply been priced for a version of the customer that no longer existed.

Emrys pulled the real numbers. About one in six of the harder contracts now needed a human paralegal to finish the review, at roughly seventeen dollars of labor each time. That escalation cost, multiplied across three hundred contracts, was quietly larger than the entire flat fee.

Eighteen dollars was never the point. The point is that even the high end of the range sits under what a paralegal already costs.

The decision I would take back is the flat fee itself, or rather, keeping it once volume and mix had both moved. It made complete sense during the pilot, when regenerating trust in the product mattered more than protecting margin on forty contracts a month. It stopped making sense the moment the account looked nothing like the pilot anymore.

Replayed with per-contract pricing in place from month one: the fee would have moved with the account instead of staying frozen at pilot-era assumptions. Three hundred contracts a month at eighteen dollars each is about five thousand four hundred dollars, comfortably above the real cost to serve, and Tallis & Vane would have seen their bill track their own usage instead of a surprise renegotiation call.

I picked the flat fee because it was the fastest thing to get a signature on. It took one finance spreadsheet, three months later, to see that fast to sell and fast to break were the same decision wearing different clothes.

BOUND, spelled out for one priceNot a discount conversation. BOUND is what tells you whether a number survives contact with real usage.

Four parts, and only one of them, the quoted proof, has nothing to do with money at all.

The fourth item is the one most price sheets never measure until it is too late.

The recap, one line per letter: break it down is cost to run versus the human alternative, own the numbers is one dollar compute against twenty eight dollars paralegal time, use a range is fourteen to twenty two dollars and never bill a hand-off, nail the sanity check is comparing the monthly bill to a paralegal's monthly cost, and direction is the escalation rate, the one number that can flip the whole model.

And if you want to be sure it really works, try it somewhere elseSame five letters, an HVAC dispatch agent instead of a legal redline. A different swing assumption, the same shape of decision.

Meridian Field Services sells an agent that triages incoming HVAC repair calls, diagnoses the likely fault from a customer's description, and schedules a technician with the right part already loaded on the truck.

Mapped onto BOUND: break it down is cost per triaged call, compute plus the cost of a wrong part guess sending a truck back out, bounded by what a dispatcher currently costs per call. Own the numbers: compute runs about forty cents a call; a dispatcher costs about six dollars a call in wages; a wrong part guess costs about ninety dollars in a second truck roll. Use a range: price between two and four dollars per triaged call, well under the dispatcher's six, as long as the wrong-part rate stays low. Nail the sanity check: at two thousand calls a month that is four to eight thousand dollars, against roughly twelve thousand dollars a month for two dispatchers doing the same volume. Direction, the swing assumption here is different from the legal case: it is not how often the agent hands off, it is how often its part guess is wrong, since one wrong guess costs as much as forty five correctly triaged calls.

Same shape of gate, a legal redline instead of a repair truck.

Swap the trigger and it still runs.
Speed: an interviewer caps you at thirty seconds. Say "price per completed task, bounded by the cheapest human alternative, never bill a hand-off," and stop.
Cost: if compute prices fall by half industry-wide, the price barely moves, because compute was never the biggest line item, escalation cost was.
The model gets better, for real: if accuracy improves so much the escalation rate drops to one in twenty, that is not a reason to cut the price, it is a reason margin quietly improves, and a smart vendor reinvests some of that into a lower price before a customer negotiates it away.

Where people run it wrong.
They price off a pilot's usage pattern and never revisit it once real volume and mix show up.
They bill for every attempt, including the ones the agent could not finish, which breaks trust the first time a customer notices.
They chase the cheapest possible compute cost instead of watching the number that actually swings the price, which is almost never compute.

How to use it live. When someone asks you to price an agent, ask yourself one thing out loud: what is the cheapest real alternative a customer already has, and does my number sit under it once real usage, not pilot usage, is the one being priced.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0