Intelligence Is Getting Cheap. Accountability Is Not.

The operations director of a wholesaler sent us two proposals for automating customer service email last month and asked which was cheaper. Both priced the model in detail: tokens per email, emails per month, a cost per thousand messages. Neither page covered who would answer to the customer when the system promised a delivery date or a refund the company could not keep, what evidence that person would have, or how the promise would be corrected.

The model line was the smallest number on either page, and it is likely to shrink again next year. The missing line is the one that will decide the cost of the project, and it is not on the same curve.

The price curve is real

When GPT-4 launched in March 2023, it cost $30 per million input tokens. In July this year OpenAI released GPT-4o mini at 15 cents per million input tokens. It is a smaller model, but for the reading, sorting and extracting that make up most operational work it is good enough, at a two-hundredth of the price. Anthropic, Google and the open-weight models are on the same curve.

David Cahn's essay "AI's $600B Question" (Sequoia, June 2024) asked where the revenue will come from to justify the infrastructure being built. For a company buying AI, the same numbers say something simpler: model capability is starting to be priced like a commodity input, as bandwidth and storage are. Few companies build a strategy on having cheaper bandwidth than a competitor. For most enterprises, the model is unlikely to be where they win or lose.

What does not get cheaper

In February, the Civil Resolution Tribunal of British Columbia decided Moffatt v. Air Canada. The airline's website chatbot had told a customer he could claim a bereavement fare after travelling. The airline's policy said otherwise, and the airline argued that the chatbot was responsible for its own statements. The tribunal disagreed: the company is responsible for everything on its website, whoever or whatever wrote it. The damages were a few hundred dollars. The principle was not. A reply that promises a delivery date on Thursday is the same kind of statement, made in the company's name, and the customer will plan around it.

That principle is the second cost line. Every decision a system takes on a company's behalf has to be answered for by someone in the company, to a customer, an auditor, a regulator or a board. Answering needs four things, and none of them is on the token price curve:

• A named person. Someone whose job includes this class of decision and who agreed to the rules the system follows.
• Evidence. What the system saw, which rule applied, who approved, what was written where. Kept long enough to survive an audit.
• Reversal. A way to undo the decision in the system that holds it, by someone who knows how.
• Review. People with the time and the information to check the cases that need checking, and a way to know which cases those are.

These costs scale with the consequence of a decision, not with its length in tokens. A wrong answer about a bereavement fare costs a few hundred dollars. A wrong decision that moves money costs the money. Making the model ten times cheaper changes neither.

Cheaper answers mean more to answer for

There is a second effect, and it runs the wrong way for anyone budgeting on the token price. When an answer costs a fraction of a cent, it becomes worth automating decisions that were never worth automating before: the routine order status question, the address change, the low-value credit note. Each one is another decision someone must answer for. As the price of intelligence falls, the number of decisions a company hands to software rises, and the accountability bill rises with it.

Planned on the token price, AI costs can rise while the price per answer falls. Planned on accountability, it is clear where the money goes: into the review, evidence and reversal around each class of decision. That is also where the money can be saved, because accountability can be engineered to be cheaper. A decision that can be undone in seconds needs less review than one that cannot. Evidence captured as a side effect of the work costs nothing to collect at audit time. Review aimed at the cases that need it costs a fraction of review spread across everything.

The lever most budgets miss is time. What a wrong decision costs depends on how often it happens, what it costs when it does, and how long it goes unnoticed. The first two get the attention, through better models and more review. The third is usually the cheapest to change. A wrong delivery promise found the next morning is a phone call to the customer. The same promise found when the goods fail to arrive is a complaint, and sometimes a lost customer. A daily check that compares what the system promised with what a person would expect, run on a sample, often buys more accountability than doubling the review queue.

Where the value will sit

If intelligence is cheap and accountability is expensive, value in enterprise AI moves to the places that make accountability cheap. We see three.

The systems of record. The ERP, the bank, the policy administration system and the ticketing system are where a company's authority, audit trail and money already live. Their permissions, approval workflows, period controls and change logs are decades of accumulated accountability. AI that acts inside those systems, through their own interfaces and under their own controls, inherits all of it. AI that works beside them, in a separate assistant that people copy answers out of, has to rebuild all of it or go without. We expect much of this year's chat-first work to be rebuilt for that reason.

The company's own rules and records. Every enterprise has an unwritten rulebook: which differences a clerk lets through, which customers get an exception, which carrier's dates can be promised as they stand. A system can only be answered for if its rules are written down and owned. The companies that write their rules down, and keep a record of every decision and every correction, will have something a competitor cannot buy from the same model vendor.

Whoever carries the outcome. Foundation Capital argued in April that AI will let companies sell finished work rather than software, and Sequoia's October essay on the reasoning era makes a similar case for services delivered as software. We think they are right about the direction. What follows for a buyer is less discussed: a vendor who sells the work is selling accountability, and that is exactly the part a buyer cannot fully hand over. The tribunal did not ask which vendor built the chatbot. The price of outcome-based AI will be set by how much of the error cost the vendor will carry, and the contract that matters will be about who pays, who fixes and who keeps the evidence when the work is wrong.

Agents make the gap wider

Until this year, most enterprise AI we saw suggested rather than acted. A person read a summary or a draft and decided. That kept accountability cheap, because a person was answerable for every action.

That is changing quickly. In September Salesforce announced Agentforce, agents built to plan and carry out tasks inside its platform, not only to answer questions. In October Anthropic released a model that can operate a computer's screen and keyboard. When AI moves from suggesting to acting, the cost of a mistake moves from a few minutes of someone's reading time to money that has moved and a customer who has been told something.

What to change now

None of this argues for waiting. It argues for spending in different places.

1. Rewrite the business case in three lines. The cost of intelligence, which will fall. The cost of integration into the systems of record. The cost of accountability: review time, evidence, reversal and the owner's time, which will not fall and grows with consequence. If a proposal cannot fill in the third line, it is not a proposal yet.
2. Name the person before the system. For every AI initiative, write down who answers for its decisions and what evidence they would need to do it. If nobody will put their name to it, the initiative is a demonstration, whatever the slides say.
3. Build inside the transaction. Prefer designs where AI prepares or takes an action through the system of record, under its existing permissions and logs, over designs where it sits in a separate window. Measure progress in transactions the system touches, not in users who log in.
4. Write the rules down and keep the record. Start with one workflow. List the decisions in it, the rule for each, and who owns the rule. Log every decision the system takes and every correction a person makes. This record becomes the company's evidence, its training data and its bargaining position with any vendor.
5. Negotiate the error, not the price. When a vendor offers to sell outcomes, ask who pays for a wrong outcome, how fast it is reversed, and who owns the rules and the decision log if the contract ends. The answers are worth more than any discount per document.
6. Spend on making decisions cheap to stand behind. Pick a model that is good enough and design so it can be replaced. Put the engineering effort into reversible actions, evidence captured as the work happens, and review aimed at the cases that carry consequence.

The line that decides it

We expect the model to become one of the smaller lines in most enterprise AI budgets within two years. What will separate companies is whether they can say, for every decision their software took, who stands behind it and how they know it was right. That capability is built slowly, one workflow at a time, and a better model does little to shorten it. It starts with the line both proposals left out: who answers for each decision, with what evidence, and how it is put right.