Intelligence Is Getting Cheap. Accountability Is Not.

Last month a wholesaler's operations director asked which of two proposals for automating customer service email was cheaper. Both priced the model in detail: tokens per email, emails per month, cost per thousand. Neither said who would answer to the customer when the system promised a delivery date or refund the company could not keep, with what evidence, or how it would be corrected.

The model was the smallest number on either page and will likely shrink again. The missing line decides the project's cost, and it is on a different curve.

The price curve is real

AI is priced per token, a few letters of text. GPT-4 launched in March 2023 at $30 per million input tokens. In July OpenAI released GPT-4o mini at 15 cents: smaller, but good enough for the reading, sorting and extracting of most operational work, at a two-hundredth of the price. Anthropic, Google and the freely downloadable models are on the same curve.

David Cahn's "AI's $600B Question" (Sequoia, June 2024) asked what revenue will justify the infrastructure being built. For a buyer the lesson is simpler: AI is being priced like bandwidth, and few companies win on cheaper bandwidth. The model is unlikely to be where most win or lose.

What does not get cheaper

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

Someone in the company must answer for every decision its systems take, to a customer, auditor, regulator or board. That needs four things, none on the price curve:

• A named person whose job covers this kind of decision and who agreed to the system's rules.
• Evidence: what the system saw, which rule applied, who approved, what was written where, kept long enough for an audit.
• Reversal: a way to undo the decision where it is held, by someone who knows how.
• Review: people with the time and information to check the right cases, and a way to find them.

These costs grow with what a decision is worth, not with text length. A wrong payment costs the money, and a model ten times cheaper changes nothing about that.

Cheaper answers mean more to answer for

At a fraction of a cent per answer, it pays to automate decisions that never did: the order status question, the address change, the small credit note. Each needs someone to answer for it, so as intelligence gets cheaper, the accountability bill rises.

Budgeted on token price, AI costs can rise while the price per answer falls. Budgeted on accountability, the money goes to review, evidence and reversal, which is also where it can be saved: a decision undone in seconds needs less review, evidence recorded during the work costs nothing at audit, and review aimed at risky cases costs a fraction of reviewing everything.

The lever most budgets miss is time. A wrong decision's cost depends on how often it happens, what each one costs, and how long it goes unnoticed. Models and review target the first two; the third is usually cheapest to change. A wrong delivery promise found next morning is a phone call; found when the goods fail to arrive, it is a complaint or a lost customer. A daily check of sampled promises against what a person would expect often beats doubling the review queue.

Where the value will sit

Value moves to three places that make accountability cheap.

The core business systems. The ERP (the main system for orders, stock and accounts), the bank, the policy administration system and the ticketing system already hold a company's authority, audit trail and money, with decades of permissions, approvals, locks on closed months and change logs. AI acting inside them, under their controls, inherits all that. AI in a separate assistant people copy answers from must rebuild it or go without, so we expect much of this year's chat-first work to be rebuilt.

The company's own rules and records. Every company has an unwritten rulebook: which differences a clerk lets through, which customers get exceptions, 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. Do that, and record every decision and correction, and you hold what no competitor can buy from the same vendor.

Whoever carries the outcome. Foundation Capital argued in April that AI will let companies sell finished work rather than software; Sequoia's October essay on the reasoning era says the same of services delivered as software. We agree. But a vendor selling the work sells accountability, which a buyer cannot fully hand over; the tribunal never asked who built the chatbot. Outcome-based AI will be priced by how much error cost the vendor carries, and the contract that matters says who pays, fixes and keeps the evidence when the work is wrong.

Agents make the gap wider

Until this year, most company AI we saw suggested and a person decided, which kept accountability cheap. In September Salesforce announced Agentforce, agents that carry out tasks in its platform; in October Anthropic released a model that operates a computer's screen and keyboard. Once AI acts, a mistake is money moved and a customer told something.

What to change now

None of this argues for waiting, only for spending differently.

1. Write the business case in three lines: the AI, which gets cheaper; connecting to the core systems; and accountability (review, evidence, reversal, the owner's time), which does not and grows with what is at stake. Without the third line, it is not a proposal.
2. Name the person before the system, and the evidence they need. If nobody will sign for its decisions, it is a demonstration, whatever the slides say.
3. Build inside the transaction. Let AI prepare or take actions through the core system, under its permissions and logs. Count transactions touched, not users logged in.
4. Write the rules down and keep the record. In one workflow, list the decisions, the rule for each and its owner. Log every decision and correction. That record is your evidence, training data and bargaining position.
5. Negotiate the error, not the price. Ask who pays for a wrong outcome, how fast it is reversed, and who owns the rules and log if the contract ends. Those answers beat any discount per document.
6. Make decisions cheap to stand behind. Pick a good-enough model, keep it replaceable, and spend on the levers above: reversal, recorded evidence, targeted review.

The line that decides it

Within two years we expect the model to be a small line in most AI budgets. Companies will differ in whether they can say, for each decision their software took, who stands behind it and how they know it was right. That is built one workflow at a time; a better model barely shortens it. It starts with the line both proposals left out.