What AI can actually do for a business, and what it costs

The answer, and the two price tags behind it

OpenAI sells the same model at half price if you can wait 24 hours for the answer (Batch API). Anthropic charges a fraction of the base input rate for tokens it reads out of cache instead of processing fresh (prompt caching docs). Those two lines on two pricing pages say more about what AI can do for a business than any capability list:

AI moves a specific class of work — reading, drafting, classifying, and multi-step tool use over text — from being priced in salaried hours to being priced in tokens, and your job is to find the workflows that survive being done by something fast, cheap, non-deterministic, and sometimes confidently wrong.

That is the whole answer.

Four capabilities that have measured results, not anecdotes

I built Metadata.io from $0 to $15M ARR on machine-driven campaign execution and hold six patents in AI-driven marketing, and I am building a zero-human company now — agents doing the work, me paying the bills. So I have opinions. They are worth less than the handful of things that have actually been measured in a controlled setting.

Read the last row twice. It is the single most useful number in this field for an operator. An agent chain that needs eight hours of unattended judgement is not a product yet; an agent that does a twenty-minute job and hands back a diff is.

Audio is the quiet one. Transcription runs at a listed $0.006 per minute on OpenAI's pricing page — roughly thirty-six cents an hour of recorded speech. Every sales call, support call, and interview your company has ever recorded is now searchable text for less than the coffee budget.

Where the money shows up, and where it usually doesn't

The headline case is public and self-reported. Klarna said its AI assistant handled 2.3 million conversations in its first month — about two-thirds of chats — cut average resolution time from 11 minutes to under 2, did the equivalent work of 700 full-time agents, and was expected to drive a $40M profit improvement in 2024 (Klarna press release, Feb 2024).

Treat that as a company statement, not an audit. It is a press release, published by the company, with no methodology attached.

My position: headcount reduction is the least reliable place to look for AI return, and coverage is the most reliable. Headcount savings require the AI to be as good as the person it replaces on the hard 10% of cases, which is exactly where the METR 50% number bites. Coverage requires only that the AI be better than nothing, because nothing is what you currently do.

So the work that gets funded first in a company that already knows its numbers is unglamorous. Every support ticket gets read and tagged instead of a 2% manual sample. Every recorded sales call gets a structured summary scored against your own qualification fields. Every inbound lead gets enriched and researched before a human sees it, at a marginal cost measured in cents. Nobody was ever going to staff any of that, which is the point: none of it fires anyone, and all of it produces a dataset you did not have last quarter — the input to every later automation decision. Adoption is still early enough that this is a real edge rather than table stakes. Fewer than one in ten U.S. firms reported using AI to produce goods or services in the Census Bureau's Business Trends and Outlook Survey through 2025.

The cost side has been moving in your favour hard. The 2025 AI Index reports a roughly 280-fold drop in the inference cost of reaching a GPT-3.5-level score between November 2022 and October 2024 (Stanford HAI). A workflow you priced out as uneconomic in 2023 is trivially cheap now. Re-run the arithmetic before you conclude anything.

What a working agent actually costs per run

Here is a scenario worth pricing: a ticket-triage agent. It reads a support ticket, pulls the customer record, searches your docs, decides a category and priority, and writes a suggested reply. Call it six model turns, about 8,000 input tokens per turn once you count the system prompt and retrieved context, and 700 output tokens. Run it on 20,000 tickets a month.

Move the sliders and the model picker. The point is not my numbers — it is the shape of the curve when you change model tier.

Three levers change the total more than any prompt engineering.

The thing that wrecks a budget is not the per-token price. It is turn count. An agent that should take six turns and takes sixty because it keeps re-reading the same page has multiplied your bill by ten and produced nothing.

The failure modes you are buying

Unbounded loops. The most common production failure in tool-using agents is the model repeating a step, re-reading the same source, or ping-ponging between two tools until something runs out. The fix is a hard step cap and a wall-clock timeout enforced outside the model, in the orchestration layer — not requested politely in the prompt. Durable-execution systems exist for exactly this: timeouts, retries, and idempotency keys as infrastructure (Temporal docs).

Prompt injection through retrieved content. If your agent reads a web page, an email, or a PDF a stranger wrote, that text is now instructions competing with yours. Prompt injection and insecure output handling are the top entries in the OWASP Top 10 for LLM Applications for a reason. Any agent with both untrusted input and a write-capable tool is one crafted sentence away from doing something you did not ask for.

Schema-valid nonsense. Structured outputs guarantee the shape, not the truth. You need a validation step that checks values against reality — does that customer ID exist, is that refund amount inside policy — before anything is written anywhere.

You own the output. Legally and commercially, the company is responsible for what its automated system says and does. The EU AI Act entered into force on 1 August 2024 with obligations phasing in on a published schedule (Regulation (EU) 2024/1689), and NIST's AI Risk Management Framework exists as a public way to document what you decided and why. Write the decision down before an incident, not after.

Every automation decision reduces to two dimensions: how expensive it is to undo, and how fast you can tell whether it was right.

Top-left is where you start. Bottom-right is where pilots go to die.

What I would not use

I would not give an agent spend authority. Not a card, not an ads budget, not a procurement flow. I took Metadata.io through $50M of raised capital and I am running autonomous agents in production on my own projects today, and I still put irreversible money movement behind a human approval or a hard-capped, pre-authorised limit. The reason is the loop failure above: the same bug that burns sixty turns instead of six burns sixty thousand dollars instead of six thousand.

I would not start with a multi-agent swarm. The default architecture people reach for — a planner, a researcher, a critic, a writer — multiplies token cost and failure surface before anything works once. Anthropic's own engineering guidance says to find the simplest composition that solves the problem and only add agentic complexity when simpler patterns fall short (Building effective agents). One model, three tools, a step cap. I wrote up the honest version of what a multi-agent system is and is not in multi-agent systems, defined without the hype, and the minimum viable build in three tools, a step cap, a JSONL note.

I would not buy a platform that hides the prompt and the logs. If you cannot see the exact prompt sent, the tool calls made, the tokens spent per run, and the failures, you cannot debug it and you cannot price it. That is not a vendor preference. It is the difference between an asset and a black box you rent. For tool plumbing, Model Context Protocol is the open option worth standardising on rather than a proprietary connector format.

I would not automate a process nobody can describe. If a human cannot write down the inputs, the tools, and the pass/fail check for a task, an agent will not discover them. Ambiguity is not a prompt problem.

I would keep GUI-driving agents off systems of record. Pixel-level computer use is real and occasionally the only option for software with no API, but the error modes are worse — a misclick writes to the wrong field with no schema to catch it. The trade-offs are in a computer use agent drives a GUI with pixels, not APIs.

Where humans stay, judged task by task and not by job title

"Can AI do sales?" is the wrong shape of question. Prospect research, list building, call summarisation, CRM hygiene, and first-draft follow-up are all in the top-left quadrant above. Negotiating a non-standard contract term with a customer who is already annoyed is not. Split the role into tasks and score each one on reversibility and verification cost, which is what I do in will sales be replaced by AI? Answer it task by task.

The same logic settles the strategic question underneath all of this. If your competitor can call the same model, on the same public pricing page, with a similar prompt, the model is not the advantage. What survives is your proprietary data, your distribution, your workflow depth, and the accumulated record of which automations worked — a point I argue in an AI moat is what survives when your rival uses your model.

The part I would act on this week is instrumentation. Every agent run should write one line: inputs, tools called, turns used, tokens in and out, outcome, and whether a human had to fix it. It costs nothing but the decision to log it. Without that file you are guessing, and guessing is the most expensive thing in the stack.