Author: SavantCat
If you have been reading the business news these past few days, you will have seen the same word wrapped in two completely opposite moods.
On one side, capital is euphoric: customer-service AI agents carry the highest valuation premium in the whole sector — roughly 219 times a company's annual revenue. A company two years old reaches 100 million US dollars of annual revenue on one AI product; in the traditional software era that took seven to ten years of grinding, and now it can be done in two.
On the other side, institutions pour cold water: Gartner predicts that by the end of 2027, more than 40 percent of agentic AI projects will be cancelled. The agent you invested in may be strangled internally before it ever runs.
Same sector: who is swimming naked, and who is really making money?
Three sets of numbers to see the truth first
The first set comes from the deployment side: the statistic circulating in the industry is that 88 percent of AI proof-of-concept projects never reach large-scale deployment. On average a company runs 33 pilots, and only 4 ever reach the production line.
The second comes from the financial model: in Galaxy Securities' estimates, the average customer acquisition cost (CAC) for an AI agent is about 50 US dollars per user, while a user's lifetime value (LTV) is only 20 to 30 dollars. Cost is a good deal higher than revenue — which means most agent products lose money on every sale.
The third comes from buyer expectations: the share of enterprise software that has genuinely integrated autonomous AI was under 1 percent in 2024, and is expected to reach only 33 percent by 2028.
Read together, the picture is clear: the agent sector lacks neither money nor imagination. What it lacks is deployment that adds up on a spreadsheet.
Why do most agents fail to make money? Because they are a single point
Look at the projects that stall, and the problem is usually not the technology: the agent works at one point and never gets into the company's whole process.
For a piece of code to go from written to genuinely usable, it passes through requirement clarification, breakdown, development, review and testing. Many companies give their programmers AI coding tools and find six months later that it does not seem much faster. The reason is not that AI writes slowly, but that the company never embedded AI in the process — the whole delivery line is still human-led, with layer upon layer of approval. AI saves you that 20 percent of development time, but fully six tenths of the bottleneck is not in writing code at all; it is in people and organisation.
In one sentence: it is not that AI is not good enough, it is that your process is not ready for AI to come in.
There is a more subtle trap as well: many people are dazzled by the idea of a general-purpose agent. A do-everything assistant that can order takeaway, book hotels and arrange your calendar sounds sexy and has enormous imaginative space — but what it solves is usually a shallow pain point. When you order takeaway, you care about what to order, not how fast the order goes through.
So who did make money? The vertical, narrow, deep players
What actually produced business value this year are exactly the vertical agents that look plain and not sexy enough:
- An insurance customer-service agent that gets reply accuracy on common enquiries to 98 percent, compresses claims approval from several days to 1 day, and saves the insurer 5 million US dollars a year;
- AI in the legal field, aimed at the single point of drafting legal documents — manual research on a case averages 20 hours, and it reaches 90 percent accuracy, bringing visible efficiency gains to the firm, with annual revenue of 100 million US dollars;
- Customer service, marketing and data analysis — jobs with clear processes and fixed rules have become the cash machines agents reach first.
The pattern is blunt: the dirtier, more vertical and less glamorous the work, the more likely it is where AI makes money. Big companies are good at platforms and demos, but getting the customisation, compliance and dirty work of a vertical industry right takes a team that knows the trade grinding it out on the ground — and that is exactly the opening for smaller players.
The rules of the game are changing too: from selling software to charging for outcomes
Software used to make money on subscriptions, per seat, per year. But the value an agent brings is non-linear — it can finish a whole job autonomously, and at that point charging by what one account costs per year makes no sense at all.
So a new charging model is growing up: payment by outcome. Let it process an invoice, close a deal, finish a compliance document, and it charges you by the quality and business impact of what was delivered, not by how long you used it.
That is good for everyone: for the client, the money goes where it counts, and no result means no payment; for a service provider that can genuinely get things done, the pricing ceiling opens up — if you save a client half a million, charging a fee in the tens of thousands is entirely reasonable.
Down to you and me, three sentences
First, do not chase a general-purpose AI assistant; go find one job with a clear process. Customer service, bid documents, reconciliation, expense claims, invoice review — wherever the rules are clear, repetition is high and someone used to have to watch it, that is the easiest place for AI to take over. You do not need an all-capable AI person; you need an AI colleague who can replace one specific post.
Second, when buying AI, do not listen to the technology, listen to the accounts. A boss who does not understand Transformers is fine — he definitely understands what hiring half a person less means. Do not pitch a feature list; pitch whether the money comes back within six months.
Third, do not wait for a stronger model before starting. Four in ten projects being cancelled does not mean AI is useless; it means AI for AI's sake is useless. First hand it one smallest job with clear rules and get that running — use it first, then talk about scaling up.
The first half of AI agents was storytelling; the second half is doing the maths. Nobody can say what a story is worth, but the numbers in the ledger do not lie.
Is there work around you that has extremely rigid rules and eats an enormous number of people? Tell us in the comments. The next post AI replaces may well be money you save.
#AI commercialisation #intelligent agents #Agent #small business #cost reduction