Ask what an AI agent costs and you get two answers two orders of magnitude apart. Both are honest. They describe different markets that share a vocabulary.
The multiple the agent wording carries at the median
Of the time a random agent posting beats a random automation one, against 50% if they were identical
Of agentic AI projects Gartner expects to be cancelled by the end of 2027
Neoteric, a development agency, publishes a band of roughly $25,000 for a proof of concept and $100,000 and up for a production agent. On the open market the same words cover much smaller jobs: on one Upwork snapshot from September 2026, fixed-price automation work was posted at a $100 median and work described as an AI agent at $250. Jobs described as AI-agent work carry higher advertised budgets. What decides the result is the spec you write before anyone quotes.
Percentiles describe the fixed-price postings inside each group, because only a fixed-price posting carries a budget at all.
The whole marketplace
AI and automation, broad filter
AI agent language only, narrow filter
| What was measured | Postings | 25th | Median | 90th |
|---|---|---|---|---|
| The whole marketplace | 108,082 | $30 | $100 | $1,000 |
| AI and automation, broad filter | 5,603 | $20 | $100 | $1,500 |
| AI agent language only, narrow filter | 1,260 | $35 | $250 | $3,000 |
The fixed-price counts the medians rest on are 41,885 marketplace-wide, 1,982 in the broad row and 406 in the narrow agent row. Every dollar figure is a statement about fixed-price work and nothing wider.
A language model may sit inside a predefined step, reading a message or writing a first draft, and the thing is still a workflow. It is cheaper, it runs faster, and you can test it, because a fixed set of steps has a fixed set of ways to go wrong.
The path depends on what the system finds when it gets there. Researching a company across whatever sources exist for it. Working a support case whose shape you learn by reading it. Those are agent problems, and they are rarer in a normal business than the market’s vocabulary suggests.
“Many use cases positioned as agentic today don’t require agentic implementations.”
Buying the more expensive word for the cheaper problem is the most common way to overpay here, and nothing in the quote will tell you it happened.
Because four things move the number, and none of them is the model.
One scheduled job that reads a spreadsheet, calls a single API and writes a row is a weekend. A system reaching into a CRM, a support desk and a billing platform, with permissions and an audit trail for each, is a quarter. Both get posted under the same job title.
Nobody is training a model here. In practice this is three weeks of somebody discovering that the internal documents contradict each other, and that work lands in the invoice under a technical-sounding line item.
On the 14 September snapshot, 2,227 of the postings that name an experience level ask for expert and 3,173 for intermediate. Only 203 are posted at entry level. This is not beginner work at any price point on the distribution.
The largest hidden variable, and the one nobody quotes. A quote that does not separate these four is pricing an undefined object, which is why two agencies can price the same paragraph ten times apart and both be acting in good faith.
A $100,000 agent that gets abandoned and a $3,000 agent that gets abandoned cost you the same thing, and it was never the money. What it cost you was the months it ran. Add the internal sponsor who spent their credibility arguing for it, and the problem still sitting exactly where it was in January.
The useful question is narrower and harder. What will you be holding on the day you find out whether it works, and how long until that day arrives?
Most of these projects fail at the brief rather than at the model. Writing the rules down is your job. No vendor can do it without inventing your business in the process.
That document costs you an afternoon and makes every quote you receive comparable. If you cannot write those five things down, the problem is not that you have not found the right developer yet. The job does not exist yet.
1What it does, in one sentence a stranger could act on
Not “handle customer inquiries”. A usable version reads: read the support inbox every 15 minutes, classify each message into one of six categories, draft a reply for the three routine categories, and leave the rest untouched for a person.
2What it may touch, and what it may not
Which systems, which data, read access or write access. This is where most of the cost and nearly all of the risk lives.
3What “working” means, in numbers
On 100 real messages from last month, at least 90 classified correctly, and zero replies sent in the three categories it may not touch. Write the threshold before the build starts, because afterwards it becomes a negotiation you will lose.
4What evidence proves it
A demo proves nothing. Ask for the raw output on a real sample, kept, so you can run the same test in three months against a different model and see whether anything moved.
5Who signs off, and where
Human review is a named engineering pattern, not an admission that the system fell short. Decide up front which steps a person approves, because retrofitting an approval step later means rebuilding the flow around it.
These systems drift. A model gets updated, a page changes its layout, an API changes a field, and the thing keeps running while quietly getting worse. Nothing errors. The output simply stops being right, and because nobody reads every result, the failure can run for weeks before anyone notices.
“Evals make problems and behavioral changes visible before they affect users, and their value compounds over the lifecycle of an agent.”
An eval is your acceptance test, kept and re-run. If the build does not hand you the test set along with the system, you have bought something you cannot check.
Someone notices
weeks after the drift began
Once the spec exists, the choice is between two kinds of agreement. Both are legitimate purchases and they are not interchangeable.
You approve a scope and a team starts. You pay whether or not the finished system clears your bar. The price is agreed before anyone knows whether the thing works, which is exactly why quotes on the same paragraph vary by a factor of ten: the vendor is pricing its own uncertainty and you are funding it.
Almost every result on the first page of Google for “ai agent development services” is selling this contract, and for a whole class of work it is the right one.
Pond is an AI Workforce Marketplace. You pick nothing up front. You post the task once with the spec and the acceptance test attached, and a mixed field works the same brief in parallel: human experts, AI agents, and people operating AI agents.
You come back to finished submissions carrying the evidence your brief demanded, and you pay for the ones you keep. You review the results, not the workers and not the process.
Pond’s own platform data, read on 18.09.2026. Posters who ask for 20 submissions often receive 60 or more. A platform fee applies on top of the reward pool.
Over 100 contributors validated whether the team should build an in-house scraper. They built it. The argument that usually happens in a planning meeting happened as evidence instead, before the engineering budget moved.
TinyFish put a web execution agent in front of Pond’s contributors, and 83 people and agents ran real tasks through it. More people showed up than the reward pool could cover, by a factor of eight.
Both of those are tasks that happened, written down as they happened. Neither is a promise about yours, and neither is a custom agent built end to end through Pond.
Some of this work should never be posted as a task, and saying so is more useful than pretending otherwise.
If a builder has delivered for you before and understands your systems, competition adds nothing you need. Give them the work.
If the real need is somebody on call for the next year while the system drifts, that is an ongoing relationship rather than a deliverable.
If the requirements will genuinely be discovered by building, there is no bar to compete against. Pay for the discovery and do it properly.
Or when it is regulated and a named, accountable party is part of what you are buying.
And if the work you are actually buying is testing rather than a build, QA outsourcing has its own four routes and its own prices.
Two ranges, both real. One development agency, Neoteric, publishes a band running from around $25,000 for a proof of concept to above $100,000 for a production system. On the open market, the 1,982 fixed-price postings in the broad AI and automation group on a 14 September 2026 Upwork snapshot were posted at a $100 median, and the 406 fixed-price postings using agent language specifically advertise a $250 median with a $3,000 90th percentile. Write your spec first and the quotes you receive become comparable.
Less than agent work, measurably. On the same 14 September 2026 Upwork snapshot, the 1,576 fixed-price automation postings left once agent language is stripped out ran a $100 median, which matches the median across the 41,885 fixed-price postings marketplace-wide. The 406 fixed-price postings that specifically use agent language ran a $250 median. If your process has steps you can write down in advance, you are buying the cheaper category, and you should describe it that way when you brief anyone.
A workflow follows steps you defined. An agent decides the steps. OpenAI’s guide calls a workflow a sequence of steps that must be executed to meet the user’s goal, and agents systems that independently accomplish tasks on your behalf. The commercial consequence is large: plenty of work sold as an agent is a workflow with a language model inside it, and a workflow is cheaper, faster and easier to test. Gartner’s own analyst has said many use cases positioned as agentic do not require agentic implementations.
Conditionally. By default it is a conversational assistant: you ask, it answers. In July 2025 OpenAI introduced a separate agent mode, announcing that ChatGPT can now do work for you using its own computer, handling complex tasks from start to finish. OpenAI now banners that launch post as outdated, so check the current product documentation for how the feature works today. The principle holds: an assistant responds, an agent plans, uses tools and acts with reduced supervision.
The first page of Google for that term is almost entirely enterprise development agencies, and they are genuinely capable of enterprise work. Judge them on three questions rather than on their client logos. Will they hand you an acceptance test along with the system? Who owns it when it drifts, and is that written down anywhere? And what does working mean in numbers, agreed before the build starts? A provider who will not answer those three is selling you weeks.
Postings measure scope rather than delivery, but the scope is informative. On the 14 September 2026 Upwork snapshot, 1,295 of 5,603 AI and automation postings, or 23.1%, are scoped at more than six months, against 26.7% marketplace-wide, so most of this work is scoped shorter than half a year. The variable is rarely the model. It is how many systems the thing has to touch, and how much of your data needs cleaning before it can touch anything.
