AI Agency Pricing: What You'll Actually Pay to Build an Agent

The first real number usually arrives after a discovery call, and it rarely matches what a search turned up. Agency rates for agent work sit above general software rates, and a first build can run from a few thousand dollars into six figures depending on how many systems it touches. This post is how to read that quote. It is not a price for our work.

Platform pricing, model by model, is on the comparison page. What an outside team charges to build the workflow is a different bill. The ongoing costs a platform quote leaves out are in the hidden costs of agents.

What do agencies actually charge to build an agent?

Clutch's software development pricing guide, built from verified client reviews rather than agency self-reporting, puts the broad median at $24 to $49 an hour. US agencies run $50 to $99. Agencies in Canada and Australia run $100 to $149. Clutch notes that specialized skills, including AI and machine learning, command a premium above those bands. It does not publish a separate AI-only rate.

On the project side, Clutch reports an average software project cost of $132,480.29 across roughly 13 months, with a typical smaller project between $10,000 and $49,000. A single-purpose agent, one workflow and one or two integrations, usually lands in the lower half of that typical range. A build that touches three or more systems, with its own evaluation and guardrails, tends to push toward the average or past it.

The premium shows up as a multiple on top of the wage.

The U.S. Bureau of Labor Statistics puts the median hourly wage for software developers (occupation 15-1252) at $65.38 and the median annual wage at $148,100, from 1,687,890 developers in the May 2025 survey. An agency's billed rate is not the wage. It is a multiple of the wage that covers overhead, benefits, sales cost, and margin. A useful check: an employee paid near that median often bills out around two to two and a half times the wage, which is roughly where Clutch's $100 to $149 bands sit relative to $65.38. The AI premium sits on top of that multiple. An engineer who has shipped a production agent is already paid above the median, and the same overhead multiple applies. That is why a quote can land well past $150 an hour even when the agency's general software rate sits inside Clutch's bands. Ask whether you are paying a specialist rate or the standard rate with a markup for the word AI. Those should not cost the same.

Do the multiplication in the meeting, not afterward. Start from $65.38, apply the two to two and a half times overhead check, and you land near the top of Clutch's regional bands before any specialization. Then ask what the agency pays the person who will actually write the agent, because that wage is the floor under your rate, not the BLS median. If they cannot tell you whether the rate is the specialist rate or the ordinary rate with a label added, you are comparing a number that has not been explained. Clutch's guide is the public anchor because the figures come from verified reviews rather than from what agencies choose to advertise. It still will not name an AI-only rate, so the premium has to be justified in the quote itself.

Why does agent work cost more per hour than typical software?

Two forces. Demand is the first. Gartner's 26 August 2025 forecast says 40% of enterprise applications will feature task-specific agents by 2026, up from less than 5% in 2025. The number of engineers who have shipped a production agent has not grown at the same pace. In a shortage, the people who can do the work set the price.

The second force is that the work is less predictable than a typical feature. Alavi, Nozari, and Luangrath (arXiv:2511.11761, November 2025) found that small shifts in how a prompt is worded change how many tokens a model generates, and therefore what a task costs to run, without the business or the agency fully controlling the shift. Traditional software behaves the same way every time. An agent's per-task cost moves with inputs nobody fully specified. Agencies that have been burned by that build a margin in, rather than pricing the work like a static feature.

Agencies price a margin around a cost they cannot fully set.

A normal feature has a cost you can rehearse. The same input produces the same path, and the hours to build that path are what the quote is for. An agent does not behave that way once it is on live wording. Alavi, Nozari, and Luangrath showed that small changes in how a prompt is phrased change the number of tokens the model writes back, which changes what that task costs to run. Neither you nor the agency fully chooses that shift in advance. An agency that has already eaten a month of runaway usage will not price the next build as if the cost per task were fixed. The margin in the hourly rate, or in the fixed fee, is how they stop the next surprise from landing only on them. When you see a rate well above the agency's ordinary software rate, ask which of the two forces it is: scarce people, or a buffer for a running cost that moves. Those are different reasons, and only one of them shrinks after the first build is done.

What does a fixed-price build typically cost?

Anchor a fixed quote against Clutch's project data, not a round number from a search result. The typical software project runs $10,000 to $49,000. The average across all sizes, including large builds, is $132,480.29. Where a specific agent lands inside that spread depends on scope: how many systems it reads and writes, whether the data needs cleanup first, and whether the agency is building evaluation and guardrail logic or wiring an existing platform. A single-channel agent on one clean CRM integration is a different build from one that drafts, pulls analytics, and posts across three platforms with a person approving each step. A quote far outside either anchor, in either direction, deserves a direct question about what is driving it. The content workflow that often sits in the middle of those quotes is content and search.

Three facts about the work move the number more than the label does.

The word agent does not place a build inside Clutch's range. The work does. Count the systems the agent must read and the systems it must write back to, because each one is a connector that can fail and a mapping someone has to maintain. Ask whether your data is already clean enough to use, because cleanup before the agent can trust a field is a separate block of time that a cheerful quote often skips. Ask whether the agency is writing the checks that decide when the agent is wrong, or connecting a platform that already has them. A single-channel agent on one clean integration sits toward the lower part of the $10,000 to $49,000 typical project. A build that drafts, pulls numbers, and posts across three places, with a person approving each step, moves toward the $132,480.29 average and can pass it. A quote far below the lower anchor can be missing the cleanup or the checks. A quote far above the average can be real, or it can be a bundle you have not been shown. Either one is a question, not a feeling.

Why are agencies moving off a pure hourly clock?

Demand and an unpredictable per-task cost push the same direction. More agencies are moving to a fixed build fee with a written change-order process, or a smaller build fee plus a retainer tied to usage. Both split the token-cost risk instead of leaving it on an open clock that is hard to budget when the cost per task is not fixed.

Hourly time still makes sense for a short discovery engagement, while the scope itself is being defined. Once the build is defined, a fixed number with a change-order clause protects both sides better than an open clock. Some agencies now offer a fixed fee for the build plus a small per-outcome charge once the agent is live, tied to something you already track, such as a resolved conversation or a published draft. That ties their ongoing revenue to whether the agent works. It is worth asking for if a quote arrives as pure hourly time with no fixed piece.

Use the clock while the work is undefined, then freeze it.

Hourly time is a reasonable way to pay for a short discovery, because neither side yet knows how many systems, how dirty the data, or how many exception paths the agent has to survive. Leaving the build itself on that clock means every extra week of prompt revision, and every spike in tokens, is an open bill. A fixed fee with a written change order does the opposite. The number holds unless you change the work, and the change has to be named before the hours accumulate. The hybrid some agencies now offer goes one step further after launch: a fixed fee for the build, then a small charge per outcome you already count, such as a resolved conversation or a draft a person approved. Their later revenue then depends on the agent doing the thing you hired it for. If a proposal arrives as open hourly time with no fixed piece and no change order, ask for one of those two structures before you compare it with anyone else's number.

What does an agency quote usually leave out?

Ask whether discovery is billed before the build fee starts. A discovery phase alone can run several thousand dollars before any of the agent's logic exists. Ask how many revision rounds are included before extra work is hourly, and get a number, not "a reasonable amount." Ask who owns the prompts, workflows, and logic when the engagement ends. Some agencies keep the right to reuse what they built, which matters if you later want to move the build in house or to another vendor. Ask whether model API costs are billed through the agency at a markup or passed through at the vendor's rate. That one answer can move the real monthly cost once the agent is in production. The staffing side of the same decision is in building in house or hiring an agency.

Four omissions change the number you think you accepted.

Read the quote for what it never mentions. Discovery billed before the build fee can be several thousand dollars before a single rule exists, and it is easy to treat that invoice as a surprise rather than as the start of the project. Revision rounds described as reasonable are not a number, so the fourth round becomes hourly with no warning you can budget. Ownership matters the day you want to leave: if the agency keeps the right to reuse the prompts and the workflow, moving the build in house or to another vendor means rebuilding the part you thought you had paid for. Model costs are the quiet monthly line. Passed through at the vendor's rate, they are visible. Billed through the agency at a markup, they grow with volume and you will not see the markup unless you asked. None of these is a reason to reject a quote on sight. Each one is a line that should be in the document before you sign, because it will not get cheaper to add later.

How much should you budget for a retainer after launch?

Most builds need tuning after launch. The agent meets cases discovery did not surface, and prompts or model versions drift. Using the same wage-multiple reasoning as the build rate, a reasonable monthly retainer is a fraction of the build cost, scaled to how often the agent needs adjustment rather than to a fixed percentage.

A single-workflow agent with stable inputs might need two to four hours of tuning a month once it settles. A multi-channel agent in its first 90 days often needs more. Get the retainer terms in writing when you sign the build, not after launch. Confirm whether it is a fixed monthly fee, an hourly bucket that expires, or a usage charge. A fixed fee protects the budget and may under-deliver in a heavy month. A usage charge tracks the need and brings back the unpredictability a fixed build fee was meant to avoid.

Name the retainer while you still have a choice.

After launch, the agent is already in the path your team uses, and your ability to renegotiate is weaker than it was on the day you compared three proposals. Put the retainer in the same agreement as the build. A single workflow with stable inputs often settles at two to four hours of tuning a month. A multi-channel agent in its first 90 days, still meeting cases discovery missed, often needs more than that, closer to a weekly look than a monthly one. Then choose the shape on purpose. A fixed monthly fee keeps the budget still and can leave you under-served in a heavy month if the hours are capped. An hourly bucket that expires punishes a quiet month by taking the unused time away, and punishes a heavy month by running out. A usage charge follows the real need and reintroduces the moving bill the fixed build fee was supposed to end. Pick the risk you prefer while more than one agency is still in the conversation.

How do you tell whether a quote is fair?

Ask what is inside the fixed fee and what is billed separately: discovery, integrations, revision rounds, and support after launch. Ask what happens if the scope grows, and get the change-order process before you sign. Ask whether they have shipped a comparable build, in your industry or a close one, and what went wrong the first time. A team that cannot answer that last question specifically has not shipped enough of these to price one accurately.

Divide the fixed quote by the hours they disclose. If the implied hourly rate lands far outside Clutch's $50 to $149 range for the agency's region, adjusted up for the specialization premium Clutch itself notes, ask what is driving the difference. Sometimes the answer is real: unusually hard integrations, a compressed timeline, or a senior team because the build is harder than average. Sometimes the answer is padding. Either way you have a specific question.

Specific answers are what separate a real quote from a placeholder.

A useful quote gets more specific when you push on it. Ask what the fixed fee contains and what arrives as a second invoice: discovery, each integration, the revision rounds, and the support after the agent is live. Ask what happens when the work grows in the middle of the build, and get the change-order steps in writing before you sign rather than after the extra work has started. Ask for a comparable build they have already shipped, in your industry or a close one, and ask what broke the first time they tried it. A team that answers the last question with a real failure has done this enough to know where the hours go. A team that answers with a general claim of experience has not given you a reason to trust the number. The division check comes after those answers. Hours disclosed, divided into the fee, compared with Clutch's $50 to $149 for that region, then adjusted for the specialization premium Clutch already warns about. A gap is not automatically padding. Hard integrations, a short timeline, or a senior team are real reasons. You just need the reason named.

An illustrative example, not a client result. A team planning a single-channel content agent gets three quotes: $8,000, $14,000, and $41,000, all for roughly the same 120-hour build. The $41,000 quote implies over $340 an hour, well above Clutch's bands even with a premium. Asked directly, that agency says the fee includes a dedicated account manager and unlimited revisions for a year, not just the build. That may still be worth paying. It is only worth paying once it is named. These figures are not a result from a customer, and they are not what we charge.

If you want a starting point for the workflow before you take a number to an agency, get your free Systems Plan.

How we know this.

Clutch's Software Development Company Pricing Guide (September 2026, verified reviews) is the source for the hourly bands and the $132,480.29 average and $10,000 to $49,000 typical project figures. BLS Occupational Employment and Wage Statistics, software developers 15-1252, May 2025, 1,687,890 developers, is the source for the $65.38 and $148,100 medians. Gartner's 26 August 2025 press release is the source for the 40% and under-5% forecast. Alavi, Nozari, and Luangrath, November 2025 (arXiv:2511.11761), is the source for prompt wording changing token cost. The three-quote example is illustrative arithmetic, not a client engagement. The links for Clutch and the Bureau of Labor Statistics are the pages those figures come from.