AI agent builders charge per run, per credit or per seat, and some sell more than one of these. As of November 2026, n8n charges per workflow run, Microsoft Copilot Studio per credit, and Salesforce Agentforce per credit, per conversation or per seat. For one invented workload of 2,000 runs a month, the list prices come to €20, $400 and $800 a month. The worked example below shows what each of those figures leaves out.
I read these prices on each vendor's own page on 3 November 2026 and link that page. They can change after that date. I chose these three because they charge in different ways, so this is not a list of the whole market. I offer agent builds myself, which gives me a stake in whether you buy a builder at all.
This note covers the platform fee. What it costs to have a custom agent designed and built is a separate number, and how much an AI agent costs to build deals with it.
The options side by side
All prices are list prices as of November 2026.
| Vendor and option | You pay for | List price | When the allowance runs out |
|---|---|---|---|
| n8n Starter | Workflow runs | €20 a month, billed annually, for 2,500 runs | Trigger and webhook runs fail at the limit. Chat runs are not blocked |
| n8n Pro | Workflow runs | €50 a month, billed annually, for 10,000 runs | Trigger and webhook runs fail at the limit. Chat runs are not blocked |
| Copilot Studio credit pack | Credits, by feature used | $200 a month per 25,000 credits | Agents are blocked at 125% of prepaid capacity unless pay-as-you-go covers the overage |
| Copilot Studio pre-purchase plan | Prepaid credit units | "Save up to 20%", no unit price on the page | The plan card offers pay-as-you-go coverage when prepaid credits run out |
| Copilot Studio pay-as-you-go | Credits used | No per-credit rate on the page | No allowance to run out, billed at the end of the billing period |
| Microsoft 365 Copilot add-on | Seats | $30 per user a month, paid yearly | No credit pool for covered use, within fair usage limits |
| Microsoft 365 Copilot Business add-on | Seats | From $21 per user a month paid yearly, from $18 for the first year under an offer ending 31 December 2026, $25.20 paid monthly | Not named in Microsoft's "No charge" column, see below |
| Agentforce Flex Credits | Actions | $500 per 100,000 credits, 20 credits an action | Billed at your contracted rate, with no overage penalty |
| Agentforce add-ons | Seats | $125 per user a month | Unmetered for employees |
Per run: n8n
n8n's pricing page says pricing is "based on monthly workflow executions, regardless of complexity". It defines an execution as "a single run of your entire workflow", however many steps that run has. All plans include unlimited users, so adding people does not change the bill. Starter and Pro are both hosted by n8n, and the page lists their prices in euros.
The plans differ in more than the run count. The same page lists a maximum execution duration of 5 minutes on Starter and 40 minutes on Pro. Metric-based evaluations cover 1 workflow on Starter and are unlimited on Pro.
The plan table shows overage fees as "Capped" on both plans, and its help text says "Starter and Pro are capped and have no overages". So there is no extra bill. n8n's support article says executions at the limit "fail immediately with an error", and that they are not "queued, paused, held, or retried automatically". That covers runs started by a trigger or webhook, along with retries, error-handler workflows and CLI runs.
Chat agents get an exception. On a paid plan, the article says "manual executions started from the editor, chat executions, and sub-workflow executions are not affected". It doesn't define a chat execution, so ask n8n how it treats your chat setup. An agent started by incoming orders needs an alert well before the limit.
n8n's documentation also describes a free Community edition that you host yourself. The server and its upkeep then become your job.
Per credit: Copilot Studio and Agentforce
With credits, the design of the agent sets the bill, because each thing the agent does draws from a pool.
The Copilot Studio pricing page has two plan cards, Pre-purchase and Pay-as-you-go. Both say "An Azure subscription is required to use agents". The Pre-purchase plan is prepaid Copilot Credit Commit Units, and the card says "Save up to 20%". Pay-as-you-go bills the credits you use at the end of the billing period. The page gives no rate for either. The one credit price it shows sits in its compare table and FAQ, at $200 a month per pack of 25,000 Copilot Credits, so my example below uses packs. Microsoft's rates charge 2 credits for a generative answer and 5 for an agent action. An agent on a reasoning model pays those rates plus 10 credits per 1,000 tokens. The rates apply to the language models Copilot Studio provides. A model you bring yourself is billed separately.
Agentforce Flex Credits cost $500 per 100,000, and an Agentforce action uses 20 credits. That makes $0.10 an action, which matches the one-action example on the same page. Salesforce counts updating a record and summarizing a case as actions. Unused credits do not roll over into later subscription terms. Salesforce also sells customer-facing agents at $2 per conversation, but one org cannot use both Conversations and Flex Credits.
Per seat: Microsoft 365 Copilot and Agentforce add-ons
A seat price suits people who work with an agent during the day. Every seat price here comes with conditions.
Agentforce add-ons for Sales, Service, and Field Service and Operations cost $125 per user a month, with unmetered usage for employees. Salesforce's cheaper $5 user licence is a different product. It requires Flex Credits and its employee usage is metered.
Microsoft lists two Copilot add-ons. Microsoft 365 Copilot costs $30 per user a month, paid yearly, and you need a qualifying Microsoft 365 plan to buy it. Microsoft 365 Copilot Business is only for existing customers on an eligible Business plan. It starts from $21 per user a month paid yearly, or from $18 under an offer that ends on 31 December 2026 and covers the first year only. Paid monthly, it costs $25.20.
Microsoft's rates show "No charge" when a licensed employee uses an agent and the agent operates under that person's identity, within fair usage limits. That column reads "Used by Microsoft 365 Copilot licensed user" and doesn't name the Business add-on. The Business pricing page ticks "Agent usage" for the add-on, and its FAQ says a Microsoft 365 Copilot license lets licensed users "build and use internal agents within Microsoft 365" at no additional cost. That points the same way, but I could not confirm the "No charge" column for this add-on on any Microsoft page I opened, so ask Microsoft before relying on it. The Business page also warns that "Some AI experiences require usage-based billing".
For agent flows, Microsoft says the exemption "doesn't apply when the agent runs autonomously, for example after an event trigger". I would not count on a seat licence to cover an agent that starts on its own.
One invented workload, worked through
This example is invented and no client stands behind it. Replace the inputs with your own.
An agent handles 2,000 orders a month. Each order is one run with four tool calls: check stock, check the shipping date, approve the order, email the customer. I took that shape from the order-processing example in Microsoft's rates document, which comes to 4 x 5 = 20 credits.
| Option | Units per run | Units a month | List price a month |
|---|---|---|---|
| n8n Starter | 1 execution | 1 x 2,000 = 2,000 | €20, billed annually, with an allowance of 2,500 |
| Copilot Studio | 4 x 5 = 20 credits | 20 x 2,000 = 40,000 | 2 packs x $200 = $400 |
| Copilot Studio, reasoning model | 20 + 141 = 161 credits | 161 x 2,000 = 322,000 | 13 packs x $200 = $2,600 |
| Agentforce Flex Credits | 4 x 20 = 80 credits | 80 x 2,000 = 160,000 | 1.6 x $500 = $800 |
Here is how I counted.
Copilot Studio packs. I rounded up to whole packs. 40,000 / 25,000 = 1.6, so 2 packs, and 322,000 / 25,000 = 12.88, so 13 packs.
The reasoning row. It borrows a token count from Microsoft's reasoning-model example, where one follow-up question to a support agent bills 14,100 tokens. That adds 14,100 x 10 / 1,000 = 141 credits to the 20. The same example bills 461 credits when the question starts a new conversation. At that figure a run costs 20 + 461 = 481 credits and the month 962,000, which is 39 packs, or $7,800. Microsoft calls its numbers illustrative, and its example is a different agent from this one, so measure your own token count.
The Copilot Studio trigger. Microsoft's definition lists triggers among agent actions, yet its order example charges only the four calls. If your trigger bills as a fifth action, the run costs 25 credits and the month 50,000. That is exactly 2 packs, so the price stays at $400.
Agentforce buying models. Salesforce's price card reads per 100,000 credits, and its worked examples price credits pro rata, so the table does too. The page lists three ways to buy. Pre-Purchase is "a set amount of usage for the full contract term", paid upfront. PayGo is "Pay only for what you use", billed monthly in arrears, and the page gives no PayGo rate, so the $800 may not be your PayGo bill. Pre-Commit is marked "Widely available later this year". The page states no minimum purchase. If you end up buying whole blocks of 100,000, one month's credits cost 2 x $500 = $1,000.
Model costs. These figures are not like for like. n8n's €20 is the plan fee. It doesn't cover model use beyond a small sign-up credit: the page says Gateway credit usage "is deducted from a prepaid balance; new accounts get a small free amount". Otherwise you bring your own API key and pay the model's maker. Microsoft's rates cover the models it provides. Salesforce's page lists no separate token charge. Its worked examples exclude "other costs like Data 360 credits or other consumption services", and the $800 leaves those out too, along with any Salesforce licence you need.
One more tool call. Add a fifth call to each run and n8n's bill stays the same, because an execution has unlimited steps. Copilot Studio uses 5 x 2,000 = 10,000 more credits, which takes 40,000 to 50,000 and still fits in 2 packs. Agentforce uses 20 x 2,000 = 40,000 more credits, which is $200 at the same rate.
Seats. If 20 employees used the agent instead, which is another invented number, the Agentforce add-ons come to 20 x $125 = $2,500 a month. On yearly payment, the $30 Copilot add-on comes to 20 x $30 = $600 a month, and the Business add-on to 20 x $21 = $420, or 20 x $18 = $360 in the offer's first year. All of these are before the underlying plans. These figures answer a different question from the table above, because they count people instead of runs.
Which way of charging fits which job
Per run is the easiest to forecast when a workflow has many steps and you know how often it starts. Credits work when each run takes few actions without a reasoning model and your team already works in Microsoft 365 or Salesforce. Seats fit people who use an agent under their own login through the day.
Then check the limit behaviour, because the three differ. n8n fails trigger and webhook runs, though not chat runs on a paid plan. Copilot Studio blocks the agents at 125% unless pay-as-you-go is set up, and Agentforce keeps billing. Find out which one you are buying before an agent stops at month-end.
Before you trust any of these numbers
Count the steps in a real run, because vendors count units differently. A shadow run does the whole job on live inputs and writes nothing, which gives you that count before go-live. Once you have it, Microsoft publishes an agent usage estimator and Salesforce a Flex Credit Calculator, linked from its pricing page. Then ask each vendor what counts as a unit and whether failed steps count.
The fee also says nothing about where the agent's limits sit. A prompt is a request and a guardrail is a wall, and you have to check that on any platform.
When a builder is enough
A builder is probably the cheaper answer when the workflow touches one or two tools with ready connectors, its mistakes are easy to undo and someone on your team will maintain it. In that case you don't need me.
I would look at a custom build when the agent takes actions nobody can undo, when its limits have to sit outside the agent, or when the prompts must end up in your own repository. That is what an agent build from me is for, quoted per workflow after a ten-day operations map.
Bring your step count and monthly volume to a thirty-minute scoping call. We run this sum on your numbers and see whether a builder covers the job.