What an AI agent costs to build and, more importantly, to run every month.

An AI agent has a build cost and a running cost, and almost every quote you receive will state the first and wave at the second. The build cost is a one-off: somebody designs the agent, connects it to your systems, tests it against real inputs and hands it over. The running cost arrives every month for as long as the agent works: model usage, the messages and calls it sends, the platform it runs on, and the person who watches it. In our own production systems the language model has been a single-digit share of variable AI spend, with messaging the largest line, which is the opposite of what most buyers expect and the reason the second cost deserves its own guide.

This page does both. The build cost comes from UK agencies’ published 2026 prices. The running cost comes from the vendors’ own pricing pages, all read on 14 September 2026: OpenAI, Anthropic and Google for models, Twilio for UK telephony and SMS, Resend for email, and n8n, Make and Zapier for the platform. Then two worked examples, an inbound enquiry agent and an outbound chasing agent, priced at stated volumes with every unit price sourced, so you can substitute your own numbers.

Prices in dollars and euros are converted at the Bank of England spot rates for 11 September 2026, $1.3530 and €1.1649 to the pound, and are marked as conversions. UK buyers add VAT. Volumes and token counts in the worked examples are stated assumptions, chosen to be typical of the small businesses that ask us this question; substitute your own and the unit prices still hold. Nothing here is a client account, and no figure describes the size of our own operation.

The published prices for a single agent or workflow sit in a narrow band once you separate "reads messages and takes one action" from "reads messy inputs and updates several systems". The first is a few days of work on a platform; the second is integration work, and integration is what you are paying for. Our fuller survey of twelve agencies is in the automation cost guide; these are the rows relevant to a single agent, as read in September 2026, with what each agency says the price includes where it says anything.

All figures exclude VAT. The build price is set by the number of systems the agent must read from and write to, and by how messy the inputs are, not by the AI.

Models are priced per million tokens, separately for what you send (input) and what comes back (output). A token is roughly three quarters of a word. The three vendors’ list prices on 14 September 2026 are below; each also sells cheaper cached input and a batch discount, which matter at scale but not for a first agent. The point of the table is the spread: the cheapest capable model is 250 times cheaper per output token than the most expensive, and most of an agent’s work, classifying a message, extracting a date, drafting a reply, runs happily on the cheap end.

List prices as shown on each vendor’s pricing page on 14 September 2026, in US dollars, before caching and batch discounts. Google’s page notes some newer Flash models are priced at $0.75 input and $3.75 output until 31 December 2026, doubling from 1 January 2027.

This is where the money goes. A model call for one conversation costs a fraction of a penny; a text message costs 4p and a minute of AI voice costs around 20p. The table gives every unit price we use in the worked examples, from the vendor’s page and in the currency the vendor bills. Voice has two layers: the telephony (Twilio’s per-minute carriage) and the AI stack that listens, thinks and speaks, for which the published UK guide figure is 15p to 20p a minute all-in; OpenAI’s realtime models are priced per audio token, and its transcription models per minute, which we list so you can build the stack yourself.

Conversions at the Bank of England spot rates for 11 September 2026, $1.3530 and €1.1649 to the pound. Twilio notes that prices may change without notice and that carrier fees may apply. All figures exclude VAT.

A fifteen-person lettings agency receives 400 web and email enquiries a month and 300 phone calls outside office hours. The agent reads each written enquiry, answers from the property list, books a viewing into the calendar, confirms by text, and hands anything unusual to a person. Out of hours it also answers the phone. The volumes and the token counts are stated assumptions; the unit prices are from the tables above. This is an illustration, not a client account.

The written enquiries first. Assume five turns per conversation and 6,000 input tokens across them (the system prompt and history are resent each turn) with 800 output tokens. On Claude Sonnet 5 that is 6,000 × $2 ÷ 1,000,000 = $0.012 plus 800 × $10 ÷ 1,000,000 = $0.008, so $0.02 a conversation and $8.00 for 400, about £5.91. On gpt-5.4-mini the same conversation is $0.0081 and the month is $3.24, about £2.39. The confirmation texts are 400 × $0.056 = $22.40, about £16.56. Email replies stay inside Resend’s free 3,000 a month. The n8n Starter plan at €20, about £17.17, covers the executions.

Now the calls. 300 calls at an assumed three minutes is 900 minutes. Twilio’s inbound carriage is 900 × £0.01 = £9.00 plus £3.50 for the number. The AI voice stack at the published 15p to 20p a minute is £135 to £180. Add it up: £5.91 model, £16.56 texts, £17.17 platform, £12.50 telephony and number, £135 to £180 voice AI, which is about £187 to £232 a month. Drop the phone channel and it is about £40. That is the whole point of the sum: the model is 3 per cent of the bill, and one design decision about voice moves the monthly cost by a factor of five.

A recruitment agency wants 1,000 applicants a month contacted within minutes of applying, chased until they reply, and booked when they do. The sequence: one email, up to two texts, and one call attempt to their mobile, stopping the moment the person responds and never contacting anyone on the do-not-contact list. Assumptions: every applicant gets both texts, 400 reply by text, every applicant gets one call averaging 1.5 minutes including the ones that reach voicemail, and the agent makes three model calls per applicant at 3,000 input and 300 output tokens each.

Texts: 2,000 × $0.056 = $112, about £82.78. Inbound replies: 400 × $0.0075 = $3.00, about £2.22. Calls: 1,000 × 1.5 minutes = 1,500 minutes × £0.0305 = £45.75 of carriage, plus £225 to £300 for the AI voice stack at 15p to 20p, plus about £1.85 for a mobile number. Model: 3,000 calls × 3,000 input tokens = 9 million input tokens at $0.75 on gpt-5.4-mini is $6.75; 3,000 × 300 = 900,000 output tokens at $4.50 is $4.05; $10.80 in total, about £7.98. Email stays inside Resend’s free tier. Platform: 1,000 applicants at three to six runs each is 3,000 to 6,000 executions, so n8n Pro at €50, about £42.92.

Total: about £409 to £484 a month, of which the model is £7.98. Remove the call step and it is about £138. Our own published outcome shares for outbound calling explain why that step is worth questioning: roughly half of calls reach voicemail and about one in eight reach a person. An agent that texts first and calls only the ones who do not reply costs a fraction of one that dials everybody, and the quote should show you that sum before you sign.

Put the two costs together and the pattern that governs the buying decision appears. The inbound agent costs £4,000 to £15,000 to build at the agencies in the table (£750 to £5,000 if it does one thing on one system) and £2,244 to £2,784 a year to run with voice, or about £480 without it. The outbound agent costs the same to build and £4,900 to £5,800 a year to run with the calling step, or about £1,660 without. Over three years the running cost of the voice version equals or exceeds the build cost, which is why the running cost has to be modelled before the build is priced, not after the first invoice.

Three rules follow. First, route the work: an agent that sends every message to the frontier model at $50 per million output tokens instead of a small model at $1.25 has multiplied the smallest line by forty for no benefit to the reader. Second, design the channels before the model: the decision to text before calling, or to answer the phone at all, moves the bill more than any prompt. Third, cap it: a per-day and per-month ceiling that has been watched stopping the agent in production, because a limit nobody has seen fire is an assumption.

There is a fourth rule for the build price, which is to place it on the ladder before you accept it. The agencies in the table agree that a one-system agent is under £5,000 and a several-system agent is £4,000 to £15,000, so a quote of £12,000 for an agent that reads a form and creates a CRM record has either misread the scope or priced something it has not described. Ask what the extra systems are. If the answer is a retainer, the running cost and the build cost have been blended, and you should ask for them separately.

The worked examples above are what a quote should look like, with your volumes in place of ours. Any agency that has built an agent has this information, because it had to choose a model, a channel and a platform to build it, and the unit prices are on public pages. A quote that omits the running cost is not incomplete by accident; it is incomplete because the sum is unflattering or because nobody did it. Either is worth knowing before you sign, and the list below is the minimum to ask for.

Since we are a row in the table: a first agent at £4,000 to £15,000 fixed, a multi-agent system at £8,000 to £25,000, with an audit at £2,500 to £6,000 credited in full against the build. Running costs are stated before go-live at your volume, using the unit prices above, and capped in production. We run the same agents in our own products and pay the same per-message and per-minute prices you will, which is why the running-cost model comes with the quote rather than after it.

Our prices sit at the higher-integration end of the single-agent band because our agents connect to several systems and keep a person in the loop for the risky steps, which is more build than a one-system workflow at £750. If your case is the one-system workflow, we will say so and point you at the cheaper rung, because an agent that is over-built for its job costs more to run as well as to build.

Build prices are from UK agencies’ own pages and 2026 pricing guides, read on 13 and 14 September 2026, with dates shown; the full twelve-agency survey is in the automation cost guide. Unit running costs are from the vendors’ pricing pages read on 14 September 2026: OpenAI, Anthropic and Google for models; Twilio for UK voice and SMS; Resend for email; n8n, Make and Zapier for platforms. Twilio shows UK voice in sterling and UK SMS in dollars, and both are recorded as shown.

The all-in AI voice figure of 15p to 20p a minute is AI Agency Plus’s published UK estimate, used because no vendor publishes a per-minute price for a complete voice stack; OpenAI’s realtime and transcription prices are listed so a reader can build the figure from parts. Currency conversions use the Bank of England spot rates for 11 September 2026. Token counts and volumes in the worked examples are stated assumptions. The outbound call outcome shares are our own, published in full in the linked calling guide. All prices exclude VAT.