Agentic AI total cost of ownership: an honest guide for UK SMEs

If you are a UK SME owner trying to work out what agentic AI will actually cost you, most of the writing online is not going to help. Vendor pages show the seat price and stop. Industry analyst pieces quote dollar figures aimed at companies with a hundred staff and a procurement department. The Google AI Overview that probably brought you here cites a £50,000 to £200,000 range that has nothing to do with a Norfolk food producer or a Suffolk professional-services firm.

The question you are actually asking is more specific than that. It is not "which is cheaper" but "what will this cost me over the next two or three years, including the parts nobody quotes". This piece is an honest answer to that question, with a worked sterling model at SME scale, and a clear view of when off-the-shelf is the right answer and when a custom build earns its cost.

What total cost of ownership really includes for agentic AI

The licence price is the part most comparisons fixate on. It is also the part that matters least once you are past month three of running an AI system at any meaningful scale.

A complete TCO picture for agentic AI covers eight cost lines:

1. Licence or subscription. The seat price, the platform fee, the base subscription. The number on the vendor pricing page.
2. Per-action or per-message metering. The cost that scales with usage. Per AI-resolved conversation, per API call, per agent task. This line is invisible at trial volumes and dominant at production volumes.
3. Integration and data plumbing. Connecting the AI to your existing systems: your CRM, your accounting software, your e-commerce platform, your inbox. Either through built-in connectors at extra cost, through middleware like Zapier or Make, or through custom integration code.
4. Ongoing maintenance and prompt or model updates. AI behaviour drifts when the underlying model is updated, when your business changes, when edge cases surface. Someone has to maintain the prompts, the routing rules, the fallback logic.
5. Internal staff time. Hours spent supervising the AI, reviewing its outputs, handling escalations, training the rest of the team. This cost is real even when it is not invoiced.
6. The cost of being wrong. Customer issues caused by agent errors, hallucinations that get sent to suppliers, time spent unpicking actions the AI took that it should not have. Bounded systems reduce this cost. Unbounded ones do not.
7. Vendor lock-in and switching cost. Templates, configurations, integration logic, and accumulated knowledge that have to be rebuilt if you move. Industry typical for moving away from a configured AI helpdesk product is £2,000 to £4,000 in rebuild time alone.
8. Cost of scaling. What it costs you to double the volume, add a workflow, or extend the system to a second team. This line is where off-the-shelf and custom diverge most sharply.

A comparison that counts the first line and ignores the other seven is missing two thirds of the picture. The rest of this post counts all of them.

The off-the-shelf cost curve

Off-the-shelf agentic AI tools have a low entry price and a fast time to first value. Sign up this afternoon, configure it tomorrow, productive by the end of the week. For many small businesses this is the correct first move and the rest of this section should not put anyone off doing it.

The curve starts gentle. A mid-market AI customer-support tool runs at roughly £60 to £90 per seat per month. Two seats over the first year is around £1,800. Add some basic setup time and a Zapier-tier subscription to wire it into your accounting software, and year one all in lands somewhere between £2,500 and £4,000 for a small business doing modest volume.

What changes the picture is metering. Most agentic AI platforms charge per resolved conversation, per AI-handled message, or per agent action. The unit prices look small individually (£0.30 to £0.50 per AI conversation is typical) but compound quickly. A business handling 1,000 AI-resolved customer conversations a month is looking at £4,800 a year in metering alone, on top of the seat cost. The seat price was never the issue.

Two other curves matter. The first is the tier ceiling. Most platforms reserve the features you actually need at scale (advanced routing, custom workflows, deeper integrations, audit logs) for higher tiers, and you are pushed up within the first two years. The second is the standard 10 to 15 percent annual list-price increase that most SaaS products quietly apply.

Off-the-shelf is cheap to start. It scales less cheaply than the seat-price suggests.

The custom cost curve

A bounded custom AI agent has an inverted curve: meaningful upfront cost, then a much flatter line.

The upfront cost is the build itself, plus the integration into your systems, plus the scoping work that decides what to build in the first place. For a single SME workflow (something like customer-email triage, or trade-order processing, or supplier communication logging), a scoped custom build typically lands between £3,500 and £8,000 depending on the integrations and the precision required. The scoping step (a structured day on-premises producing a costed plan) sits at £1,495 on the MoonBoots AI Blueprint Day page.

After the build, the run cost is small. Hosting an agent on a modern serverless platform runs around £20 a month for an SME-scale workload. Claude API usage for the volumes a 35-person business generates is typically £40 to £80 a month. A maintenance retainer that covers prompt tuning, edge cases, and the occasional new category usually sits around £150 a month from month four onwards. That is roughly £2,500 a year in total running cost, with no per-message tax.

The risk is that a custom build becomes a money pit if it is unbounded or poorly scoped. A team that says yes to every feature request, or starts building before deciding which one workflow to instrument first, can spend twice the budget for half the value. This is why the project starts with a scoping step and why Custom Business Platforms are sized to a clear problem, not an open-ended ambition.

The bounded approach is not a brand slogan. It is the difference between a custom build that earns its cost and one that does not.

A worked TCO model

Here is an illustrative 36-month model for a real-shaped UK SME. The figures are illustrative and rounded to the nearest hundred. They are not a MoonBoots quote; MoonBoots pricing lives on the Custom Business Platforms and AI Blueprint pages.

The business. Fenland Foods Ltd is a fictional 35-person specialist food producer in Wisbech, Cambridgeshire. Family-owned, operating for 22 years. Turnover £4.2 million. 60 SKUs across chutneys, preserves, and seasonal gift ranges. Orders arrive via Shopify, phone, and email.

The workflow. Customer email triage and drafted-reply assistance. The business receives around 2,000 customer emails a month, of which roughly half (around 1,000 in early months, rising to around 1,250 by year three) need an AI-assisted response. The other half are straightforward, auto-categorise, or pass directly to humans.

The horizon. 36 months.

Cost line	Off-the-shelf (36 mo)	Custom build (36 mo)
Setup or build	£600	£4,000
Scoping (Blueprint Day or equivalent)	included	£1,495
Seats and subscription	£6,360	£0
Per-message metering	£16,200	£0
Hosting and model usage	£0	£2,460
Integration tooling	£1,080	£1,200
Maintenance and iteration	included	£4,950
Switching cost if you leave	£2,000 to £4,000	£0 (you own it)
36-month total	£26,240 to £28,240	£14,105

Off-the-shelf workings. Two seats at typical mid-market rates of £75 per seat per month, accounting for the standard 10 to 15 percent list-price drift over the period and the tier upgrade that volume usually forces by year two. Per-message metering at £0.40 per AI-resolved conversation, averaging around 1,125 conversations a month over the 36 months. Zapier-tier integration at £30 a month to wire the tool into Sage and Shopify. Switching cost is the typical industry figure for moving away from a configured AI helpdesk product (template rebuild, integration teardown, knowledge export, team retraining).

Custom workings. The build figure (£4,000) is the midpoint of the email-assistant range used in the Fenland Foods Blueprint Day worked example. The scoping figure (£1,495) is the live AI Blueprint Day price. Hosting plus Claude API usage at this volume runs around £68 a month over 36 months. Sage and Shopify integration is one-off connector work built into the assistant, with no ongoing third-party subscription. Maintenance retainer at £150 a month from month four onwards covers prompt tuning, edge cases, and occasional new categories.

The cross-over. Off-the-shelf wins for the first year because the upfront cost is much lower. The cross-over (the month at which the custom path's cumulative total falls below the off-the-shelf cumulative total) happens around month fourteen, driven by the per-message metering on the off-the-shelf side. From month fourteen onwards the custom path is ahead, and the gap widens every month after that. By month thirty-six it is ahead by roughly £10,000.

The honest caveat. This finding is volume-driven. At very low volumes (under around 200 AI-resolved conversations a month), the metering line on off-the-shelf is small and the seat cost dominates. In that case off-the-shelf stays cheaper for two or three years easily. The cross-over only kicks in at meaningful volume. The next two sections name when each path is the right answer in plain terms.

When off-the-shelf is the right answer

Off-the-shelf agentic AI is the correct first move when:

• Volume is low. Under around 200 AI-resolved interactions a month, the metering tax is small and the per-seat cost dominates. The numbers stay in your favour.
• The workflow is generic. Customer support email triage, meeting transcription, document summarisation, basic CRM enrichment. If a mature product already does the thing you need, building your own is reinventing the wheel.
• Integration depth is shallow. If your business runs on a small number of standard tools (Shopify, Xero, Mailchimp) and the off-the-shelf product has native connectors, the integration friction is honestly low.
• You need it working this week. A custom build, however well scoped, has a minimum lead time of four to six weeks. If the answer is needed now, off-the-shelf is the only honest path.
• The team has no appetite to maintain anything. A custom system is a small piece of software your business owns. Someone has to care for it. If that is not realistic, the maintenance burden of a custom build will outweigh its TCO advantage on paper.

A small business that ticks four of those five should buy, not build. There is no shame in this, and an honest consultancy will tell you so. The companion blog post Custom AI Agents vs Off-the-Shelf sets out the decision framework in more detail.

When a custom build earns its cost

A bounded custom build earns its cost when:

• Volume is meaningful. Above around 500 AI-resolved interactions a month, the per-message metering on off-the-shelf compounds fast. A custom build with no metering and a flat run-cost overtakes inside year one.
• The workflow is specific to your business. If off-the-shelf forces compromises (the tool's tagging schema does not match yours, the routing logic cannot do what you actually need, the response tone is wrong) you will spend more time working around the tool than the tool is worth.
• Integration depth is real. If the AI needs to read your Sage data, write to your bespoke order system, or look up information in your specific customer database, deep custom integration is cheaper and more reliable than the third-party connector tax.
• The horizon is multi-year. A custom build amortises over time. If you are committing to this workflow for the next three years or more, the upfront cost is recovered well within the horizon.
• You need bounded behaviour. If the AI is going to be in front of customers, children, regulated data, or anything where its mistakes have real cost, you need control. Off-the-shelf tools give you very limited control over what the model will and will not do. See the Bounded AI post for what that means in practice.

A business that ticks three of those five is in the territory where custom earns its cost. An AI Blueprint Day is the standard way to scope it properly before committing.

The costs nobody puts in the quote

Several costs sit outside both the off-the-shelf and the custom quote and matter as much as anything in the table:

Agent sprawl. A business that adopts one AI tool tends to adopt three more within a year. Each one has its own seat cost, its own integration, its own learning curve. The cumulative subscription bill is rarely audited and is often double what anyone realises.

The maintenance cliff. Off-the-shelf vendors push price increases and feature changes on their own timeline. A custom system that has been left untouched for a year accumulates prompt drift, model-version churn, and dead code. Both paths have a maintenance cost. The difference is who controls when it happens.

The supervision tax. An unbounded agent that occasionally surprises you with what it does in production needs supervision. The internal time cost (someone reviewing flagged conversations, handling escalations, fixing the things the AI got wrong) is real and rarely captured in TCO models. Bounded systems reduce this cost; off-the-shelf tools rarely give you the controls to bound them.

Vendor pricing changes. Several major SaaS products in this space have raised prices by 30 to 50 percent inside a single renewal cycle. If you are locked into the platform, you absorb the increase. If you own the build, you do not.

How to actually decide

A real TCO decision starts with a few specific numbers about your own business, not a generic comparison.

Five questions to answer honestly:

1. How many AI-handled interactions per month is realistic for this workflow at steady state?
2. Is the workflow generic enough that a mature off-the-shelf product already handles it well?
3. How deep is the integration into your existing systems?
4. What is your horizon for this workflow: 12 months, 36 months, indefinite?
5. Who in your business will own the system once it is running?

If you have clear answers to those five, you have most of what a decision needs. If you do not, two MoonBoots products are designed for exactly this step. The AI Readiness Scorecard is a free four-minute self-assessment that names where your business sits. The AI Compass is a fixed-fee three-hour structured session that produces a written one or two page summary inside 48 hours and points you at the right next step. Both are designed to make the decision clearer, not to push you towards a build.

Frequently asked questions

How much does a custom AI agent cost for a UK SME?

A scoped custom AI agent built for a single workflow at SME scale typically costs £3,500 to £8,000 to build, plus £150 to £250 per month to run and maintain. The exact figure depends on the workflow, the integrations, and the data the agent needs to handle. The MoonBoots Blueprint Day worked example puts a customer-email triage build at £3,500 to £4,500.

Is off-the-shelf agentic AI cheaper than building custom?

Cheaper to start, yes. Cheaper over 24 to 36 months, often no, once per-message metering and per-seat charges compound. At volumes above roughly 500 AI conversations per month, a custom build usually overtakes off-the-shelf inside the first year.

How long until a custom AI build breaks even?

At meaningful volume (around 1,000 AI conversations per month), the cross-over with off-the-shelf is typically around month fourteen, early in year two. At lower volumes the break-even pushes out further. At very low volumes a custom build may not pay back inside the comparison horizon at all.

What are the hidden costs of agentic AI platforms?

Per-message or per-conversation metering, forced tier upgrades when you hit a feature ceiling, integration tooling between the AI tool and your business systems (Sage, Shopify, your CRM), internal time supervising an unbounded agent, and the switching cost when a vendor changes pricing or you decide to move. Most quotes show the seat price and stop there.

When should a small business choose off-the-shelf over custom AI?

When volume is low, the workflow is generic and well served by an existing tool, integration depth is minimal, the team has no appetite to maintain a custom system, or you need something working this week. Off-the-shelf is the correct first move for many small businesses and should not be apologised for.

What is included in the total cost of ownership for agentic AI?

Licence or subscription, per-action or per-message metering, integration and data plumbing, ongoing maintenance and model or prompt updates, internal staff time, the cost of agent errors in production, vendor lock-in and switching cost, and the cost of scaling seats or volume. A comparison that only counts licence cost is missing two thirds of the picture.

A closing note

The honest answer to "what does agentic AI cost" is "it depends on your volume, your workflow, and your horizon". This post has tried to make those dependencies concrete enough that a UK SME owner can run the numbers on their own situation.

If you want to run this model against your own figures, the AI Readiness Scorecard is free and takes four minutes. The result is yours to keep regardless of whether you ever speak to us. The first conversation, when it comes, is always free and never wasted.