What is the true five-year cost of ownership for a custom AI solution?

17 October 2026

What should a five-year budget include?

A custom AI project has at least seven cost lines: discovery and design, the initial build, system integrations, data preparation, hosting and model usage, ongoing support, and staff adoption. The purchase order for the build covers only part of the bill. If an agency quotes £20,000 for development, ask what happens when your CRM changes, the model provider changes its terms, a staff member needs a new permission, or the assistant begins giving poor answers. Those are ordinary operating events, not exotic failures.

For a focused UK SME workflow, a useful planning illustration is £25,000 for discovery, build and launch, then £9,000 a year for hosting, model use, monitoring, support and training. Five years costs £70,000 before material upgrades. A more integration-heavy system might cost £60,000 to build and £25,000 a year to operate, or £185,000 over five years. These are planning examples, not market quotations or Precise Impact AI prices. Obtain a scoped quote against your own data, users and service level.

The Office for National Statistics found in its July 2026 analysis that AI use among UK businesses with 10 or more employees rose from roughly 12% in late 2023 to roughly 35% by June 2026, but only 10% of adopters reported extensive use. That gap matters for budgets: buying or building a tool does not guarantee that people will incorporate it into daily work. Set aside time and money for adoption, not just software. Source: ONS, Artificial intelligence in UK businesses: 2023 to 2026.

How much do running and support costs add?

Model charges can be tiny for a low-volume internal assistant and substantial for a high-volume system that reads long documents, processes images or retries failed calls. Ask for a workload estimate in documents, requests, users and peak periods, then price at low, expected and high demand. Include storage, backups, logging, monitoring, security tools and any third-party licences. A £200 monthly test environment is not evidence that production will also cost £200.

Support should have an owner and a budget. Someone must investigate bad answers, maintain integrations, update permissions, test provider releases and respond to incidents. For a modest workflow, allow an illustrative £300 to £1,000 a month for technical support and governance, plus actual usage charges. A business-critical service with out-of-hours cover may cost much more. If the provider promises free lifetime support, ask exactly what is included, the response time and who pays for change requests.

Budget for the human side as well. A manager who spends two hours a week checking outputs spends around 100 hours a year doing so. At a fully loaded £35 an hour, that is about £3,500 a year, or £17,500 over five years. That review may be exactly the right safeguard, but it belongs in the economic case. Separate cash spend from staff time so the board can see both. A system that saves 12 hours a week but consumes eight hours of checking is not delivering the headline saving.

Which hidden costs are most often missed?

Data is the first hidden cost. If product codes, customer names or policies are inconsistent, the system needs cleaning rules, ownership and a process for correcting source material. The second is integration. An API that works in a demonstration may not support the permission model, audit trail or exceptions your live process needs. The third is change. Vendors alter APIs, your business changes workflows, and new staff need training. An annual improvement allowance of 10% to 20% of the initial build value is a sensible scenario to test, not a universal tariff.

Compliance can be another real operating cost. Where personal data is involved, document the lawful basis, retention, access controls and whether a data protection impact assessment is needed. The Information Commissioner’s Office guidance explicitly addresses accountability, transparency, lawfulness, accuracy and fairness through the AI lifecycle. Its guidance is under review following the Data (Use and Access) Act, so check the current version before a new deployment. Source: ICO guidance on AI and data protection. Budget staff or specialist time for those controls instead of treating them as paperwork after launch.

Finally, price an exit. You may need to export records, replace a model, migrate prompts and workflows, or retrain staff on a replacement system. Ask the supplier who owns the code, prompts, configuration, logs and derived data, in what format you can export them, and how much transition help costs. If those answers are missing, the attractive year-one price may conceal expensive lock-in.

What does a realistic five-year example look like?

Consider a UK service business with 25 employees building an internal assistant to classify incoming enquiries, retrieve approved answers and draft responses for human approval. Suppose discovery and data work cost £6,000, development and integration £19,000, and launch training £3,000. Initial cost is £28,000. Suppose hosting and model use average £300 a month, support £500 a month, annual retraining and process review £2,000, and an upgrade allowance £3,000 a year. Annual cash operation is £14,600. Five-year cash ownership is £101,000: £28,000 plus five times £14,600. Those numbers are an illustrative worksheet, not a quoted offer.

Now include a supervisor spending three hours a week on quality review at a fully loaded £35 an hour. That is about £5,460 a year or £27,300 over five years, taking the economic cost to approximately £128,300. If the system saves 15 hours a week of work worth £25 an hour for 48 weeks, the gross annual time value is £18,000. Five years is £90,000. On those assumptions the project does not break even on labour time alone. You would need to value faster responses, fewer missed enquiries or better consistency, reduce operating cost, or choose a simpler tool.

Change one assumption at a time. What if usage triples? What if only half the expected hours are saved? What if the CRM integration needs rebuilding in year three? A business case should show a base case and a downside case. Never quietly count every minute of theoretical saving as cash. Staff capacity becomes cash only when you avoid external spend, avoid hiring, increase output or improve retention in a measurable way.

Would off-the-shelf software be cheaper?

Often, yes. Microsoft 365 Copilot, ChatGPT Business and automation platforms such as Make can cover generic drafting, summarisation or simple routing without a bespoke build. Compare their current UK pricing and data terms directly because plans and rates change. An off-the-shelf tool wins when the workflow is common, the integration is supported, the business can accept the vendor’s product roadmap and there is little competitive value in owning the implementation. A custom system wins when narrow permissions, specialist data, unusual hand-offs or differentiated customer experience justify the extra ownership burden.

Run a four-week test with the standard tools first. Record the number of transactions handled, time saved after review, errors, rework and staff satisfaction. If a £50-a-month subscription and a straightforward automation cover the requirement, do not buy a £50,000 bespoke system. If the off-the-shelf route forces people to copy confidential data manually, breaks at every exception or cannot produce the audit trail you need, a tailored build may be justified.

The comparison must use five-year totals for both routes. Subscriptions rise, internal configuration consumes time, and replacing a vendor can also be costly. Conversely, custom software needs ongoing technical ownership. Ask both suppliers for an itemised implementation price, year-two through year-five charges, usage limits, service levels, security responsibilities and exit arrangements. The cheapest first invoice is not necessarily the cheapest operating model.

When is a custom solution not right for you?

Do not commission a custom build if you cannot name the process owner, the baseline cost of the current work, the data source, the human approval point and the expected volume. Do not build if the underlying process changes every week or if the source records are too unreliable to support safe decisions. Fix the process and data first. A custom model is also unnecessary when an existing software feature solves the problem adequately.

Start with a smaller pilot if the five-year downside case exceeds what the outcome could reasonably return. Put a stop rule in the contract: if the pilot fails agreed accuracy, time-saving or adoption thresholds, pause before paying for a full rollout. For regulated or high-impact decisions, obtain appropriate legal and data protection advice and keep meaningful human oversight. If nobody in your organisation can own the system after launch, either pay for a clearly defined managed service or avoid creating a dependency you cannot maintain.

If you are ready, ask for a five-year cost schedule with each line separated into fixed charges, usage-based charges and optional changes. Include the assumptions behind every number. A supplier willing to show those assumptions is easier to work with than one who only quotes the exciting build phase. If you want an independent sense-check of that schedule, we can discuss it without assuming a bespoke build is the answer.

Is This Right For You?

A custom build is a sensible option when a measurable, repeated workflow needs specialist integration, narrow permissions or features that standard products cannot deliver. Give it a named owner, a budget for ongoing support and a clear stop rule.

It is not right if the process is unmeasured, the data is unreliable, standard software already handles it, or the downside budget is larger than the likely benefit. Test the simpler option first.

Frequently Asked Questions

Is maintenance included in the build price?

Usually not unless the contract explicitly says so. Ask for included hours, response times, security updates, integration fixes and rates for change requests.

What percentage of the initial cost should I reserve each year?

Test a scenario with annual improvements of 10% to 20% of build cost, then add hosting, model usage, support and training separately. This is a planning scenario, not a standard supplier rate.

How do I estimate model usage charges?

Measure expected requests, document sizes, users, peak demand and retries in a pilot. Price low, expected and high usage against the provider’s current tariff.

Do I need to include staff review time?

Yes. Multiply review hours by a fully loaded hourly cost. It may be necessary for quality and safety, but it is still part of ownership.

What happens if we change AI model provider?

A modular system can reduce migration work, but prompts, evaluations, permissions and integrations still need retesting. Put export rights and transition help in the contract.

Can a subscription be better value than custom AI?

Yes. If a standard tool meets the need safely and reliably, a subscription often costs less and avoids maintenance ownership. Compare five-year totals, not first invoices.

Should we capitalise or expense the build?

That depends on your accounting facts and policy. Ask your UK accountant to classify development, subscriptions and ongoing support correctly rather than assuming all costs receive the same treatment.