A founder running a 15-person consulting firm books a vendor demo for an AI tool that will help the team produce client proposals faster. The demo is impressive. The pricing slide shows £600 per month for the team licence. Onboarding, the vendor says, takes two weeks. She signs up for a pilot at £8,000. Six months later, she is trying to work out why the project has cost her firm closer to £35,000 and is still not reliably embedded in how the team works.
She is not alone in this. The vendor price she saw was real. The costs she didn’t see were also real. That gap between the opening quote and the eventual bill is one of the most consistent findings across AI project research, and understanding it before you commit is worth the time.
What actually goes into the cost of an AI project?
The licence fee on a vendor’s pricing page covers one thing: access to the software. Four further cost layers sit underneath it: cleaning your data so the tool can use it; connecting the AI to your existing systems; training your team; and ongoing maintenance to keep outputs reliable. Studies of professional services deployments find the total consistently landing 3 to 5 times above the opening quote.
Dan Cumberland Labs, which works with mid-market professional services firms, reported in 2024 that enterprise implementations typically cost 3 to 5 times the initial estimate “when you account for integration, customisation, and operational overhead.” A £40,000 pilot, on that reckoning, implies a realistic total of £120,000 to £200,000 to reach a stable, embedded solution. Harvard Business School Online’s primer on AI implementation notes that model development is usually the minority of total cost; the bigger expense is “infrastructure to support it, systems to integrate it, and people to maintain and improve it.”
For firms in regulated sectors, compliance adds a further layer. UK GDPR requires a lawful basis for any personal data you feed into an AI tool, and organisations using AI for decisions with legal or significant effects on individuals may trigger Article 22 safeguards, including the right to human review. The ICO demonstrated its willingness to act in 2022 when it fined Clearview AI £7.5 million for unlawful biometric data processing. For firms in financial services, the FCA makes clear that regulatory responsibilities remain with the firm regardless of what a third-party tool does.
Why does the cost gap catch owner-operated firms off guard?
The decision to proceed gets made on the opening number. The full bill arrives later, often in stages that feel manageable at the time. By then, internal resource has been committed, the integration is half-built, and the cost of stopping is as uncomfortable as the cost of continuing. Owner-managed businesses are particularly exposed because they rarely have a dedicated project team to absorb overruns and are usually running the engagement alongside their day jobs.
The risk compounds if the project fails to deliver. Gartner’s research found that a large proportion of AI projects fail to reach stable deployment, primarily because of unclear business value, data quality problems, and poor scoping. IBM’s 2024 analysis found that average computing costs in AI-adopting organisations are expected to rise 89 per cent between 2023 and 2025, with generative AI as the primary driver. Even a flat licence fee can conceal a variable infrastructure bill that grows with usage, a separate cost that vendor demos rarely emphasise.
Where do these costs actually show up?
The overruns in AI projects tend to surface in three places: the time spent preparing your data before any tool can use it, the work needed to connect the AI to the software your business already runs, and the hours your team spends learning, supervising, and correcting the system. Each of these is largely invisible in a vendor quote.
Data preparation is where the underestimation bites hardest. Cumberland Labs estimates it consumes 50 to 70 per cent of project time and 15 to 35 per cent of costs, even in firms that already have reasonably tidy data. For a 15-person services firm, that typically means extracting documents from shared drives and email, removing personal data that cannot be fed into third-party tools under UK GDPR, and standardising formats across spreadsheets, PDFs, and CRM notes. That is skilled, time-consuming work and it almost never appears in a software vendor’s opening price.
Integration costs are a second underestimated layer. If your AI needs to connect to legacy accounting software, a specialist case-management system, or a custom database, Cumberland Labs found that integration typically costs 2 to 3 times more than a greenfield build where everything is cloud-native. A vendor quote for £80,000 on a fresh deployment can reasonably become £160,000 to £240,000 once complex integration is factored in. When a vendor tells you integration is straightforward, ask for that in writing and request a fixed-price or capped-time arrangement before you sign.
People and ongoing costs complete the picture. Staff training, process redesign, and the internal time spent reviewing AI outputs before they reach clients are real costs that rarely appear in proposals. Once the system is live, plan for 15 to 30 per cent of infrastructure cost per year for monitoring, retraining, and correcting for model drift: the gradual degradation in output quality that happens as the underlying data patterns in your business shift over time.
When does the full-cost picture apply, and when doesn’t it?
The 3 to 5 times multiplier applies when you are building something that requires your data, your systems, and your specific context. The more the AI needs to read your client files, connect to your line-of-business software, or produce outputs that feed into regulated workflows, the more the layers beneath the licence price compound.
The picture changes for simple, bounded use cases. A Microsoft 365 Copilot licence used for drafting emails and summarising documents has minimal integration overhead. AI transcription tools used only for meeting notes are close to commodity territory, where competition has driven prices down and setup is straightforward. If you already have clean, well-structured data in a single system, your data preparation costs will be materially lower than average. The hidden-cost argument is real; it applies at the custom and integrated end of the buying spectrum, not across every AI purchase a firm might make.
What other cost and contract questions connect to this?
Understanding the five cost layers changes the questions you ask at the sales stage: ask for integration costs to be quoted separately, ask how data preparation will be resourced and by whom, and ask what ongoing support is included and at what price. None of these questions are aggressive; they are simply the ones a well-prepared buyer asks. Asking them upfront shifts the risk profile of the project before you commit.
Related posts in this series cover how to estimate total cost of ownership before you sign, what to look for in AI vendor contracts, and how UK GDPR shapes what you can and cannot feed into third-party tools. For firms in regulated sectors, the ICO’s guidance on AI and data protection and the FCA’s discussion paper on AI in financial services set out what staying within your existing obligations actually means in practice. The EU AI Act, now in force, adds a further consideration for any UK business serving clients in Europe. These are worth reading before the pilot contract lands on your desk. If you’re currently scoping an AI project and want to pressure-test the numbers before committing, Book a conversation.
