The demo looked good. The AI handled the use case cleanly, the pricing seemed reasonable, and the salesperson answered every capability question confidently. The questions that got skipped came up later, when the contract was already running: where the data would live, who the data processor would be, what the exit terms looked like.
That pattern is common. Owner-managed businesses tend to evaluate AI suppliers on features first and get to governance questions late, or not at all. The result is contracts harder to exit than expected, data sitting in jurisdictions the owner never agreed to, and accountability gaps that stay with the business owner rather than the vendor.
This guide covers the three distinct shapes of AI supplier decision, when each makes sense, and the questions that separate a workable deal from an expensive one.
What choice are you actually facing?
Three categories of AI supplier sit in front of many owner-managed businesses at some point: a narrow point solution for one bounded job, a general platform such as Azure OpenAI or Google Vertex that you build multiple workflows on, or a bespoke build tailored to your specific processes and compliance needs. Which shape fits depends on your use cases, your data sensitivity, and how quickly those uses are likely to grow.
The governance implications vary significantly across these three shapes. Point solutions typically ship with standard UK GDPR terms, a data processing agreement you may or may not read carefully, and a privacy policy that has already opted you into something you might prefer not to be opted into. Platforms require you to configure data handling explicitly: which regions data transits through, whether your inputs are used to improve the model, and which teams inside the vendor organisation can access your requests. Bespoke builds shift responsibility further. You are the architect, so decisions about data residency, access controls, and audit logs sit with you from day one.
The British Chambers of Commerce and Cisco found in 2023 that 48% of UK owner-managed businesses cited lack of understanding as their primary barrier to AI adoption. That figure reflects a features gap. The governance gap is less visible and typically more expensive when it materialises.
When is a point solution the right call?
A narrow AI tool makes sense when you have a clearly defined, bounded workflow, data sensitivity is low to medium (anonymised support logs, marketing content, publicly available information), and you want to be operational in weeks rather than months. Typical SaaS AI tools in this tier run from £20 to £100 per user per month, implementation risk is low, and cancellation is straightforward if the tool does not deliver.
What makes point solutions attractive is also what makes them risky at scale. Proliferation is the main hazard: three tools become six, each with its own data processing agreement, its own retention policy, and its own sub-processors operating in regions you have not audited. The ICO is explicit that as the data controller, you remain accountable for what your processors do with personal data, regardless of how many there are or how small each integration appears.
Before adding a point solution to your stack, read the data processing agreement rather than just the privacy policy, confirm where data is stored and whether your inputs are used for model training, and log the decision in your data protection register if you are keeping one. Many point solutions offer opt-outs on training data; it is rarely the default setting.
When is a general platform the right call?
A general platform becomes the right call when you expect multiple AI use cases across the business in the next 12 to 24 months and have the technical capability to build on APIs, either in-house or through a development partner. Azure OpenAI, Google Vertex and the OpenAI API give you reusable building blocks, more control over data routing and model versions, and the ability to standardise governance across a single vendor relationship.
The trade-off is depth of dependency. The Competition and Markets Authority’s review of AI foundation models flagged the risk of envelopment strategies, where AI capabilities get bundled into existing productivity suites and exit becomes materially costlier over time. Ofcom’s cloud market study found businesses reporting five-figure bills to move workloads out of one provider to another. The same switching economics apply to AI workflows built heavily into a single platform’s stack.
Bespoke builds sit at the far end of this spectrum. They make sense when use cases are genuinely unique, when off-the-shelf tools cannot meet domain or data residency requirements, or when the regulated nature of the decisions involved (credit scoring, clinical triage, recruitment screening) demands explainability and audit trails that commercial tools cannot provide. The investment threshold is typically a six-figure benefit case, which rules out many owner-managed businesses unless the use case is genuinely high-value and the regulation demands it.
What does getting it wrong actually cost?
Getting the supplier choice wrong costs more than the licence fee you end up writing off. IBM’s 2023 Cost of a Data Breach report found that third-party supplier breaches cost an average of $4.76m and take 12.5% longer to identify and contain than internal breaches. The ICO’s enforcement record confirms that you cannot outsource accountability: the data controller stays liable regardless of which processor caused the breach.
On regulatory exposure, the FCA expects regulated firms to maintain full accountability for any outsourced technology under its operational resilience rules, and that includes AI. For IFAs, lenders, or brokers using AI for credit scoring or compliance monitoring, the expectation covers testing, governance, validation, and documentation, whether the model is built in-house or supplied by a third party.
The Experian enforcement notice in 2020 resulted in a £20m fine centred on data governance failures in a third-party data environment. The British Airways breach in 2017 and 2018, which involved third-party vulnerabilities and inadequate controls, drew an initial proposed fine of £183m (ultimately reduced to £20m) plus compensation claims. In both cases, the contractual relationship with the supplier did not insulate the data controller from enforcement.
Lock-in adds a separate financial dimension. The CMA’s review of AI foundation models found that dependence on a small number of providers reduces buyer bargaining power and makes switching expensive, particularly when AI workflows are embedded in productivity tools the business already depends on for everything else.
What should you ask before you sign anything?
Before signing any AI supplier contract, ask five questions: where your data is stored and whether it crosses UK or EEA borders; what security certifications they hold; whether you can opt out of model training; what the data export format and egress cost looks like if you leave; and whether they can name a reference customer in your sector who will take a call.
On data location: the ICO requires Transfer Risk Assessments for personal data moving outside the UK or EEA. A supplier who cannot tell you where your data is processed fails the first test.
On certifications: the NCSC recommends preferring suppliers who hold ISO 27001 or Cyber Essentials Plus and can provide penetration test summaries on request. Asserting security and evidencing it are different things.
On model training: OpenAI, Microsoft, and Google have all introduced opt-out controls following regulatory pressure, but these are not uniformly applied across product tiers. Free and lower-paid tiers frequently include training use by default; enterprise tiers generally do not. Get the confirmation in writing, specific to the product tier and version you are actually buying.
On exit terms: the CMA’s review found that egress fees and proprietary data formats are the main switching barriers in AI and cloud platforms. Ask for data export in open formats (JSON, CSV) as a contract term, not an afterthought.
On references: a supplier without UK or EU customers in your sector cannot demonstrate domain-appropriate compliance. Logos on a case studies page without contact names available are not references.
If you are about to shortlist AI suppliers, run these questions before the demo, not after. The features will be compelling. The accountability questions are where the real cost lives.
