You’ve typed “how much does AI cost” into a search bar and found answers ranging from “£25 a month” to “£75,000 for a full implementation”. Both figures are real. The trouble is they are answering different questions, and until you know which applies to your business, you cannot set a useful budget.
The spending splits into two things that often get conflated: the software licence and the implementation work needed to make that software useful in your actual operations. They have different price tags, different timescales, and different failure modes.
What is the choice you’re actually facing?
AI spending for an owner-managed business falls into two distinct buckets. The first is software licences, tools like Microsoft Copilot at £24.70 per user per month or Google Workspace AI at similar prices. The second is implementation: connecting those tools to your systems and getting them used. The confusion starts when businesses set a budget before working out which bucket their problem actually requires.
For a business with one to ten people, a realistic structured adoption budget covering off-the-shelf tools, internal time, and light training runs to roughly £2,000-£10,000 over three to six months, based on gigCMO’s analysis of owner-managed business AI deployments. For a business with ten to fifty people that needs AI connected to a CRM, accounting system, or case management platform, the total typically runs to £15,000-£75,000 over six to nine months.
If your mental budget sits somewhere between those two bands, you are probably blending a software question and an implementation question without realising it.
When does off-the-shelf AI give you what you need?
Off-the-shelf tools are the right call when your team’s most time-consuming tasks already live in email, documents, spreadsheets, or a CRM. Adding AI licences to what you already have is often sufficient. You’re paying for productivity on top of familiar tools, not for a new system. Many owner-managed businesses with under ten staff can stay in this lane for months without needing anything more complex.
The US Small Business Administration recommends starting with AI features already built into your existing tools, many of which offer free or low-cost entry tiers. Cornell Design Group’s 2026 pricing guide puts the typical running cost for this kind of setup at $50-500 per month across licences, a workflow tool such as Zapier, and modest API usage.
The main investment at this level is internal time: learning new tools, adjusting prompts, and redesigning a handful of repetitive tasks. The situations where off-the-shelf stops being enough are worth knowing. High-volume processes where staff manually copy AI outputs between systems add friction quickly. Any process that needs consistent governance across all outputs, such as client-facing communications or compliance checks, will eventually need something more structured.
When do you need integration and implementation support?
When AI needs to touch your core systems, off-the-shelf licences are not enough. If the goal is to reduce time on claims triage, contract drafting, client onboarding, or quoting, and those processes currently involve data from your CRM, case management system, or accounting platform, you are looking at integration work. That requires a consultant or developer, and a budget that reflects it.
gigCMO’s recommended budget split for businesses in this bracket is revealing: 40% to integration and data work, 30% to licences and infrastructure, 20% to training and change management, and 10% to ongoing operations. That split matters because buyers often receive a quote that covers licences and initial setup, then find the remaining work arrives as change requests.
For firms in regulated sectors, the implementation cost includes additional layers. The FCA has stated that existing obligations on governance, consumer protection, and operational resilience apply to AI systems, including those sourced from third-party providers. Financial services firms will typically need model validation and oversight documentation alongside the standard implementation work, and that takes time and budget.
A consultant-led pilot covering one to two use cases typically costs £8,000-£40,000. A full multi-function rollout for a firm of ten to fifty people runs to £15,000-£75,000 over six to nine months, with ongoing support retainers adding 10-20% of build cost annually. The 40/30/20/10 budget split is a useful sense check for any proposal you receive.
What does it cost to get this call wrong?
Getting the choice wrong usually runs one way: committing to integration work before the simpler option has been tested. gigCMO’s analysis puts the failure rate at around 70% for owner-managed business AI deployments that never reach production, with cost overruns of 20-70% common in those that proceed. For a firm spending £40,000 on a rollout, a failed project means writing off the full budget and several months of management time.
The regulatory dimension adds to this. The Information Commissioner’s Office requires organisations using AI to process personal data to conduct a Data Protection Impact Assessment and to address fairness, transparency, and explainability. Where AI makes decisions with legal or significant effects on individuals, data subjects have rights to human review under Article 22 of the UK GDPR. These steps carry enforcement risk if skipped, and they are not optional for any business using AI on customer or staff data.
The National Cyber Security Centre’s guidance on AI system development adds a security dimension: integrating AI into business systems expands the attack surface, including through API exposure and prompt injection. Those risks need a proper review, not an assumption that the vendor has handled them.
The reverse failure mode is also real. Buying licences for a problem that genuinely needed integration means spending three months wondering why nothing has changed. Less expensive, but an equal waste of the opportunity.
What should you ask before you commit the budget?
Before signing anything, you need answers to three categories of question: scope and ROI, vendor quality, and compliance. The most commonly skipped is the first. If you cannot name the specific processes the AI will change, estimate the staff hours they consume each week, and describe a measurable outcome, you are not ready to buy. The rest of the questions only matter once that one is settled.
On scope: name the specific processes, by name. Estimate the staff hours those processes take each week. Can you run a contained pilot in four to eight weeks with a defined go/no-go decision? Cornell Design Group’s guide recommends this structured pilot approach as the standard way to de-risk implementation spend.
On vendors: do they have live UK references in businesses of a similar size? Is the project priced at a fixed fee or time and materials, and what assumptions trigger change requests? Who owns the prompts, workflows, and any customised models once the engagement ends? The Competition and Markets Authority has highlighted vendor lock-in as a genuine risk in AI procurement. Paying a small premium for solutions that support portability and open standards can be worth it over a two to three year horizon.
On compliance: will this AI process personal data? If so, budget time for a Data Protection Impact Assessment per ICO guidance. Are any decisions likely to have legal or significant effects on individuals, potentially engaging Article 22 UK GDPR? Where is data stored, and does any of it leave the UK? For financial services firms, add the FCA’s model governance expectations to the checklist before the first invoice arrives.
If you can answer all of those questions before the work begins, you have a workable brief. If you cannot, keep asking until you can.
