The owner of a twelve-person professional services firm books three AI software demos in one week. The first vendor shows a polished video. The second claims their tool integrates with everything. The third offers a free trial. By Friday, she has a shortlist built on price and gut feel, with no clear idea what she’s actually comparing.
That’s the default buying pattern. The tools are plentiful, the demos are impressive, and the criteria are vague. Getting the criteria right before the demos changes the decision entirely.
What should you define before you look at any tool?
The firms that extract the most value from AI begin with a concrete list of use cases, not a list of tools. McKinsey’s 2023 global AI survey found that organisations capturing the highest AI value focus on a few high-impact use cases and scale them, rather than deploying generic tools. The UK Government’s AI Playbook follows the same logic: define the problem before you select any technology.
For a services firm with ten to fifty staff, practical starting points include drafting and polishing emails and proposals, summarising meeting notes, and first-pass review of documents. A 2023 NBER study of 5,000 call-centre workers found a 14% average productivity gain when staff used a generative AI assistant on content-heavy tasks. That’s a realistic benchmark for what good looks like on writing-heavy work.
For each use case, write a brief job story: when does this work start, what should AI produce, and how will you know it worked? Three to five of these stories become your evaluation criteria. The demo is then testing the tool against your requirements, rather than showcasing the vendor’s strongest features.
Why does data protection sit above price in this decision?
Under UK GDPR, your firm remains the data controller whenever you configure and deploy an AI tool, even if the vendor hosts everything in the cloud. The ICO makes this explicit: choosing the tool, setting the configuration, and deciding what data flows through it are all your responsibility. The Samsung 2023 incident showed employees pasting confidential source code into ChatGPT under its default consumer terms.
Consumer plans for many AI tools have historically allowed providers to use inputs for model training unless users explicitly opted out. Enterprise and team plans are different: OpenAI’s business terms contractually exclude company data from training, and Microsoft states that Copilot for Microsoft 365 does not use your organisation’s content to train foundation models. If your team handles client or personal data, you need that commitment in writing, along with a UK GDPR-compliant data processing agreement.
For regulated firms, this matters further. The FCA has confirmed that AI does not reduce Senior Managers’ responsibilities under SM&CR for outcomes and accountability. The ICO advises that deploying AI on personal data is likely to require a Data Protection Impact Assessment. If you have clients in the EU, the EU AI Act introduces additional obligations for deployers of AI systems on the European market, including businesses based in the UK.
Where will integration make or break the tools you trial?
A tool that works well in a demo but sits outside your team’s existing workflows rarely survives past the pilot. The UK Government’s AI Playbook stresses this risk: tools adopted without connecting to systems your team already uses get abandoned quickly. For a services firm running on Microsoft 365 or Google Workspace, the simplest starting point is AI built into those platforms.
Microsoft Copilot for Microsoft 365 is embedded in Word, Excel, PowerPoint, Outlook, and Teams. Google Gemini for Workspace works across Docs, Sheets, and Gmail. HubSpot’s AI features operate inside the CRM many services firms already use. Xero and Sage have built AI-assisted reconciliation and invoice coding directly into accounting workflows. Tools your team already opens every day are far more likely to be used consistently than a standalone product that requires a separate login.
When assessing vendors, look beyond the tool itself. The CMA’s 2023 review of foundation models flagged concerns about over-dependence on a small number of AI providers and the risk of lock-in. Ask who provides the underlying model and whether that could change. Ask what uptime and support commitments are in the contract. For client-facing or regulated work, check whether an audit trail of prompts and outputs exists.
When does training solve the problem, and when does it mask the wrong tool?
When adoption stalls, the instinct is to run a training session. Sometimes that’s the right call. The UK Government’s AI Playbook and the NCSC both stress that staff need practical guidance on what data they can share with AI tools, rather than deep technical skills. But if the tool generates outputs requiring more checking than the original task, training alone won’t rescue it.
For teams new to AI, a simple internal playbook covers more ground than a technical course. The essentials: which tools are approved, what data can and cannot go into them (no client names or identifiable personal data in consumer tools), and a small library of standard prompts for proposals, meeting summaries, and draft reports. 360Learning’s 2024 AI Upskilling Playbook recommends designating one or two internal AI champions who support colleagues and add to the prompt library as the team learns.
The Department for Education’s 2023 guidance on generative AI found that organisations that trained and guided staff outperformed those that banned tools outright. If, after training, staff still avoid the tool or consistently override its outputs, the issue is more likely the tool than the people using it.
What does a sensible buying process look like for a 10 to 50 person firm?
Many AI buying mistakes happen in the first two weeks, when the criteria are vague and the demos are compelling. A four-stage sequence avoids the worst of them: map use cases and data risks before you look at a single product, then pilot on a narrow task, then build internal guidance before you scale, and measure against specific benchmarks before renewing any contract.
AI productivity tools for office work commonly run at US$20 to $30 per user per month for team or enterprise plans. Harvard Business School research found that consultants using GPT-4 completed tasks 25% faster and with a 12.2 percentage point improvement in quality. On billable work, even a fraction of that gain offsets the licence cost. The implementation time, governance setup, and staff onboarding are the real cost, not the subscription.
On contracts, three terms matter above the rest: who owns the IP in your prompts and outputs, whether the vendor’s terms contractually prevent training on your data, and what liability protection exists if the tool infringes third-party copyright. For any tool handling personal data, ask for a UK GDPR-compliant data processing agreement. The ICO’s guidance for procuring AI stresses due diligence on data handling, auditability, and exit options.
Measure over 90 days: hours saved on defined tasks, error rates in AI-generated outputs, and a short staff survey on perceived usefulness. A 90-day and 180-day review gives you enough data to decide whether to expand the rollout, renegotiate the contract, or cut the tool before it becomes a sunk cost.
