Building AI skills across a small office team

This Leeds-based accountancy and advisory firm has 28 staff and a set of AI assistant licences that are going largely unused. Each day, team members encounter tasks such as drafting client correspondence, summarising documents, conducting research, and writing up meeting notes. At the moment, each person decides independently whether to use the AI tool or work manually, prompts it without any shared guidance, and judges the output without a consistent quality check. There is no written policy on what client data can safely be included in a prompt, no agreed prompting standards, and no one in the firm with the technical background to build or maintain any automated workflow.

The managing partners want all 28 staff to use AI tools confidently and safely in their daily work. The end goal is a team that handles AI-assisted tasks with shared methods and clear data rules, and that has identified two or three processes well enough documented to automate properly. Currently, roughly 3,000 of these tasks occur each year, each taking around 30 minutes on average, which adds up to meaningful staff time that better AI use could reduce. The challenge is not a missing tool but a missing foundation: the knowledge, governance, and process documentation that would make any automation trustworthy and sustainable simply do not exist yet.

The recommended path is Education via the AIRIES programme, and the reasoning starts with an honest diagnosis of what is actually broken here. Every problem the firm describes, including staff distrust of AI outputs, unused licences, inconsistent quality, accidental data exposure risk, and no one able to maintain anything built, points to the same root cause: a skills and governance gap. No technical solution addresses that gap. Deploying an AI workflow, an RPA bot, or any other automation into this environment would produce a system the team cannot check, cannot adapt, and cannot confidently use in client-facing work. It would also deepen the compliance exposure under UK GDPR and client confidentiality duties rather than resolve it.

The firm has no stated technology preference and no internal developer resource, which removes several options from practical consideration. Traditional Code requires a developer to build against documented processes, and neither condition exists. RPA requires stable, rule-based, well-documented processes and ongoing bot maintenance, none of which is available here. An autonomous AI Agent requires mature governance, clear data boundaries, and staff who can supervise agent behaviour; the firm currently has none of these, and deploying one would amplify risk significantly. AI Workflow automation is the most plausible near-term technical path, but it is only viable once processes are documented, data policies are written, and the team has the skills to specify and oversee what gets built. The Hybrid path, which combines automation with human-in-the-loop checkpoints, is the logical step after education for one or two specific processes, but it cannot be designed or governed today. Staying Manual wastes the licences already paid for and leaves 3,000 tasks per year consuming avoidable time without addressing the compliance gap either.

Education wins on every dimension that matters for this firm right now. AIRIES Level 1 for all 28 staff builds a shared baseline: how the AI tool works, how to prompt it effectively, and what the technology cannot do. This directly addresses the distrust and inconsistency that currently prevent adoption. Level 3 for team leads and process owners gives those individuals the practical skills to audit their own tasks, document processes clearly, and build simple AI-assisted workflows, which is precisely what the firm needs before it can commission any serious automation. Level 5 for the managing partners develops the judgement to evaluate automation proposals, set meaningful KPIs, and avoid commissioning projects the firm cannot sustain or oversee. These three levels, taken in sequence by the right audiences, produce the foundation that every other path depends on.

There is also a sequencing point worth making explicitly. The data and AI use policy must be written before or alongside Level 1, not after it. Encouraging more AI use before the boundaries around client data are clear would increase compliance exposure rather than reduce it. That policy is a low-effort task in terms of drafting time, but it requires a partner decision and ideally a brief check against the firm's professional indemnity insurance terms, since some policies have specific requirements around AI use in client-facing work. With the policy in place and Level 1 complete, the firm will have removed the fear that is currently blocking engagement and created the governance layer every subsequent step needs.

The projected return is real but it follows the education, not precedes it. At roughly 3,000 tasks per year averaging 30 minutes each, at a blended rate of around 45 pounds per hour, the current annual time cost is approximately 112,500 pounds. A conservative 60 percent saving on tasks amenable to AI assistance, once staff are trained and policies are in place, projects annual savings of around 67,500 pounds against programme costs in the 8,000 to 18,000 pound range. That is an attractive return, but only if the team completes the programme and actually changes how they work. The risk is not the technology; it is the managing partners treating the programme as optional during a busy advisory calendar. A nominated internal champion and a clear completion deadline are not optional extras; they are the conditions on which the return depends.

Education (AIRIES) scores a clear ten against every other path here, and the reason is straightforward: the other paths all assume a foundation that does not yet exist. Hybrid automation, which would otherwise be a reasonable next step for a firm of this size, scores only four because a hybrid approach depends on someone being able to specify the human-in-the-loop gates, write the data-handling rules, and oversee what gets built. None of those conditions are in place. Deploying a hybrid workflow now would mean commissioning external build costs with no internal person able to accept-test or maintain the result, and no written policy governing what client data flows through it. The compliance exposure would grow, not shrink.

AI Workflow sits one rung below Hybrid at three, and faces the same structural problem in a more acute form. Correspondence drafting and month-end reporting are plausible candidates for workflow automation eventually, but neither process is documented, neither has clear decision rules written down, and the team has not yet agreed what may or may not be included in a prompt. Building an AI workflow against undocumented processes is how firms end up with something that works for two weeks and then silently produces wrong outputs because a team member changed a step nobody recorded. RPA scores two and Traditional Code scores one for the same compounding reasons, with the added problem that both require either a stable, rule-based process or an internal developer, and this firm has neither.

Education (AIRIES) wins not because it is the cautious or conservative choice but because it is the only path that actually resolves the root cause. The firm already has the licences. The bottleneck is knowledge, confidence, and a shared set of rules. Once those are in place, the scores for Hybrid and AI Workflow rise significantly, and the firm can commission automation it can genuinely own and maintain. Every other path is a shortcut that skips the step the whole firm is blocked on.

The first action, running in parallel with or slightly ahead of any training, is drafting a one-page AI and data use policy. The managing partners, with input from a nominated compliance lead, need to set down in plain terms what client data may and may not be included in prompts sent to the existing AI assistant, how outputs should be handled, and who is accountable for reviewing AI-assisted work before it reaches a client. This does not need to be a lengthy document, but it needs to exist before the team is encouraged to use AI more actively. The professional indemnity insurer is worth a brief conversation at this stage, as some policies have specific language around AI use in client-facing advisory work. This policy is the governance foundation that every subsequent step depends on.

With the policy in place, all 28 staff complete AIRIES Level 1. This is not a half-day awareness session; it is a structured programme covering how large language models actually work, how to write prompts that produce consistent and useful outputs, and where the technology reliably fails. The six staff who are already using the AI assistant should be positioned as peer champions during this phase, running short informal sessions alongside the formal programme to answer practical questions about the tools already licensed. By the end of Level 1, the team has a shared vocabulary, a shared understanding of limits, and the confidence to engage with the tools rather than avoid them. The distrust and inconsistency that currently make automation impossible begin to dissolve here.

A cohort of six to eight people, including team leads and the partners most actively interested in future automation, then moves to AIRIES Level 3. This level teaches participants how to audit their own tasks, map process steps clearly, and build simple AI-assisted workflows using the existing AI assistant and, where appropriate, a lightweight workflow layer such as Microsoft Power Automate or Make, depending on the licence environment. The specific value here is that this cohort becomes the firm's internal capability: they can document processes in enough detail to specify an automation, and they can oversee and test what an external partner builds. They use Level 3 skills directly to identify the two or three processes genuinely ready for automation, most likely correspondence drafting and month-end reporting, and to produce the process maps that will serve as the brief for any future build.

The managing partners complete AIRIES Level 5 alongside or immediately after the Level 3 cohort. Level 5 focuses on setting meaningful KPIs for AI adoption, evaluating automation proposals critically, and making informed investment decisions about which processes to build against. This is the step that prevents the firm from commissioning a project it cannot look after. Once all three levels are complete and two or three processes are fully documented with clear inputs, outputs, decision rules, and data-handling requirements written down, the firm is positioned to run a tightly scoped pilot automation for the single best-fit process. That pilot, commissioned externally but owned and tested internally by the trained cohort, becomes the template for scaling to the remaining candidates. The data flow at that point runs from a documented trigger, through the AI assistant or workflow layer, to a human review checkpoint, and out to a client-ready output, with the trained team able to identify and correct any deviation at every stage.