Accounting firms are, in theory, perfectly positioned for AI automation. The work is structured. The inputs are predictable. The outputs follow templates. Many processes run on a fixed schedule.
In practice, most accounting practices are running on a patchwork of tools that do not talk to each other, manual steps that someone performs out of habit rather than necessity, and approval chains that were designed for a paper-based world. The result is that experienced staff spend significant time on work that software could handle, while genuinely complex analysis work gets squeezed.
This article covers what we have found in practice, working with accounting and fiduciary firms in Switzerland and Italy: what automation delivers, where it consistently fails, and how to think about sequencing the investment.
The four workflows that consume the most time
Before identifying what to automate, you need to know where the time is actually going. When we conduct Clarity Scans with accounting practices, the same four categories consistently account for the majority of manual overhead.
Document collection and chasing. Clients send documents late, in the wrong format, or incomplete. Staff spend hours per week sending reminders, following up on missing items, and chasing confirmation that documents have been sent. This is entirely mechanical work — the judgement required is zero — yet it is frequently handled by qualified accountants or practice managers.
Data entry and reconciliation. Bank statements, invoices, receipts, expense claims. Despite the proliferation of accounting software, a significant amount of raw data still enters the system via manual keying. The time varies by client profile, but a seven-person practice handling 40 active clients often loses 6 to 8 hours per week to this alone.
Report generation. Monthly management accounts, quarterly summaries, year-end packages. The structure is identical from client to client and month to month. The data changes; the format does not. Yet these reports are assembled manually, formatted by hand, and reviewed for presentation before sending.
Deadline and compliance tracking. Every client has a set of deadlines: VAT returns, payroll runs, annual filings, statutory declarations. Someone maintains a list. Someone checks it. Someone sends reminders. Someone confirms completion. This cycle repeats every month, with the same structure and the same risk of items falling through the cracks.
What automation handles reliably
AI automation is not a single tool or a single approach. When we talk about automating these workflows, we mean connecting existing systems (accounting software, email, document storage, client portal) with automation logic that handles the mechanical steps without human intervention.
Document collection can be automated end to end for most standard document types. A client portal sends automated requests on a schedule. Reminders go out at defined intervals without anyone having to remember. When a document arrives, it is routed to the correct folder and the status list updates automatically. The accountant only sees it when it needs review.
Data extraction from structured documents — bank statements, supplier invoices, standard expense claims — can be handled by AI document processing with high accuracy for common formats. The extracted data is written directly to the accounting system. What remains for human review is the exception: ambiguous items, unfamiliar suppliers, unusual categories.
Report generation based on structured templates is one of the most reliable applications. If the data is in the system and the format is consistent, the report can be generated and formatted automatically. The accountant reviews the output for accuracy and adds any interpretive commentary — the part that actually requires expertise.
Deadline tracking and reminders is pure workflow automation: given a client, a deadline type, and a lead time, the system sends the right reminders to the right people at the right time. No manual scheduling required.
What automation cannot reliably do
The honest answer here matters, because overpromising on automation scope is how projects fail.
AI cannot replace the judgment that comes from knowing a client. When a bank statement shows an unusual transaction, the right treatment depends on context that lives in the accountant's head: the client's business model, a conversation from last month, a known pattern. No automation layer has access to that.
AI cannot handle regulatory complexity reliably. Swiss VAT treatment of cross-border services, cantonal tax nuances, the interaction between IFRS requirements and local filings — these require current regulatory knowledge and professional accountability. Automation that touches tax positions should only be used where the rules are unambiguous and well-defined.
AI cannot manage client relationships. The practice manager who knows which clients need a personal call rather than an automated reminder, who can read when a deadline request is going to cause friction — that is irreplaceable. Automation removes the mechanical work. It does not replace the relationship layer.
The sequencing problem
Most accounting practices that attempt automation start with what seems most visible: they try to automate report generation first, because the reports are painful and the value is obvious.
This rarely works as expected. Report generation automation depends on clean, structured data in the accounting system. If data entry is still manual and inconsistent — which it usually is — the reports come out wrong and the automation creates more work, not less, because someone has to find and fix the errors.
The correct sequence is to work from input to output. Fix document collection first, so documents arrive on time and in standard format. Fix data extraction second, so the accounting system has clean data. Then automate report generation, because at that point the inputs are reliable.
This is the pattern we document in our accounting practice case study: an implementation that took eleven weeks instead of four precisely because the data quality issues upstream needed to be resolved before the downstream automation could be stable. The right diagnosis at the start would have changed the sequence and the timeline.
A realistic picture of what to expect
A seven-person accounting practice that implements document collection automation, data extraction for the top three document types, and automated monthly reporting can typically recover 8 to 12 hours per week of staff time. That is not time the practice eliminates — it is time that gets redeployed to higher-value work: client advisory, new business development, and the complex analysis that was previously getting squeezed.
The implementation takes longer than people expect, because accounting workflows have more upstream dependencies than they appear to. Starting with a proper diagnostic that maps the actual dependencies — not just the obvious pain points — is what separates implementations that compound over time from implementations that stall after the first sprint.