Automation at the Regional Accounting Firm
A regional accounting firm employs 30 people. Junior staff spend about 40 percent of their week re-keying figures from client PDFs into spreadsheets, then checking totals by hand. A partner proposes buying an automation tool that extracts figures from PDFs and enters them automatically. The tool costs $18,000 per year and is claimed to be '95 percent accurate'. Two senior accountants are sceptical, noting that a single wrong figure in a client return can trigger a correction and reputational damage. Clients range from simple sole traders to complex trusts, and PDF quality varies from clean digital exports to phone photographs of paper receipts.
1. Evaluate the '95 percent accurate' claim as a basis for the decision, and state what evidence the firm should require.
Read the worked answer
A single accuracy figure is not enough to decide, for several reasons.
- The base rate matters: 95 percent accurate per figure means 5 in 100 figures are wrong. A return with 200 figures would contain around 10 errors on average, which is not acceptable unchecked.
- Averages hide variation: the claim probably averages across document types. Accuracy on clean digital exports may be very high while accuracy on phone photographs of receipts may be poor. The firm's mix includes both, so the average may not describe their work.
- Error type matters more than error rate: a wrong digit in a total is different from a misplaced decimal or a figure assigned to the wrong category. Some errors are obvious; others look plausible and pass a casual glance.
- Cost of error: corrections and reputational damage are borne by the firm, not the vendor, so the firm needs a margin of safety.
Require a trial on the firm's own documents: a sample covering each client type and each PDF quality, with results broken down by document type, error type, and detection rate when staff check the output. Compare against the current manual error rate, which the firm should measure first.
2. Recommend how the tool should be introduced, including what human checking remains.
Read the worked answer
Introduce it as an assistant that drafts entries, with a human accountable for every figure that reaches a client.
- Pilot on low-risk work: start with simple sole-trader returns and clean digital PDFs, where accuracy is likely highest and errors are easiest to spot. Do not start with complex trusts.
- Route by document quality: automatically flag phone-photographed receipts for manual entry rather than pushing them through extraction.
- Check by exception plus sample: staff verify every figure above a threshold and a random sample of the rest, and reconcile totals independently. Totals reconciliation catches many errors cheaply because a wrong figure usually makes the total disagree.
- Keep an audit trail: record what the tool proposed and what the human changed, so errors can be traced and the tool's real accuracy measured over time.
- Train staff on the new skill: reviewing extracted data is a different skill from re-keying. Junior staff should learn to look for plausible-but-wrong figures, not just missing ones.
The 40 percent of time saved should be partly reinvested in checking quality, not treated as pure profit. The firm should set a review date and decide whether to expand based on measured results, not the vendor's claim.
3. The partner argues that even with checking, the tool saves time, so the firm should adopt it immediately. Assess this argument.
Read the worked answer
The argument is plausible but incomplete, because it assumes the time saved survives the checking.
Where it holds: if extraction is accurate on clean documents, staff replace slow re-keying with faster review, and the saving is real. Reconciliation of totals is quick and catches many errors, so the net gain could be substantial on the firm's simpler work.
Where it weakens:
- Checking can be as slow as entry when the output is unreliable, because the reviewer must compare against the source carefully. On poor-quality documents the tool may add work.
- The $18,000 is fixed while the benefit varies by client mix. If only half the work suits the tool, the payback case changes.
- Risk is asymmetric: a small time saving does not offset a serious error in a client return. The firm should be willing to lose some efficiency for reliability.
- Switching costs: training, process changes and the pilot period consume time before any benefit appears.
A fair conclusion: adopt, but conditionally and in stages, with measured results at a review point. 'Immediately and everywhere' overstates the case; 'not at all' understates it. The judgement is about where the tool is reliable, not whether automation is good in principle.