Two ways of handling
the same work
Neither approach is right for every situation. This page sets out what each involves, where each is suited, and what actually changes when automation handles part of the work.
Back to homeWhy the comparison matters
Automation is often described in ways that make the choice seem obvious. It is not. Both manual and automated handling involve real trade-offs, and the right arrangement depends on the volume, variability, and stakes of the specific task involved.
What follows is a straightforward account of how the two approaches compare across the tasks Tokeiya works with: document processing, demand forecasting, and staff readiness. The differences are real, but so are the conditions under which manual work remains the more appropriate choice.
Side by side
| Area | Manual / Traditional | AI-Assisted (Tokeiya) |
|---|---|---|
| Invoice processing speed | Minutes to hours per document depending on format and staff availability | Seconds per document for standard formats; exceptions routed for review |
| Error rates | Varies by staff, volume, and time of day; increases with fatigue and volume peaks | Consistent within trained data types; error rates reported by document category |
| Forecast method | Historical averages, buyer experience, supplier minimums | Sales history with seasonality and promotional period adjustment; compared against existing method |
| Staff confidence with tools | Learned informally; inconsistent across department | Structured programme using the department's own material; policy document provided |
| Scalability | Requires proportional staffing increases for volume growth | Handles volume increases without linear cost increase; exceptions scale separately |
| Dependency after setup | Dependent on staff retention and training continuity | Documented for internal maintenance; staff retrained on how to update the model |
Checkpoint: exceptions, edge cases, and low-confidence outputs always pass through a staff review step — regardless of which approach is in use.
What the approach is built on
No output is treated as final without a defined review step. The system identifies what it handles with high confidence and what it does not. Items in the second category go to staff — not silently past them.
Every engagement produces accuracy figures by document or product type, including failures. Where the automated approach performs worse than the existing method, that is reported. The parallel running stage exists specifically to make this comparison possible.
Effectiveness by task type
Where automation tends to do well, and where manual work remains more appropriate.
High volume, structured formats, consistent suppliers: automation handles well. Highly variable formats, occasional documents, unusual layouts: manual review remains appropriate and is retained as the exception path.
Where 3+ years of consistent sales records exist and promotions are tracked, AI-assisted forecasting improves on average-based methods. For new product lines or thin historical data, the existing approach typically performs similarly or better.
Generic AI training tends to produce surface familiarity without practical confidence. Role-specific exercises using the department's actual work produce more durable understanding — and a policy document staff can reference later.
Cost and long-term view
A three-year perspective on setup versus ongoing cost.
| Service | Setup (once) | Ongoing (internal) | Manual equivalent |
|---|---|---|---|
| Invoice Processing | ¥44,000 | Staff review of exceptions only; system maintained internally | Staff time per document × monthly volume, recurring |
| Demand Forecasting | ¥43,000 | Periodic retraining when sales patterns shift; documented process | Buyer time, overstock and understock costs from forecast error |
| Staff Programme | ¥28,000 | Policy document available; question channel open two months | Informal learning, inconsistent practice, policy gaps |
All figures in Japanese Yen. Setup cost is a single payment per engagement. There are no recurring licence or platform fees in any of the three services.
What the experience looks like
- Process depends on individual staff knowledge
- Training is informal and inconsistent across the team
- Errors identified after the fact, during review or reconciliation
- Volume increases require additional staff time directly
- No structured measurement of process accuracy
- Scoped against your actual documents and data before work begins
- Parallel running period allows direct comparison with current practice
- Exceptions routed to staff; nothing passes through silently
- Accuracy reported by document or product type, including failures
- Documentation and handover: internal team can maintain independently
Sustainability over time
Manual processes are sustainable when volume is low and variability is high. They depend on individual knowledge remaining within the organisation, which is a genuine operational risk when staff change roles or leave. Documentation of manual steps is often informal.
AI-assisted processes are sustainable when the task is high-volume and the data feeding the system remains consistent in format and availability. The engagement ends with documentation your team can use to retrain or adjust the system — reducing dependency on external support over time.
Common misconceptions
"Automation replaces the need for staff review."
Automation handles the portion of work where confidence is high. Exceptions — documents with unusual formats, data points outside normal range, low-confidence outputs — always pass to staff. The review step is retained by design, not as a fallback.
"AI forecasting is always more accurate than experience-based methods."
Not always. For new product lines, thin data, or categories with highly irregular demand, a buyer's judgment informed by supplier relationships often outperforms a model. The parallel running stage is there to measure this directly, not assume the outcome.
"Staff training on AI tools is a one-time event."
Tools evolve and staff change roles. The written policy draft and the two-month question channel exist precisely because questions arise after the sessions end. The programme provides a foundation that departments can build on, not a fixed endpoint.
When this approach is a reasonable fit
- Recurring, bounded tasks with consistent inputs
- Volume high enough to create processing friction
- Existing data that can be used to train and evaluate
- Internal team willing to maintain the system
- Very low volume tasks where manual handling is fast enough
- Processes with no consistent data or records
- Highly judgement-dependent tasks with no clear rules
- Situations where staff have no capacity to maintain or review
Have a specific task in mind?
Describing the task — its volume, its current method, and what friction it causes — is enough to begin. We can assess fit together without any commitment on your side.
Get in touch