All guides

Which ERP Task Should You Automate First?

Not every manual ERP task is worth automating. Learn how to evaluate the workload, exceptions, and costs before investing in your first AI automation.

A rounded orange shape sits inside a white fold, surrounded by smooth dark slate surfaces. The ternfold wordmark appears in the bottom right.

Consider a distributor whose repeat customers email their orders as PDFs. Its team works in an ERP that already holds customer accounts, products and prices. For each order, an employee reads the PDF, finds the matching records and enters the line items. Sales staff also prepare quotations: they select products, clarify variants and put together the quotation details in the same system.

The owner wants to use AI to take on some of this recurring work. Orders look promising because they are repetitive; quotations because each request takes more time. To choose between them, the owner needs to understand how much work the team would still have to do in each case, and how it could use the time released.

Why investigate order entry first?

For the comparison, assume the following volumes and processing times. All figures in this example are assumptions, not measured customer results.

Scroll sideways to read the full table.

Task Monthly volume Manual processing time Monthly workload
Enter PDF orders from known customers 300 orders 10 minutes per order 50 hours
Prepare quotations 60 quotations About 20 minutes per quotation 20 hours

Order entry takes more time overall. Many of the customer and product mappings are also already stored in the ERP. That makes it a reasonable task to investigate first. The owner has a starting point for a trial, but has not yet made an investment decision.

A closer look at the daily work could change the picture. A familiar product may be easy to match, while an unclear quantity calls for clarification from the customer. Quotations may involve lengthy discussions about variants and pricing, yet a narrowly defined product category could need little clarification. The owner therefore asks to see a typical case and a difficult one for each task, to understand which work an assistant could take on and which decisions would remain with the team.

What would change when an order arrives?

An assistant that can read PDFs could first present an order's contents as text or a table in a chat. The details would be easy to read, but an employee would still need to find the right customers and products in the ERP and enter the line items. What matters for this comparison is whether the assistant can prepare a usable draft in the ERP.

To do that, it would need to look up customers and products, use existing mappings and pass the details to the ERP's order-entry functions. The system would continue to apply its rules for pricing, calculations and order checks. An employee would start with a prepared draft, check its contents and resolve any open questions. The team would still approve the order and send the confirmation to the customer.

The assumed ten minutes for manual order entry cover the active work from receipt of the PDF to a reviewed ERP draft. We use exactly the same boundaries for the assisted process. Approval and customer confirmation are excluded from both estimates, so the comparison covers the same completed task.

Ternfold makes the core functionality and business logic of existing software and systems accessible to AI agents and assistants. For this distributor, that means access to customer and product data and the preparation of orders in the ERP the team already uses.

Before commissioning a trial, the owner checks whether an existing ERP feature or solution already covers the task. A connection needs access to the required functions with the appropriate permissions. The guide to technical feasibility explores this question; the owner can also discuss it with Ternfold using the specific task and ERP software as a starting point. The owner can move on to a trial if the necessary access looks feasible and the expected effort is reasonable.

How much of the 50 hours could be released?

Now assume that all 300 orders fall within the proposed scope and that AI-assisted preparation is attempted for every one. It works for 240 orders, taking an average of 4.5 minutes of employee time each, including review and corrections. For the other 60, an employee has to finish the work manually after a failed attempt. The assumed twelve minutes for each of these orders include both the failed attempt and the subsequent manual work.

Scroll sideways to read the full table.

Outcome Calculation Monthly workload
Processed with AI assistance through to a reviewed draft 240 orders (80%) × 4.5 minutes 18 hours
Failed attempt followed by manual completion 60 orders (20%) × 12 minutes 12 hours
Processing all orders 18 + 12 hours 30 hours

Those 30 hours are spread across all 300 orders, giving a weighted average of six minutes per order that includes failed attempts. Allow another three hours a month for monitoring and maintenance, and 17 hours of capacity would be released from the current 50-hour workload: 50 − 30 − 3 = 17.

The time needed to review drafts can change this calculation considerably. If the overall average across the same 300 orders is eight minutes rather than six, processing takes 300 × 8 = 2,400 minutes, or 40 hours. After the same three hours for monitoring and maintenance, only seven hours would be freed: 50 − 40 − 3 = 7.

What would that time mean for the business?

The owner now needs to consider what the team could do with 17 extra hours each month. It could clear a backlog or handle more orders. That would be useful capacity, without automatically reducing payroll expenses. If the change actually eliminated paid overtime, that avoided expense could be measured separately.

Faster responses to customers also need their own assessment. Preparing a draft may take fewer working minutes, yet the draft could still wait for approval. If the distributor wants to shorten the time until an order is confirmed, it needs to measure how long customers wait as well as the team's active working time.

There is also work and cost that the monthly time calculation does not fully capture. The example assumes a one-time setup effort of 40 internal working hours. The costs of integration, the AI assistant, licences, usage and ongoing support are still unknown. Whether the released time justifies that expense and effort can only be judged once the operational benefit and the full costs are known.

How the trial would resolve the choice

Before the trial, the owner could require order entry to release at least 15 hours a month after monitoring and maintenance. There would also be an agreed cost ceiling and a requirement to match customers, products and quantities reliably. A draft with incorrect matches could not count as correctly processed. The owner would then have clear criteria for judging the trial.

The team would record every order within the agreed scope over a period that represents ordinary operations. It would measure active working time through to a reviewed ERP draft, including questions, corrections, failed attempts and full manual processing. A comparison with the existing process needs the same task boundaries and a comparable mix of orders. Additional difficult cases or orders not previously used in testing can reveal weaknesses; monthly estimates must reflect how often each type of case actually occurs. Measured times and the monthly projections based on them remain clearly distinguished.

If such a trial confirmed the 17 hours, the assumed capacity target would be met. With reliable drafts and total costs within the agreed ceiling, the distributor could put assisted order entry into regular use. If only seven hours were released, the owner would need to consider whether correction work could be reduced enough, or whether a narrower group of orders would be more suitable. Narrowing the scope would require a new monthly calculation that includes the smaller volume and the work still handled manually.

If order entry remains too demanding, quotation preparation for a single product category comes back into consideration. The owner would need reliable figures for the number of eligible quotations, the remaining work, and the setup and operating costs too. That comparison might show that the smaller task releases more useful time for the team at an acceptable cost. If neither task offers that benefit, the owner can defer automation.

Keep your software.
Reduce the manual work.

Tell us which software you use and show us one recurring task. We can work out together which functions need checking next.

Discuss a task