Deep DiveNo. 03
How we reduced the time required for an administrative process by 75% using AI and automation
Organising information before it is reviewed reduces repetitive work and focuses human judgement where it matters most.
Gonzalo GilgesHead of Practical AI
- Reading time
- 5 min
Preparing a customs case used to take close to one hour. Today, the same process takes around 15 minutes: a reduction of approximately 75%. The system processes roughly 1,075 cases per month.
When every sale requires documents to be processed, administrative capacity can constrain growth. Reducing that burden makes it possible to handle more operations without increasing resources at the same rate. If demand exists and the rest of the process can keep pace, that can translate into more sales and stronger profitability.
01
From searching through documents to reviewing a proposal
A DUA, or Single Administrative Document, is used to declare a customs operation. Preparing it requires data from invoices, transport documents and certificates. The time compared here begins when those documents arrive and ends when the draft declaration is ready.
Previously, a person opened files, searched for the data and transferred it into the system. Their business knowledge was mixed with repetitive searching and data entry.
02
Explain what the team already knows
The technology needs to understand the business rules: which data to collect, which items to combine and when a difference needs checking. For example, when several invoices involve the same exporter and importer, a rule may state that they should be grouped and that their quantities and amounts should be added together.
As we discussed last week, this applies to both a workflow and an AI agent. The procedures and criteria used by the team need to be documented.
Just as we would explain them to a new colleague, we need to explain them to the AI.
A fund considering the acquisition of a company faces a similar challenge. AI can extract revenue, debt and margins from different documents, but the team must define how those figures are normalised, which criteria are used for valuation and which differences require review. Finding data is not the same as deciding what to do with it.
03
How we designed the solution
We developed a hybrid workflow: the process follows a series of defined steps and, at some of them, calls an LLM to complete a specific task. This combines the control and efficiency of traditional automation with AI’s ability to interpret documents that vary in format.
- Register the input. Automation collects email attachments, prevents duplicates and opens the case.
- Classify and extract. AI distinguishes between documents, separates those that are useful and locates the data, even when each supplier presents it differently.
- Reanalyse uncertain data. If a value does not reach the required confidence threshold, a more capable AI model can review it. The more expensive resource is used only where it is needed.
- Apply the rules. AI returns the data to the workflow. The workflow checks it against the business rules, groups, calculates and organises the pre-DUA: the draft declaration. A person reviews the content and its sources before the final filing.
An insurer could use similar logic to prepare policy proposals: receive documents, identify information and apply its own criteria. The professional would begin with an organised proposal to review, while retaining the decision on risk.
Fig. 1A defined workflow, with AI where needed
- 01 Input
Receives documents and opens the case.
- 02 AI
Classifies documents and extracts the data.
- 03 Rules
Calculates and organises the draft.
- 04 Person
Reviews uncertainties, corrects and confirms.
- 05 Pre-DUA
Records the confirmed content.
04
Human judgement improves the next case
If a correction reveals that the team had not defined which document takes precedence, that criterion is added to the system and tested. This is the key to the process: each adjustment helps create a more efficient and robust system.
In this case, the average fell from 10 corrections per case to 2: an 80% reduction.
As the rules cover more cases, people can focus on exceptions. The objective is to process more operations using the same structure while retaining oversight.
05
What 45 minutes means at scale
At roughly 1,075 cases per month, with 45 fewer minutes of work per case, the improvement represents about 806 hours per month: the work of five full-time people.
In figures
- Preparing a customs case, from when the documents arrive to a draft declaration ready: close to one hour before, around 15 minutes today.
- 1 hour to 15 minutes −75%
- Cases the system processes each month, roughly.
- 1,075 cases
- Corrections per case, on average.
- 10 to 2 −80%
- Work saved each month at 45 fewer minutes per case: the work of five full-time people.
- 806 hours
Based on the type of operations the solution could cover, we identified a potential scope of 75,000 cases per year: around 6,250 per month. That could become the next expansion target, provided the rules required for those cases are incorporated and validated.
If the same 45-minute reduction per case is maintained at that volume, it would represent 4,688 hours per month: the working time of approximately 29 full-time people.
The opportunity is to spend less effort preparing information and more on resolving exceptions, handling new cases and improving the rules. Some of the work that creates the efficiency also helps sustain it as volume grows.
To recognise a similar opportunity, look for a process in which the team spends considerable time organising documents before applying judgement. Separate the information it needs, the rules it uses and the decisions a person must retain. That is where the work begins to make each new operation require less effort than the last.
Thank you for reading. Until next time.