AI in e-commerce: which processes should you automate first?

Start with the work that takes time

Describe one task from beginning to end. What triggers it? Which information does it need? Who checks the outcome, and where is that outcome used?

Include searching, coordination, review and rework. A task involving ten minutes of writing may still take an hour when the necessary information is difficult to find.

Distinguish an information problem from an execution problem. If nobody knows which product specification is correct, generating text more quickly will not resolve the uncertainty. A reliable source and clear owner may be the more useful intervention.

Choose a process where improvement would matter. An infrequent task with few errors does not automatically deserve priority because an attractive AI demonstration exists for it.

Choose rules, AI assistance or an agent workflow

Rules-based automation suits predictable steps. Transferring data, applying a fixed calculation or creating a task when information is missing often requires no AI.

AI assistance can be useful where language and interpretation matter: preparing a draft, comparing texts or grouping feedback. The output still needs suitable review.

An agent workflow uses AI to carry out several steps through agreed tools. It might retrieve information, prepare a draft and place it in a review queue. Additional steps and permissions also increase the need for boundaries and oversight.

A straightforward integration may therefore be sufficient. Our Channable guide examines existing options for recurring product-data work. Introduce AI where interpretation adds something useful.

Four applications worth assessing

1. Preparing and checking product content

Use approved specifications, brand guidance, existing copy and channel requirements as inputs. AI can prepare a draft or flag inconsistencies and missing information.

The potential benefit is less preparation work and a more consistent first review. Generated product information is not a reliable source in its own right. Missing dimensions, materials or claims must not be filled in as established facts.

A product owner should check accuracy and supporting evidence. A content owner should review clarity and brand fit. Publication follows the agreed approval process rather than becoming an automatic consequence of generating a draft.

2. Summarising reports and flagging changes

Start with checked exports and consistent definitions for revenue, costs, orders and reporting periods. AI can prepare a summary and identify changes that deserve investigation.

Calculate important figures using established formulas where possible. A language model does not need to replace a calculator or financial source.

A decline may relate to availability, seasonal demand or a measurement change. A convincing explanation is still a hypothesis. An analyst should verify the figures and examine the explanation before anyone changes campaigns or commercial priorities.

3. Analysing reviews and customer questions

Use relevant reviews and enquiries, retaining only the information needed for the analysis. AI can group themes such as compatibility, ease of use or unclear instructions.

This may help identify product information or experiences worth improving. However, the feedback does not automatically represent all customers. Language, irony and context may also be misinterpreted.

Have someone read examples within the proposed themes and check the classification. Use the findings to investigate improvements, rather than dismiss complaints or manufacture positive conclusions. Keep the original evidence available for reviewers.

4. Preparing and routing internal tasks

Use agreed process rules, task types and responsibilities as inputs. AI can summarise a request, identify missing information and suggest the appropriate owner.

That may make handovers clearer. It will not resolve unclear responsibilities by itself. A request may concern several teams or contain information that should not be visible to everyone.

Send uncertain cases to a named person. Restrict access and ensure that routing a task does not accidentally grant authority to change prices, orders or live campaigns.

A practical example: from advertising export to review

An e-commerce manager gathers advertising results each week and prepares an overview for the team. A focused AI workflow can support that preparation:

1. A checked export provides the figures and established calculations.

2. AI summarises the main changes and flags products or campaigns that deserve attention.

3. Missing information and possible explanations are listed separately.

4. The e-commerce manager checks the summary and decides which actions need further investigation.

The potential benefit is less preparation work and a more consistent review process. Campaigns are not changed automatically. During the pilot, measure the time required for the entire process, including checks and corrections.

This is an illustrative workflow, not a BND client case or a guaranteed outcome.

Decide which application deserves priority

Compare potential applications through five questions:

• Frequency: how often does the work occur?

• Time: how much preparation, review and rework does it require?

• Data quality: are reliable inputs available?

• Error impact: what happens if the output is wrong?

• Reviewability: can someone check the result reasonably quickly?

This is a practical decision aid, not a proven BND scoring model. A recurring task with reliable inputs and easily checked output may be a suitable starting point.

A process with uncertain data and serious consequences needs stronger foundations first. For complex data flows, our ChannelEngine guide examines systems and responsibilities. That is a different question from deciding where AI-assisted interpretation can help.

Make authority and responsibility explicit

Assign a process owner. Specify which information the application can read, where it is processed and who can access the results.

Use only the data required. Remove personal information where it is unnecessary, and check whether the chosen environment fits your privacy, confidentiality and retention arrangements.

Keep enough information to investigate problems: source data, relevant settings, outputs, corrections and approvals. Avoid collecting additional sensitive information merely because logging is possible.

Price changes, changes to live campaigns and external publication need explicit decision rights. A first pilot can exclude these actions entirely.

Agree when the process should pause or escalate to a person, including situations involving missing sources, conflicting information or outcomes outside the approved scope.

Start with one focused pilot

First, measure the existing process: time, quality, errors and delays. Include ordinary cases and relevant exceptions.

Then choose a limited product group or recurring report, one owner and a test period. Two to four weeks may be a workable starting point if the task occurs often enough to provide useful evidence.

Define quality criteria before testing. These might include no invented specifications, figures traceable to their source and correct assignment of tasks.

Agree who reviews the results, how much checking is required and what would lead to adaptation or stopping. Compare the proposed workflow with the current approach before extending it to more products, teams or tasks.

Measure time saved after checking and rework

Consider a fictional example. A task previously took sixty minutes. With AI, preparation takes twenty-five minutes, review takes twenty and corrections take ten. The net saving is five minutes, not thirty-five. Implementation costs have not yet been recovered.

Also assess quality, software costs and reliability. Released team time becomes a direct financial saving only when an actual expense disappears. Otherwise, it creates capacity that can be used for more valuable work.

Document one recurring task this week

Choose a task that occurs regularly. Record its inputs, steps, current time requirement, common errors, desired output and owner.

Then assess whether fixed rules, AI assistance or a bounded agent workflow would fit. Keep the first decision small enough to test and review.

BNDX brings AI and automation expertise into BND’s wider approach.