Your First AI Digital Worker Should Improve the Data Before It Automates the Work

The standard advice for adopting AI in operations sounds sensible: clean your data first.

It is also where many SME projects stall.

Customer records are duplicated. Supplier names do not match. Product codes have accumulated exceptions. Approval rules live partly in the ERP, partly in email, and partly in the head of the person who has managed the process for ten years.

If “perfect data” is the starting line, the business may never start.

There is a more practical route. Do not give an AI agent full authority over a messy process. Give it a bounded role that helps the team see the process more clearly. Let it observe work, surface exceptions, recommend corrections, and create an audit trail. Then expand its authority only when the stable path has been tested.

The first digital worker should not hide your data problems by moving faster. It should help you find them.

A useful signal from industrial operations

IFS recently published findings from research it commissioned from The Futurum Group on capacity gaps and digital workers in industrial companies. The release says industrial workers spend 41% of their time on manual, repetitive tasks. It also says only 5.7% of surveyed decision-makers trust AI to act fully autonomously, even though 66% are likely to invest in digital workers within a year. These are vendor-commissioned findings, not neutral industry-wide benchmarks.[1]

IFS also says its Loops platform can automate 60% of agentic transactions end to end, with the remaining 40% including human review and approval checkpoints. In one customer example, a digital worker operating in a purchase-to-order process surfaced a part-number error that had reportedly remained in the process for about a decade. These are IFS’s claims and should be read as product and customer evidence supplied by the vendor, not independent validation.[1]

The useful point is not the percentage.

It is the operating pattern: the agent performs bounded, repeatable work, while people retain responsibility for exceptions and consequential decisions. Better data can emerge from that work because the process becomes more observable.

That is very different from connecting an agent to an ERP and telling it to “fix the records.”

Imperfect data does not justify blind automation

There are two bad extremes.

The first is waiting for a multi-year data-cleaning programme before testing any useful AI workflow.

The second is using AI to automate a broken process and discovering, too late, that it has multiplied the errors.

A governed digital worker creates a middle path. It does not treat bad data as harmless. It treats data quality as something to measure during controlled execution.

Imagine a supplier-order workflow. The digital worker reads incoming confirmations, compares quantities and dates with purchase orders, and flags mismatches. At first, it changes nothing. Every flag is reviewed by a procurement specialist.

After several weeks, the team can answer practical questions:

• Which exceptions are genuine errors?
• Which are harmless formatting differences?
• Which rules exist only in someone’s memory?
• Which suppliers repeatedly create the same mismatch?
• How often does the agent make a correct recommendation?
• What is the cost of a false positive or a missed exception?

Now the data-readiness conversation is grounded in a real workflow. The team is no longer debating whether the database is “clean.” It is measuring whether the data is reliable enough for a specific decision.

Use four stages of authority

The safest route from chat to execution is not a single leap. It is a progression of authority.

1. Observe and surface exceptions

Start read-only.

The digital worker watches one repetitive workflow and records what it finds. It can classify documents, compare records, identify missing fields, or highlight inconsistent values. It must not alter production data, send messages, approve transactions, or trigger payments.

The output is an exception queue with evidence: the source record, the rule applied, the confidence level, and the reason for escalation.

This stage tells you whether the agent understands the work without allowing it to change the work.

2. Recommend an action

Once the exception detection is reliable, allow the agent to propose the next step.

It might suggest correcting a product code, requesting a revised delivery date, or routing an invoice to a different approver. The recommendation should show its evidence and remain easy to accept, modify, or reject.

Rejections matter. They are not merely failures. They reveal missing context, weak rules, and undocumented domain knowledge.

3. Execute after human confirmation

Next, let the agent prepare the action and require an accountable person to confirm it.

This is where many SMEs can recover meaningful capacity without granting full autonomy. The digital worker handles the searching, matching, drafting, and form preparation. The process owner applies judgment before anything consequential happens.

The approval must be explicit and logged. Silence is not approval.

4. Automate only the stable path

Full execution should be narrow, earned, and reversible.

Automate cases where the data pattern is consistent, the rules are clear, the downside is limited, and repeated tests show reliable performance. Keep unusual values, low-confidence cases, financial commitments, customer-impacting actions, and policy exceptions in the human queue.

The goal is not maximum autonomy. The goal is the right autonomy for the risk.

Domain experts become the AI architects

The person who knows procurement, finance, customer service, or inventory is not just a user of the digital worker. That person defines its operating architecture.

A domain expert should specify:

• the outcome the workflow must produce;
• the data sources the agent may use;
• the actions it may prepare or execute;
• the values and cases that require escalation;
• the evidence required for approval;
• the owner of each exception;
• the rollback method if an action is wrong; and
• the measures that determine whether the workflow is improving.

This is where practical governance lives. Not in a generic policy document, but in the decision rights around a real process.

IT still owns security, identity, access, integration, and monitoring. Business leaders still own outcomes and accountability. The domain expert translates the work into rules, thresholds, and exceptions that the digital worker can follow.

Make every action inspectable

A useful digital worker leaves evidence behind.

For every recommendation or action, record the input, rule or instruction used, tools called, output produced, approval received, system changed, and final result. High-risk steps should have a checkpoint before execution. Production changes should have a rollback path.

This is not administrative overhead. It is how trust is built.

Without logs, the team cannot distinguish a model error from bad source data, a broken integration, an outdated rule, or an incorrect human approval. Without that distinction, the agent cannot improve safely.

Auditability also changes the conversation with management. Instead of reporting that “the AI saved time,” the team can show exception volume, recommendation acceptance, processing time, error rates, rework, and the number of cases escalated for judgment.

A practical first workflow for an SME

Choose one process with five characteristics:

1. It happens often.
2. It follows a recognisable pattern.
3. The current pain is measurable.
4. A subject-matter expert can explain the exceptions.
5. A wrong recommendation is recoverable before it becomes a costly action.

Good starting points may include matching purchase-order confirmations, checking CRM records for missing fields, triaging service requests, preparing invoice exception packs, or validating inventory updates.

Document the current baseline. Run the digital worker in read-only mode. Review every exception. Track where its reasoning fails. Update the rules. Only then move to recommendations and confirmed execution.

Do not begin with the most impressive use case. Begin with the workflow that will teach the business how to supervise a digital coworker.

Data readiness is a result, not just a prerequisite

Clean data still matters. Governance still matters. Human judgment still matters.

But an SME does not have to choose between perfect preparation and reckless automation.

A bounded digital worker can make the process visible, expose inconsistencies, capture expert knowledge, and generate the evidence needed to decide what can be automated next. In that model, data readiness improves through governed work rather than through a separate promise that never reaches operations.

Orchestrate the authority. Do not simply switch on the automation.

Sources

[1] https://www.prnewswire.com/apac/news-releases/industrial-workforce-capacity-gap-being-filled-by-agentic-digital-workers-302860424.html — Industrial workforce capacity gap being filled by agentic digital workers