AI process automation is worth considering when a recurring task involves variable inputs, its outputs can be checked and the benefit covers implementation and ongoing review. The assessment needs to identify the exact steps AI will handle, the exceptions people will retain and how the team will recognise a failed result.
This guide compares business process automation candidates and explains how to estimate their return. Start with a specific workflow and your own baseline; published productivity gains can suggest a test, but they cannot replace it.
What task-level studies have measured
Research has measured gains on particular support, consulting, coding and writing tasks. The figures below come from an NBER customer-support working paper, a consultant field experiment, GitHub's controlled coding experiment and Noy and Zhang's writing study. GitHub studied its own product; the tasks, samples and methods differ across these studies.
These results concern assistance on defined tasks, not the removal of an entire role or department. A production automation also needs data preparation, integration, business rules, monitoring and exception handling. Include that work when comparing a pilot with the current process.
Evaluate difficult cases before automating
The experiment with 758 consultants also found work completed roughly 25% faster and quality rated about 40% higher on tasks within the model's capabilities. On a task outside those capabilities, AI-assisted consultants were 19 percentage points less likely to reach the correct answer. That variation matters when deciding which cases a system can handle without review.
Evaluate the cases that are expensive to get wrong, even when they are uncommon in the test data.
Report failures by category as well as overall accuracy. Check unusual inputs, missing information and cases where the correct response is to escalate. A good average score can conceal a small group of costly errors.
Choose between automation, assistance and human decisions
Choose the level of automation for each step. AI can prepare information for a human decision even when it should not make that decision itself. The examples below describe starting points for assessment; the appropriate controls depend on the actual use case.
Automate
Errors are cheap or reversible, outputs are objectively testable, inputs are stable, and exceptions can be detected automatically.
- Invoice extraction and matching
- Ticket classification and routing
- Demand forecasting
- Visual defect inspection
- Internal knowledge retrieval with citations
Augment
AI prepares analysis or drafts for a qualified person who checks the evidence and makes the decision.
- Contract review and clause extraction
- Recruitment administration and document preparation
- Lead prioritisation
- Analysis, research and first drafts
- Clinical and legal interpretation
Keep human-led
Keep a qualified decision-maker responsible where consequences are serious, the situation is unfamiliar or errors cannot be reliably contained.
- Hiring, firing, promotion, pay
- Credit denial and eligibility
- Diagnosis and treatment
- Safety-critical shutdown
- High-value or unusual payment release
What makes a process suitable for AI automation?
A useful candidate has identifiable inputs, an extraction or classification step, objective checks and a controlled action. For example, an invoice workflow can extract fields, validate totals, check for duplicates and send exceptions to accounts payable. Do not rely on a model's stated confidence alone to decide whether the result is safe to use.
Assess frequency, volume, input variability, data access, task complexity, explainability, review effort, regulatory exposure and error costs. Estimate ROI after those costs. A high commercial score cannot compensate for an error the organisation has no acceptable way to prevent, detect or contain.
Business process automation use cases and measures
| Process | Possible scope | What to measure | Main concerns |
|---|---|---|---|
| Customer service | Classify requests and draft answers | Resolution time, corrected answers, escalation success | Invented policy, missed escalation |
| Accounts payable | Extract and match invoices with controls | Handling time, field accuracy, duplicate detection | Fraud, duplicate payments, approval authority |
| Recruiting | Assist with scheduling and administration | Administrative time and correction rate | Personal data, discriminatory outcomes |
| Legal review | Extract clauses for qualified review | Verified clauses and review time | Fabricated references, confidentiality |
| Software and analysis | Assist with implementation and investigation | Accepted changes, defects, total review time | Incorrect code or evidence, missed edge cases |
| Forecasting | Generate forecasts for defined decisions | Forecast error against the current method | Changing conditions, poor source data |
Calculate ROI after review and operating costs
Estimate the return from the workflow's own volumes and costs. Calculate the annual value of recoverable time, avoided errors or additional capacity, then subtract review, model usage, hosting, maintenance and support. Compare the resulting net benefit with implementation and transition costs. If the recurring net benefit is not positive, there is no payback on those assumptions.
Account for the frequency and consequence of errors, including rare but severe ones. Some risks require a control or a change in scope regardless of the expected financial return. Test false negatives, performance across relevant groups and changes in input conditions separately from the average score.
- Measure time as well as case counts: automating the easiest 60% of tickets may remove much less than 60% of the labour
- Separate the contribution of AI from other changes, such as data cleanup, training or a new CRM
- Include the time people spend checking, correcting and escalating results
Design human review that can catch errors
Assigning a reviewer does not establish whether the review works. The person needs access to the underlying evidence, enough time to check it and authority to reject the result. Evaluate review quality using known errors and record which failure types are missed.
Test the review step as part of the workflow
Measure how often reviewers catch incorrect recommendations and how long it takes. Show source evidence clearly and consider asking for an independent assessment before displaying the model's recommendation when the consequences justify it.
Where we start
Start with a frequent task whose output can be checked, such as document extraction, ticket classification or draft preparation. A limited pilot can establish data quality, review effort and integration cost while the team learns how to monitor results and handle exceptions.
Set acceptance thresholds according to the consequences of each error type. Check detection and recovery as well as accuracy. Reversible administrative mistakes and irreversible decisions need different controls.
Our Forward Deployed AI Engineers assess the process, identify a suitable scope and define what the pilot must demonstrate. The assessment may recommend AI assistance, conventional automation or a process change, depending on the evidence.