Where AI and automation create value in a business process

Begin with the journey of a case

The useful question is not where to put a chatbot, but why a case waits, returns or requires another round of data entry. A process crosses people, rules, documents and tools. Improving one visible step matters little if the case remains blocked downstream. Observing the real journey, including exceptions, gives a better starting point than a list of AI features.

This note offers a measurement method. Its examples are hypothetical situations, not Digital Place customer results. A benefit becomes credible when its scope, quality and full cost are documented.

Separate active work from waiting

Measure cycle time from receipt to closure, then the time actually spent working on the case. Cases may wait for a document, approval or another department’s reply. Faster entry will not eliminate an approval rule that blocks work for days. Automation may, however, flag missing documents earlier, route a request or make a queue visible.

Map a few journeys with the team: ordinary cases, incomplete applications, urgent requests and complex cases. At each handoff, record what is awaited, who provides it and which tool holds it. A timestamped “application complete” event is often more useful than a broad “in progress” status.

Target rework and repeated entry

Copying the same information between email, spreadsheets and applications creates opportunities for error. Check first whether an integration or rule can solve the problem. For varied documents, extraction may prepare fields while a person confirms important values before saving. Visible provenance matters as much as reading speed.

Track rework: the share of cases returning to a previous step. Identify causes such as missing fields, unreadable documents, incorrect routing or misunderstood rules. Verify that a lower rework rate does not simply move errors to closure or to another team.

Search and decision preparation

An assistant can retrieve a procedure and prepare a sourced summary. It helps when the difficulty is gathering scattered information. Its usefulness declines when procedures are obsolete or inaccessible to the people who need them. Assign source owners and make insufficiently supported answers identifiable.

Preparing a decision means assembling facts, identifying missing information and proposing options. It does not automatically authorise the system to decide. Rejections, financial commitments and permission changes need responsibilities and rules established by the organisation.

Coordination and exceptions

Automation can route requests, prepare acknowledgements and remind people of pending actions. It must also stop when it encounters an unexpected document, conflicting rules, inconsistent sources or an absent response. Exceptions should reach a human queue with enough context, rather than disappear behind a technical status.

Human-in-the-loop requires an actual operating arrangement: authorised reviewers, review deadlines, available information and a means to correct results. If the review queue grows faster than its handling capacity, the process has not improved. Measure escalation rates and the time people spend resolving those cases.

Define a small measurement set

Cycle time describes the end-to-end experience; examine both the median and slow cases. Cost per case should include human work, tools, model usage, integration and operations. Error rates need an explicit definition and a checked sample. Rework rates describe friction; SLA attainment measures whether promised deadlines are met.

Add capacity, meaning cases correctly completed over a period, and a satisfaction measure relevant to users. Track human escalations and abandoned cases too. A case removed from a queue through an incorrect rejection must not count as a success.

A calculation with explicit assumptions

Suppose, purely to illustrate the method, that a case requires twelve minutes of active work before a change. Afterwards preparation takes seven minutes and human checking takes two. Active work becomes nine minutes, three minutes less per case if quality remains equivalent. Across one hundred comparable cases, this assumption represents three hundred minutes of work, not three hundred minutes saved in cycle time.

To estimate the economic effect, use an explicit fully loaded hourly cost and subtract additional service and control costs. This calculation promises no savings; it organises a hypothesis for testing. Observed times and the distribution of exceptions may invalidate it.

Compare before and after fairly

Establish a baseline period and retain the same definitions after the change. Compare cases of similar difficulty; record volume, seasonality, staffing and rule changes. A comparable pilot group or gradual rollout can help distinguish the tool’s effect from broader changes.

Measure learning time and workarounds too. A team may appear faster because it checks less, or because another department absorbs corrections. Follow selected cases end to end and ask users what they still have to redo.

Research and its limits

Generative AI at Work studies assistance in customer support and reports effects that vary across workers. It does not provide a universal multiplier for every organisation. Conversely, METR’s study of experienced developers shows that AI assistance can slow work in a particular setting.

The February 2026 METR update also highlights participant and task selection problems as tool use changes. Its newer observations do not support a reliable, directly comparable estimate of uplift. The early-2025 result should therefore not become a general conclusion about tools available in 2026.

These findings support field measurement. Neither a model benchmark nor a perceived gain in speed replaces observing comparable cases. An impressive capability may be the wrong intervention for the chosen process.

Define the pilot decision in advance

Choose a bounded scope, an owner and stopping criteria. Define what the tool prepares, what a human confirms and how to return to the previous workflow. Set acceptable quality and escalation thresholds before reviewing initial results. Protect confidentiality and access rights from the first trial.

Digital Place’s approach connects processes, data, integration and control. The intended value is a case handled better, with less friction and an explainable decision. Measurement should reveal both improvements and remaining work.