The hospital celebrated renewal of its clinical dictation service with one simple slide: jobs had been retained. The supplier offered speech recognition, faster templates, and transcripts ready before a consultation ended. In the broad picture there were no redundancies and no formal reduction in hours. Yet Marta, an assistant on an internal-medicine ward, began spending the final hours of each afternoon correcting drug names, dates the system had altered, and negations that had disappeared from a sentence. She also had to justify every incident in a new field because the dashboard attributed delays to “input quality.” The contract could count jobs. It did not count who repaired the text, who carried the clinical risk, or how much of a shift no longer belonged to them.
That distinction matters because automation can preserve headcount while profoundly changing the quality of work. A balance sheet that lists only contracts, savings, or volume answers a question that is too narrow: how much labour remains. To understand impact, follow the route of the tasks. What does each group now do that it did not do before? Which decisions have lost room for judgement? Which errors reach a particular person, with what time and what authority to resolve them? In healthcare, these are not social footnotes to technology. They shape the safety of the record and whether care becomes little more than maintaining a stream of documents.
Begin with an account of tasks#
A useful impact account can trace a working day before and after dictation. A doctor prepares a note, the system creates a transcript, an assistant checks terms, coding staff spot inconsistencies, and a nurse consults the report while arranging administration. At every point, describe time, decision, dependency, and possible repair. This makes it visible that a task did not disappear: it moved hands, arrived later, or broke into small corrections that are difficult to see. Average reports per hour may rise while unassigned minutes of interruption multiply.
A useful impact account can trace a working day before and after dictation. A doctor prepares a note, the system creates a transcript, an assistant checks terms, coding staff spot inconsistencies, and a nurse consults the report while arranging administration.
The account must include groups rarely shown in a procurement presentation. Assistants, administrative staff, residents, nurses, archive workers, and temporary staff may receive different parts of the same uncertainty. Marta may correct because she knows the context of the ward, while lacking permission to amend an ambiguous order or protected time to ask a question. That combination is dangerous: practical responsibility without sufficient autonomy. Recording impact by group prevents the convenience of the person dictating or signing the contract from becoming the sole measure of a tool that crosses many people’s work.
Repair time is clinical and labour data#
Correcting a transcript is not a neutral keystroke. It involves reading, comparing with the consultation, recalling abbreviations, finding someone who can clarify a doubt, and deciding whether a change affects an instruction. Repair time should be measured without turning every assistant into an object of surveillance. There is no need to log every keypress or watch a screen. Samples of shifts, categories of incident, and reasonable time estimates agreed with the people doing the work are enough. The question is how much capacity is needed to repair what the tool introduces, not who works fastest under observation.
When repair happens matters as well. If it is concentrated at the end of a shift, when staffing is thin or an urgent response is needed, the same number of minutes carries a greater cost. If it interrupts a patient relationship or delays work that cannot wait, the balance must show it. Automation that speeds the first draft while moving corrections to the edges of the day has not removed work. It has moved pressure to people with the least power to reorganise it. Naming that shift makes it possible to discuss resources, sequence, and limits before exhaustion becomes normal.
Autonomy with support, not risk without defence#
Dictation can give an assistant more ability to flag errors while also confining their judgement inside a rigid procedure. Those are different things. A good tool lets someone pause a note, mark ambiguity, ask for clarification, and record that correction was needed. A poor tool requires the result to be cleaned up so the report closes on time, without displaying the doubt or offering an escalation route. In the latter case, the person is exposed: correcting may be treated as altering a note, while not correcting carries the consequence of letting it pass.
An impact account should therefore ask which decisions each role retains and which have moved to the system, a supervisor, or a supplier. It should state what Marta can resolve herself, what needs clinical validation, and how long that answer takes. It must also allow a person to challenge the balance. If management says correction time is minimal and staff observe that it is rising, there should be a route to present evidence, request a review of the sample, and receive a reasoned response. The right to challenge corrects the asymmetry between an aggregated dashboard and the situated experience of those doing the work.
Learning and the distribution of knowledge#
Another question contracts omit is what the team learns while using the system. People who join the ward can develop clinical skill by comparing a transcript with attentive listening to a consultation. If they receive only already-polished text, they may take longer to recognise a dangerous abbreviation or an error pattern. Conversely, a well-designed tool can free time for teaching, provided that time is not immediately absorbed by volume targets. The balance should record whether shared review spaces exist, whether recurring errors feed training, and whether knowledge is distributed rather than concentrated in a few experts.
Repair can also teach the design process. If the system repeatedly confuses a trade name or fails with certain accents, it is not enough to tell each assistant to be more careful. Incidents should be grouped, their cause checked, and a decision made about changing vocabulary, limiting use in one speciality, or returning a category to manually reviewed dictation. That response turns an error report into improvement infrastructure. Hiding it to protect an adoption rate leaves the hospital paying again and again for the same fault.
A balance that can be disputed and revised#
A responsible balance can be short, but it must be readable. It can show which tasks were automated, which new ones appeared, how much repair time was observed, which groups carried it, which decisions can stop the flow, and which incidents caused change. Aggregated data should protect people: no ranking of assistants or intimate narratives are needed to reveal unequal burden. Relevant transparency lets labour committees, clinical leads, and affected teams debate the tool using shared evidence.
Review needs dates and consequences. At thirty days it may adjust a template or training; at ninety it may redistribute protected time; at six months it may decide that a function does not meet the required level of safety or autonomy. If the only evaluation moment is contract signature, the hospital confuses acquisition with governance. A dictation service is not justified because an invoice exists or because headcount has not fallen. It is justified if it improves the record without handing the work of repairing its limits to people least able to contest them.
Look at Marta’s afternoon#
When retained jobs are shown again, someone can ask to see Marta’s afternoon as well. Not to make her a symbol of sacrifice, but to pose a checkable question: which task reached her, why, with what time, what authority, and what ability to say that the process is failing? That question opens concrete choices: protect review minutes, provide access to clinical escalation, correct a template, limit one category of dictation, or stop a function that produces too much ambiguity.
Measuring by task and group does not reject automation. It prevents it being declared successful before we know who sustains its results. The contract may continue to count employed people, but the hospital must also count repair, autonomy, learning, and the right to challenge. Once those elements enter the balance, dictation ceases to be an abstract promise of efficiency. It becomes a practice accountable for the work it reorganises and for the safety of the words it puts into a clinical record.
