Human judgment can be worth keeping in automated workflows

The work left after automation may be part of its return

Removing the last human tasks from an automated workflow can make the investment perform worse. At Wonderful Dental, an ecommerce business selling dental products, people still finish advertisements and review customer-service replies because that work contributes to the result. The company’s experience challenges a common budgeting assumption: the hours left after automation are necessarily the next expense worth eliminating.

An unnamed Wonderful Dental insider with a creative background describes the experience, drawing on their work in advertising approval, automation decisions and support escalation. The insider has a direct interest in the approach succeeding: the company benefits when its advertisements perform better and its support costs fall. The insider calls the approach semi-automation, meaning deliberately keeping human work whose removal would cost too much or weaken a capability the business needs.

Keeping that work can look like leaving a project unfinished, especially when a development plan still lists possible improvements. Yet a budget can count creative or review time as an expense while overlooking what that time produces. Wonderful Dental’s advertising comparison shows why the work matters: people approved both approaches, but only one involved substantial human creative finishing.

Human creative finishing changed the advertising result

Wonderful Dental compared 20 advertisements generated across images, hooks (opening ideas meant to catch attention), headlines and copy (advertising text) with two advertisements that people finished using AI-generated graphics. The fully generated batch looked suitable for paid distribution: its composition looked competent and its branding seemed appropriate. “I approved those ads,” the insider says. So approval was already part of the process that performed worse.

The generated batch also offered more chances to find a successful advertisement. With 20 variations against two, it supplied 10 times as many versions, including more hooks and approaches that might attract attention. Wonderful Dental reports using the same channels, audiences and performance metric as its other advertising. Even with that numerical advantage, the fully generated advertisements were the weakest performers among everything the company ran during six months.

In that comparison, the insider describes the company’s advertisements as winning by “40%” and describes the generated batch as “40% under on click-through.” Two human-written headlines outperformed all 20 AI alternatives on click-through, the share of viewers who clicked an advertisement. The production process shows what went into the two human-finished advertisements.

Their production began with substantial automation. Wonderful Dental trained a generative image model on its brand, products and existing images, then used the insider’s guidance to produce multiple graphics. The model supplied a base image that people developed into an advertisement with a specific message and presentation.

Developing that image required decisions that approval alone could not supply. People wrote the copy, retouched the image in Photoshop, selected the fonts, arranged the text and chose calls to action, or CTAs, which prompt viewers to take the next step. They then purchased media through the company’s established process. Creative direction, brand knowledge, imagination and copywriting remained part of production; approving a finished advertisement meant assessing choices already made.

Retaining that work still left Wonderful Dental with a large production saving. A graphic of this kind had previously cost hundreds of dollars and required weeks; the new process generated the starting material in minutes, followed by a few hours of cleanup. The insider identifies reduced production time as the main benefit. Human finishing preserved that gain while contributing to the better-performing advertisements.

The insider attributes the generated batch’s weaker performance to viewers becoming familiar with recognizable AI patterns. “My guess at why the 20 lost is that people can smell it now,” the insider says, pointing to recurring characteristics in composition, lighting, typography and copy. In the insider’s explanation, viewers were less accustomed to these characteristics a few years ago, but repeated exposure has made them easier to recognize.

That familiarity creates a problem for companies using similar tools to produce polished advertising. As more companies generate comparable treatments, viewers have fewer reasons to notice any individual advertisement. The insider saw a generic AI finish in Wonderful Dental’s generated advertisements despite their use of the company’s branding. Recognizing the brand and finding the advertisement distinctive therefore require separate judgments.

The cost of making those judgments becomes more visible as initial production gets cheaper. “AI gets you to the middle for almost nothing,” the insider says. An acceptable first version becomes inexpensive, while making it useful for a particular business still takes time. Removing those remaining hours may improve the production budget while weakening the advertising it buys.

Cheaper first versions also create more work to assess. Across design, writing, advertising, photography, video and software, the insider reports spending more time deciding which work deserves further development as output becomes easier to produce and more polished. Additional versions sometimes help, but the insider gives the example of a strong first version followed by 19 worse alternatives. Producing more work creates a selection task of its own.

The insider calls the business-specific judgment needed for that task “taste.” It includes recognizing work that looks professional but does not suit the company, copy that resembles everyone else’s, or an attractive image that makes the product appear cheap. These judgments connect creative choices to business consequences. As the insider puts it, “My 20 ads were polished.”

For Wonderful Dental, that means continuing to use AI for base assets while retaining human copywriting, retouching and layout. Screening for generic patterns before buying media adds another check, while advertising usefulness and performance determine whether the process works. The insider uses the tools daily and expects to increase their use next year. As its use of AI grows, the company is preserving the contribution of creative finishing.

Automating rare support exceptions can undo the savings already achieved

Customer service gave Wonderful Dental a different reason to retain human work: further automation could cost more than the remaining task justified. The company repeatedly received questions about order location, product suitability and differences between products. The insider gives examples of customers asking, “Where’s my order?”, “Can I use this product for this?” and “What’s the difference between these two?” Even a straightforward reply required an employee to gather information before writing.

Gathering that information meant finding the customer, checking the order, retrieving tracking details, searching the knowledge base and composing a response. More difficult messages combined an order issue with a product question or required detailed product knowledge. Because many replies needed this preparation, automating it offered an immediate benefit.

Wonderful Dental built an internal support tool, Robonos, to connect company systems and assemble much of that information. The company trained it on product information and previously sent emails so its responses would reflect the company’s way of communicating. Within a few weeks, it had removed much of the easy support work. Employees spent less time searching, transferring information and rewriting familiar answers.

Once that work was automated, the remaining cases were harder to handle consistently. Robonos encountered unusual orders, missing knowledge-base entries, unfamiliar questions and disagreements between internal information sources. Some responses were factually accurate but used a tone that did not suit the company. Each category gave the company a reason to add more rules, even as the individual situations became less frequent.

The cost of pursuing those cases became concrete when Wonderful Dental paid a developer for three days to automate a situation that occurred twice a month. An employee could handle each occurrence manually in 60 seconds. “We called that finishing the job. We were spending real money to delete a minute of work,” the insider says. Completing the system had displaced the question of whether that particular improvement justified the development effort.

Earlier success helped make continued development seem worthwhile. The first improvements had removed hours of repetitive searching, copying and rewriting, but later changes addressed increasingly unusual situations that took more effort to understand and implement. So each exception needed a fresh economic decision. Being part of an already useful support system did not establish the value of automating it.

The cost also continued after development. Each new exception introduced rules, maintenance obligations and opportunities for unexpected behavior, making the tool harder for employees to predict. The insider reports that repeated surprises reduced employees’ confidence and led them to work around it, erasing the intended time savings. “We spent money building the thing that undid the savings,” the insider says.

Those workarounds made employee behavior part of the investment’s result. Staff were spending effort avoiding a system intended to save them time, so another feature could add to the complexity causing the problem. Before commissioning more development, the company needed to reassess the remaining work and the cost of handling it. That reassessment also clarified which responsibilities people should retain.

Support review creates value by recognizing uncertainty and finding expertise

Wonderful Dental retained a review step in which Robonos assembles information and prepares a reply, then a person reads it, makes occasional corrections and identifies questions requiring deeper knowledge. Information scattered across systems, documents and employees can be brought together in a few seconds. The reviewer uses that easier access while remaining responsible for recognizing uncertainty.

Recognizing uncertainty requires knowing when to seek help. “The person reviewing the email often doesn’t know the answer either,” the insider says. The reviewer’s valuable action is to recognize that limit and send the question to the right person. Depending on the issue, that might be a dentist, a manufacturing specialist with detailed product knowledge, the company’s chemist, someone in operations or the insider.

Once a specialist answers, the company decides whether to add the response to its knowledge base. If the same question returns six months later, the employee handling the inbox can reuse an answer grounded in specialist knowledge. Making both the route to an expert and the decision to save an answer explicit gives the company a repeatable process for improving support.

The insider reports that this process lets junior employees handle a broader range of questions and reduces repeated interruptions to experts. The aim is to avoid asking a specialist the same question 15 times, keeping their involvement mainly for genuinely new issues. Expertise remains spread across the organization, while saved answers make more of it available to the people handling customer messages.

Dental-product questions make that division of responsibility important to Wonderful Dental. The insider wants a person to recognize an information gap before the system supplies an unsupported answer. Review therefore contributes value even when most drafts need little correction. “A short review isn’t a pointless one,” the insider says.

To explain why duration alone is insufficient, the insider gives a hypothetical job reduced from eight hours to 20 minutes. If the remaining minutes involve copying information between systems, further automation may be worthwhile. If they involve reading, correcting and routing unfamiliar questions, the company is willing to retain them. What happens during those minutes determines whether they are an avoidable expense or a responsibility worth funding.

Decide what to automate next by frequency, cost and consequence

The distinction between copying information and reviewing an unfamiliar question gives an automation budget a useful starting point. Repetitive production or information transfer may justify further automation; rare exceptions may cost more to automate than they consume; judgment and escalation may contribute directly to the result. Each category needs a separate spending decision because removing it changes the workflow in a different way.

The insider frames the financial assessment with three questions: “How much time is this taking, what does the leftover work cost, and what would it cost to remove it?” Answering them requires the task’s frequency, time per occurrence, current employee cost and labor savings already achieved. Review time belongs in the calculation, along with implementation, maintenance and the complexity introduced by additional rules.

Frequency can substantially change that calculation. The insider asks, hypothetically, how much automation spending makes sense for 10 emails a day compared with a hundred. Existing employee costs and the time already returned by the system affect the answer. A proposed $20,000 investment needs an assessment of the specific work it would remove and the operating improvement it would buy.

The insider’s other hypothetical examples show how differently that assessment can turn out. A task repeated thousands of times a day can justify development that saves only a few seconds per occurrence. A workflow occurring 20 times a day, with 10 minutes of total remaining effort, can make months of exception development a poor trade. The comparison depends on the remaining burden and the full cost of removing it under that business’s conditions.

Cost also depends on the consequences of changing who does the work. In support, Wonderful Dental considers how often unfamiliar questions arise, how long review takes and what an incorrect response would cost. In advertising, it considers whether people receive a useful starting asset and whether the finished work performs. Employee confidence matters too, because unpredictable behavior can create extra work elsewhere in the process.

These consequences can help set a cost-versus-frequency threshold for developing exception handling. Leaders can revisit that threshold as task duration, remaining labor cost and the cost of further automation change. Better models, changing demand, improved connections between systems or the disappearance of a workaround can make another investment worthwhile. Further automation becomes a stronger candidate when those changes lower its full cost while preserving the judgment the business needs.

Key takeaways for leaders

• Preserve creative finishing: Wonderful Dental’s two human-finished advertisements outperformed 20 fully generated alternatives on click-through. Marketing teams evaluating AI production savings need to account for the performance contribution of copywriting, retouching and layout.
• Price support exceptions individually: Wonderful Dental spent three developer days automating a task that took one minute twice a month. Support owners evaluating further automation need to include exception frequency, development costs, maintenance and employee workarounds in the calculation.
• Keep review connected to expertise: Human reviewers add value by recognizing unfamiliar questions and routing them to specialists. Support managers can make that process repeatable by defining escalation paths and saving reusable specialist answers in the knowledge base.
• Set an economic threshold for further automation: Budget owners need to compare the remaining task’s frequency, labor cost and consequences with the full cost of removing it. Revisit that decision as demand, model capabilities and integration costs change.