The Hidden Work of AI Transformation: Lessons From the People Doing It
I was recently asked to write about AI and digital transformation. I reached out via Help a Reporter Out (HARO) for thoughts, but my article leaned more on frameworks and anecdotes. I took the dozens of responses I received and used ChatGPT to analyze them, extracting value by identifying common themes across the interviews. The most revealing line came from someone describing a mistake. Chris Wright, founder of Fifty Five and Five, said an automated pass over his company’s website changed a word’s spelling inside a link, breaking it. The same pass rewrote an article’s description to say the opposite of what the article said. Both changes looked plausible at a glance.
“Checking is a lot of my job now. Did not see that coming.”
That observation connected with a recurring theme across the responses. AI can remove work, expand capacity, and make previously impractical ideas worth trying. It also creates work that organizations need to recognize, fund, and assign. Verification, context maintenance, exception handling, and coordination become part of the operating model.
These responses are practitioner accounts, with many coming from people who sell AI products or services. They help identify patterns and questions worth pursuing. They do not establish general productivity gains or prove that any particular implementation caused a reported business result.
What they do suggest is that leaders need a more complete picture of the work before declaring transformation a success.
Recurring Lessons
Repair handoffs before accelerating tasks
Several respondents described improvements that depended on changing how work moved between people and systems.
Harsh Patel, founder of MRI technologist training company TeslaMR, described an AI assistant that helped staff draft feedback for student evaluations. Initially, staff requested the draft in Slack and copied it to an evaluation page. That handoff lost the ticket’s tracking. The revised process saves the draft on the ticket, where staff can review and edit it. Sending requires a separate action through an explicitly selected staff account.
“But if it’s a complex multi-step process with hand-offs involved, I’d rather fix the process first, and then automate it.” — Marta Pawlik, co-founder and director, Laik.
The useful change was in the relationship between the draft, the record, and the person authorized to send it. Faster writing alone would have left the tracking problem intact.
Jeremy Su, co-founder and CEO of Bearly Art, described a related discovery across his e-commerce operations. Connecting data from several platforms made problems visible that individual systems had made difficult to diagnose. In his account, the trouble lay in the human handoffs between them. Rehan Ahmed of Smart Web Agency offered the same foundational lesson through a farm booking project: making one record useful across the visit came before considering where AI might help.
For leaders, the implication is straightforward. Map where information changes hands, where responsibility becomes unclear, and where people reconstruct context. Investigate those places before selecting a model or buying another tool.
Budget for verification and correction
Then follow the work into verification. Tudor Brad, managing director and co-founder of software testing company BetterQA, echoed Wright’s experience. Brad argued that faster production creates more to check. Brian Peterson, Dialpad’s CTO and co-founder, raised the same concern in software development: faster implementation can overwhelm review, even as security, architecture, and production quality still need attention.
A team that generates twice as much code or content has created a capacity question for everyone who must evaluate it. That includes the recipient asked to absorb another long document. Organizations should account for review queues, correction cycles, and the attention required to distinguish a plausible answer from a useful one.
“If implementation accelerates while review becomes overwhelmed, you’ve only moved the bottleneck, not eliminated it.” — Brian Peterson, CTO and co-founder, Dialpad
This also changes how leaders should think about expertise. If experienced people become the final check on a growing stream of output, their time becomes an explicit constraint. A productivity plan needs to show how it will manage that constraint. Assigning review to someone already at full capacity is a staffing decision, even when the project budget treats it as free.
Give reviewers evidence, time, and authority
Accountability requires more than a name beside an approval box. Martin Mulyadi, an accounting professor at Shenandoah University, put the requirement clearly: “Human review needs evidence, time, and authority to change the result.” His accounting example separated a proposed journal entry from authorization of a material posting, with supporting information, model versions, and overrides available for later investigation.
Rachel Tonner, VP of product marketing at Amplience, approached the issue from a content operations perspective. She emphasized that experts should be able to change outputs and contribute their knowledge of the product, brand, and customers. A review interface limited to approving or rejecting can underuse the expertise it supposedly preserves.
“Human review needs evidence, time and authority to change the result.” — Martin Mulyadi, professor of accounting, Shenandoah University.
Together, these observations make human oversight a requirement for operation. A reviewer needs to understand the evidence, know the decision’s boundaries, and have permission to stop or alter the work. The organization also needs to specify which actions can proceed autonomously and which require additional authorization. A person nominally available to intervene provides little protection when they cannot see the problem or act in time.
AI Transformation: When AI Changes the Business Relationship
The main article reports patterns that consistently emerge across the interviews. I also looked for outliers. Some distinctive practitioner accounts suggest that AI can change relationships with suppliers and buyers, as well as work within an organization. Here are three examples.
A prototype can change a renewal negotiation
Chris Wright, founder of Fifty Five and Five, described responding to a SaaS price increase by using AI-assisted development to build an alternative around the portion of the product his company actually used. He didn’t cancel the existing service. Having an alternative changed the budget conversation about price, requirements, and replacement.
The strategic question is what it would take to operate and maintain a credible alternative. A working prototype can help a buyer examine that question, even when the eventual decision is to renew.
The workflow itself can become an asset
Francis Brero, VP of AI strategy at HG Insights, said the internal agents and workflows MadKudu developed were part of what interested its acquirer. His account suggests that an organization’s accumulated operating knowledge can gain additional value when captured in reusable systems.
That raises questions about who controls the instructions, corrections, integrations, and context that make those systems useful. Brero’s account doesn’t assign a value to the workflows, but it suggests managing them as business assets.
The displaced work may sit outside the payroll
Evgenii Arsentev, CEO of AskDocDoc, described bringing Facebook targeting under an AI agent that he manages through voice commands. The startup had previously used an outside agency because it could not afford an employee dedicated to the work.
In his account, the change displaced an external service relationship. An analysis confined to employee headcount would miss it. Leaders assessing AI’s effects should examine agency spending, contractors, and other purchased services alongside internal roles—and account for the supervision and maintenance they take on when bringing work inside.
Build shared context people can depend on
Shared context needs owners, too. Barr Moses, co-founder and CEO of Monte Carlo, described an early period when employees built useful agents and ran private AI sessions without creating much collective benefit. The company later identified reusable workflows, built shared data and context layers, and established cross-functional ownership. Its internal go-to-market hub became a way to share those capabilities.
Chad Stamm, director of AI integration at TMC Digital Media, described the maintenance side of the same problem. His team compares AI drafts with the versions that actually get published and feeds the differences back into shared context. It moves shared files into a repository where people review changes.
“Individual AI saves a person time, but institutional AI changes how an entire company operates.” — Barr Moses, co-founder and CEO, Monte Carlo
The lesson extends beyond storing approved prompts. Organizations need a process for deciding which corrections become shared knowledge, who approves changes, and how stale instructions are retired. A useful personal workaround becomes an organizational capability only when other people can find it, understand it, and depend on it. That transition requires ongoing work.
Moses captured an adoption principle worth noting: “We made the governed path easier than opening a private AI window.” Governance can support use when the approved approach gives people better access to the context and capabilities they need.
Choose where AI adds value
Another recurring lesson concerns the place of AI within the solution. Rob Smith of OSF Digital described combining AI interpretation with conventional automation in a refund process. AI can help interpret what the customer means; established business rules can determine whether to authorize a refund. Andreas Huttenrauch, founder and CEO of InfoLobby, similarly described deterministic workflows with AI steps introduced where interpretation adds value.
Mike Dymott, principal consultant at ISM, offered an even more direct qualification. In a previous CRM role, his team faced pressure to pursue savings through AI and found that standard automation could address the need. His account challenges the assumption that an AI initiative must end with AI performing the work.
“Also, don’t assume the answer is AI. Focus on the business needs.” — Mike Dymott, principal consultant, ISM.
That distinction belongs in the business case. Compare the alternatives: change the process, use conventional automation, add AI to selected tasks, or combine them. The choice should follow the requirements for consistency, interpretation, maintenance, and cost. An organization gains little by making a repeatable transaction less predictable just to increase AI in the workflow.
Measure the completed workflow
Measurement must cover the completed work. DuBose Cole, co-founder of New Classic, described looking for reliable client deliverables that do not require someone to rescue the process. Mulyadi called for baselines that include turnaround time, material errors, review hours, and total cost. Mark Wonderlin of Mosaic Media Films cautioned against attributing revenue changes to AI when staffing, marketing initiatives, or sales processes also changed.
Their common concern is the gap between activity and results. Logins can indicate adoption. Token consumption can show use and cost. Neither demonstrates that customers receive better answers or that employees finish useful work with less effort.
“Counters measure activity; you have to look at the work itself to know what’s happening.” — Francis Brero, VP of AI strategy, HG Insights.
The practical unit of measurement is the workflow through completion, including checking, correction, exceptions, and recovery. Where feasible, compare similar work with and without AI. Where several changes happen together, describe the result as a result of the redesigned workflow and be precise about what can be attributed to AI.
People can disagree on how far that attribution needs to go. Some respondents emphasized the combined business result; others wanted stronger comparisons. Leaders still need enough evidence to decide which components deserve further investment and which can be simplified or removed.
Make vendor dependency a deliberate business decision
Finally, vendor dependency needs an explicit business decision. Francis Brero, VP of AI strategy at HG Insights, emphasized keeping workflow orchestration separate from model choice and retaining ownership of workflows that embody the organization’s distinctive way of operating. Peterson asked practical questions about exporting data, changing permissions, continuing through an outage, and moving to another provider.
Smith placed more weight on moving quickly and developing internal skills, accepting that switching costs and rework may be worthwhile. That disagreement is useful. No universal level of dependence fits every workflow.
A routine support capability, a source of competitive advantage, and a process that can make consequential commitments deserve different assessments. The decision should include the cost of operating the system, the capabilities needed to maintain it, and the effort required to change direction. Building internally creates dependencies as well.
Plan for the work that follows generation
Across these accounts, the strongest lesson is the organization’s ability to absorb the change it creates. Producing more becomes valuable when the business can verify, use, maintain, and take responsibility for the result. Time released from one task becomes capacity only when people can apply it productively elsewhere.
Before expanding an AI initiative, leaders should be able to trace what happens after generation: who checks the work, what evidence they use, what authority they have, how corrections improve the shared system, and what it costs to reach an acceptable outcome. Those answers describe the work of transformation. They belong in the plan from the beginning.
“Buying speed today should come with a clear understanding of what changing direction will cost tomorrow.” — Brian Peterson, CTO and co-founder, Dialpad
For more serious insights on AI, click here.
Did you find this analysis, The Hidden Work of Here: Lessons From the People Doing It, useful? If so, please like, share, or comment. Thank you!
The cover image is AI-generated from the author’s prompt.
Leave a Reply