Before You Hire an AI Agent, Define a Finished Job
For a small media business, the meaningful comparison is not a subscription versus a salary. It is the cost of work good enough to deliver, and the responsibilities that remain afterward.
THE WORK. A production tool belongs inside a workflow, not in place of its acceptance standard. The question is what reaches the client ready to use. Illustrative photograph; not a Neuvieu employee or an AI-productivity experiment. Photograph: Nubelson Fernandes / Unsplash.
The decision in brief
Define an acceptable result before delegating. Count briefing, supervision, failed attempts and corrections alongside software charges. Then decide whether the best arrangement is a person, an agent, or a person working with an agent. The framework below is a proposed management method, not a measured Neuvieu savings claim.
Imagine a photographer asking a personal agent to update a portfolio, prepare client replies and find new contracts. An hour later, the agent reports that everything is done. The photographs are uploaded, the replies are polished, and the opportunity list is long.
Now imagine opening the results. The images are in the media library but not the gallery. One email promises a turnaround the photographer never approved. Two opportunities have expired. This is a hypothetical scenario, not a report of a Neuvieu trial. It exposes the difference between visible activity and usable work.
The business question is not whether the system stayed busy. It is whether delegation reduced the work required to produce an acceptable result.
A task is not a job description
An agent can take successive actions through tools and adapt to intermediate results. Anthropic’s engineering guidance distinguishes the record of those actions from the final state they produce: a booking claim is not the same as an actual reservation. This is vendor guidance about evaluation, not proof that its products outperform competitors. [1]
For a studio, apply that distinction before making a staffing comparison. Uploading an asset, approving its description and publishing it are separate responsibilities. So are drafting a reply and agreeing to a client commitment. Buying help with one step does not establish that an entire role has been replaced.
Neuvieu’s earlier guide examines which decisions should remain human. This companion asks the next question: once a task is eligible for delegation, how will the business know whether the arrangement is worth keeping?
Your invoice is not your productivity report
The agent market is not limited to usage pricing. Intercom bills its Fin service for defined outcomes. Its current chat-and-email documentation lists $0.99 for a resolution or configured Procedure handoff, while a qualification is $9.99. Those are selected prices, not a complete cost estimate for a business. [2]
The definitions matter as much as the rates. A resolution may be confirmed by the customer or assumed when the customer leaves after an answer without seeking more help. Intercom says the charge is deducted if the customer returns to that conversation needing assistance. A deliberately configured Procedure handoff can be billable; ordinary escalation is treated differently. [2]
There is nothing inherently wrong with defining a billable event. But an invoice measures the provider’s agreement with you. Your operating review must also ask whether the client’s need was met, how much follow-up remained and whether the outcome advanced the business. Silence, routing and satisfaction should not be treated as synonyms.
THE BRIEF. Define the deliverable, its evidence and the approval boundary before choosing who or what will produce it. Illustrative photograph; no participants in the cited studies are shown. Photograph: Quaritsch Photography / Unsplash.
The research does not supply a universal savings rate
There is credible evidence of useful assistance. A peer-reviewed study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond examined a staggered rollout to 5,172 customer-support workers. The final 2025 paper reported a 15% average increase in issues resolved per hour, with stronger gains among less experienced and lower-skilled workers. Humans remained responsible for conversations and could edit or ignore suggestions. It studied one firm, not autonomous media production or the number of people a studio should hire. [3]
A different setting produced a different result. METR’s July 2025 randomized study followed 16 experienced developers tackling 246 issues in familiar open-source repositories. With the early-2025 tools tested, tasks took 19% longer when AI was allowed. That is a historical finding about selected developers and tasks, not a current verdict on every coding assistant. [4]
The follow-up is essential. In February 2026, METR said participation, task-selection and timing problems made its newer experiment an unreliable measure of current productivity. Researchers thought greater speedups were likely, but described the evidence for their magnitude as weak. Reporting the old slowdown while ignoring that update would misrepresent the research. [5]
Neither customer-support results nor coding experiments can be pasted into a photography business’s forecast. They justify measuring the complete workflow under actual operating conditions, with consistent quality requirements, rather than assuming that a faster first draft produces a cheaper finished job.
“Done” has to survive review
METR’s March 2026 research note found that roughly half of certain test-passing, AI-written software patches would not be accepted by repository maintainers, even after adjustment for variation in review decisions. The agents had no chance to revise after feedback, and the note explicitly rejects treating the result as a fundamental capability limit. This is a software evaluation, not a rejection rate for creative work. [6]
The analogy is nevertheless useful as a management question: does the test reflect what the recipient actually needs? A correctly sized video can still miss the brief. An email can be grammatically sound and factually wrong. A working link can lead to an irrelevant opportunity.
Anthropic’s evaluation article reports that Descript organizes editing-agent evaluation around preserving the project, following the request and doing it well. That is a reported product-team approach, not an independent business-results study. It illustrates why creative quality requires more than confirming that a tool executed. [1]
Define four deliverables before choosing the tool
The following acceptance criteria are proposed examples. Adapt them with the person who owns the work and, where relevant, the client. Set them before reviewing the output, not after seeing what the agent happened to produce.
Portfolio maintenance
Accept the job when the approved images appear in the intended gallery, existing work is preserved, captions and alt text are accurate, and the public result has been checked. Record which devices or views were actually tested. An upload receipt alone does not meet that brief.
Client correspondence
Accept a draft when its services, prices, availability and scope match confirmed business facts. Missing information must be flagged. Sending the message or accepting additional work is a separate permission, not an automatic consequence of producing good prose.
Opportunity research
Accept a shortlist when every entry has a working original-source link, relevant eligibility, a current deadline and a reason it fits. Distinguish verified details from questions requiring contact with the organizer. A shorter usable list is preferable to padding the deliverable with uncertain leads.
Publication preparation
Accept a package when factual claims trace to supporting material, quotations match their sources, image permissions and credits are documented, and the intended page or draft contains the approved assets. Do not count fabricated interviews or unverified citations as completed research.
THE LAST MILE. A timeline, export or completed tool call can show progress. Review establishes whether the result satisfies the assignment. Illustrative photograph, not a demonstration of a named AI agent. Photograph: Georgia de Lotz / Unsplash.
Price the accepted deliverable
Use two views of the same assignment: cash paid and the economic value of human time consumed. An owner’s unpaid evening still has a cost, but valuing that time does not turn it into an actual payroll saving.
Cost per accepted deliverable = total allocated software, usage, setup, human work and recovery costs ÷ accepted deliverables.
This is a proposed accounting framework, not an industry certification. Include retries and abandoned attempts in the costs. Count each accepted deliverable once. A later correction belongs in the record; a withdrawn acceptance should change the denominator. If nothing is accepted, report a failed pilot rather than a productivity success.
Allocate shared subscriptions transparently and do not count bundled usage twice. For labor, use either an appropriate loaded employment cost or a stated owner-time assumption, not both for the same hour. Record one-time setup separately so a pilot does not conceal it or charge the entire setup anew to every task.
A hypothetical client-draft calculation
Suppose a batch produces ten acceptable client-response drafts. Allocate $20 of software subscription cost and $5 of additional usage to the batch. Briefing, checking and corrections consume three hours, valued at an assumed $40 per hour. The total economic cost is $145, or $14.50 per accepted draft. These numbers are invented for explanation, not observed Neuvieu results or quoted vendor rates.
Suppose the comparable manual workflow produces the same ten acceptable drafts in four hours with no incremental software expense. At the same assumed hourly value, it costs $160, or $16 per draft. Delegation saves one human hour and $15 of modeled economic cost. It adds $25 of software and usage spending; the one-hour reduction is not automatically money removed from payroll.
One extra hour of corrections would raise the assisted batch to $185, or $18.50 per draft. That sensitivity is the point: inexpensive generation can coexist with expensive supervision. Neither outcome should be presumed before measurement.
Keep elapsed time separate from human attention. An agent may run while you edit another project, but waiting that blocks delivery still matters. Track what the recovered time is used for. Extra production capacity becomes extra revenue only when there is demand and the business actually sells the work.
Hiring is a capacity decision, not a token calculation
The International Labour Organization’s 2025 occupational-exposure index evaluates tasks and concludes that job transformation is more likely than wholesale replacement because many occupations retain human-input requirements. Exposure estimates are not counts of people who have lost their jobs. [7]
For a media business, compare equivalent assignments before comparing entire roles. Consider three arrangements: a person with conventional tools; a person using an agent for selected steps; and a more delegated workflow with a named human reviewer. Hold quality, deadlines and permissions constant.
A more delegated arrangement is a candidate when the task is bounded, errors can be detected, and an affordable recovery path exists. Human-led work is the safer starting point when success depends on undocumented client expectations, access that should not be shared, or decisions the business has not authorized an agent to make. These are proposed selection rules to test, not findings that every task in those categories behaves alike.
Ask what workload remains after automation. If every output queues behind one senior reviewer, that person may become the constraint. If client development, creative direction and reliable delivery still need an owner, the hiring need may shift rather than disappear.
THE CAPABILITY. Staffing decisions should account for craft, judgment, relationships and learning, not just the tasks a tool can perform. Illustrative photograph; the operator is not identified as a Neuvieu worker. Photograph: Vanilla Bear Films / Unsplash.
Do not eliminate the path to expertise
Training is part of that calculation. Anthropic’s January 2026 randomized experiment involved 52 mostly junior software engineers learning an unfamiliar Python library. The AI-assisted group averaged 50% on an immediate quiz versus 67% without assistance, a 17-percentage-point gap; the overall speed difference was not statistically significant. The study did not establish long-term skill loss or effects in media work. Its qualitative associations between interaction styles and learning were not causal tests. [8]
That does not cancel the customer-support study’s finding that less experienced workers benefited. The tasks, environments and learning measures differ. Treat both as reasons to examine how assistance changes learning, not as grounds for assuming that junior workers either become unnecessary or always improve. [3] [8]
A practical response is to retain opportunities for junior creatives to make an initial judgment, explain an edit and diagnose a failed result. Evaluate those abilities alongside output speed. This is a proposed development practice, not a proven cure for skill erosion.
The ILO’s media-and-culture brief also treats compensation, creative control, consent and worker dialogue as central issues in the transition. A cheaper workflow is not a complete workforce strategy. Involve the people doing the work in defining what quality and sustainable capacity mean. [9]
Start with a bounded pilot, not a staffing promise
Select one recurring, low-risk assignment and write its acceptance checklist. Compare matched examples with your usual process. Record all attempts, not just successful demonstrations: task, tool and version when visible, permissions, cash charges, human minutes, elapsed time, revisions, acceptance decision and unresolved failures.
Test on material cleared for that environment. Check how a service handles client files before granting access. Use the minimum permissions necessary, preserve originals and require explicit approval for spending, sending client messages or publishing. A small efficiency gain should not override an unacceptable confidentiality or authorization failure.
Decide beforehand what would make you stop, revise or expand the pilot. Require actual reductions in total effort or another explicit benefit, such as better accessibility or quality. Do not lower the standard merely to convert activity into accepted work. Recheck after meaningful changes to the model, task or integration.
A small pilot cannot establish a universal productivity percentage or a safe headcount reduction. It can expose hidden review costs, reveal a task worth delegating and show where a human capability still carries the business.
The promise of personal agents is not that a founder never works again. It is that more people may be able to delegate work they previously had to absorb themselves. Make that promise accountable to the finished job.
Buy assistance. Measure accepted work. Keep responsibility visible.
Sources and method
This is a document-based business guide prepared with AI assistance in research, drafting and website production. No original interviews or controlled Neuvieu productivity trial were conducted. Task examples, the proposed pilot and the cost calculation are illustrative, not measured company results. Research findings retain their original settings and dates; software and support studies are not presented as media-industry forecasts. Product billing documentation was checked October 18, 2026.
All four photographs are illustrative, credited individually and listed by their source pages as free to use under the Unsplash License. They do not depict Neuvieu staff, clients or the experiments discussed.
Comments