When An AI Tool Costs More Than It Saves
A communications team introduces an AI agent to monitor media coverage, prepare a morning brief and draft suggested responses. The tool produces the first version quickly. A senior employee then removes irrelevant mentions, checks the sources, corrects the tone and rewrites the recommendations before the brief reaches management. The company has not automated the task, it has added a new production stage.
Similar calculations are now appearing across marketing, communications, research and professional services. AI was initially approved because it promised lower costs and faster output. Many organisations are discovering that the direct technology bill represents only part of the expense. Human review, duplicated subscriptions, weak source material, unnecessary content and additional approval rounds can absorb the expected saving.
In some cases, a smaller AI model will solve the problem. In others, conventional software or a better workflow will be sufficient. There are also assignments where a competent employee remains cheaper and more reliable than an autonomous agent.
Companies need a clearer test: what does the task cost today, what will AI remove from the process and who will still need to check the result?
Calculate the full cost of the task
AI costs are difficult to interpret because companies rarely pay one predictable price for one finished piece of work. They pay subscriptions, platform fees and variable charges based on usage. An agent may call several models, search databases, revise its own output and repeat steps without showing the user how much processing took place.
A request that looks like one instruction can create dozens of billable actions.
Only 35 percent of organisations in a KPMG survey could fully capture and continuously monitor their AI spending. Another 42 percent had only partial visibility, while 13 percent learned the total after receiving the invoice. Half had already restricted or delayed AI-agent projects because of cost.
The calculation should not stop with the invoice. A useful business case includes the model and platform fees, implementation and integration, employee training, time spent checking and correcting output, legal and data-security review, additional work created for clients or approval teams, and the cost of errors or unusable material.
Consider a monthly report that previously took an analyst eight hours. An AI tool reduces drafting to one hour, but the analyst spends four hours checking the data and rebuilding the conclusions. The saving is three hours, not seven. When the tool, implementation and supervision cost more than those three hours, the business case fails.
The comparison needs to use actual operating figures rather than an estimated percentage of time saved.
Decide what the work is worth before choosing the model
Many companies use the most powerful model available for routine assignments. It simplifies procurement and gives employees access to one familiar system, but it can make simple work unnecessarily expensive.
A premium model may be justified for a difficult research assignment involving conflicting evidence, technical language and several jurisdictions. It is rarely necessary for transcription, formatting, approved-copy adaptations or a basic meeting summary.
A practical allocation could work as follows.
Use conventional software or a small model for transcription and formatting. These tasks follow clear rules and do not require advanced reasoning.
Use a small or locally hosted model to summarise approved internal documents. The source material is already defined, and the main requirement is controlled extraction rather than open-ended analysis.
Use a small or mid-range model for variations of approved copy. The wording can change, but the claims, tone and boundaries should already be fixed.
Reserve advanced models for complex research across conflicting sources. Their higher capability may be justified when the assignment involves technical interpretation, several markets or incomplete evidence.
Keep regulated external statements under experienced human control. AI can help compare documents, test wording or prepare a first structure, but an accountable professional should decide what the organisation is prepared to publish.
Leave final reputation, policy and client decisions with people. These assignments depend on judgement, institutional context and responsibility rather than the quality of the first draft.
The allocation should reflect the organisation’s own risk profile. A pharmaceutical company will classify tasks differently from a consumer brand. A bank may require protected infrastructure for material that another business can process through a standard cloud service.
The aim is to stop treating every assignment as a test of the latest technology. Model capability should correspond to the commercial value and consequences of the work.
Compare AI with the real alternatives
The alternative to AI is not always a person completing the same process manually.
A rule-based system may handle a repetitive calculation more accurately. Search may locate an approved document more reliably than a language model. A standard template may produce a stronger client report than an agent instructed to create a new structure each week.
The process itself may also need to be simplified.
A company publishing five weekly reports can ask whether all five are read and used before automating their production. A marketing team generating dozens of social-media variations can reduce the number of versions rather than deploy AI to produce even more. A consultancy with inconsistent source files should fix the knowledge base before placing a conversational interface over it.
AI can make low-value activity faster. It does not make the activity valuable.
Each proposed use case should be compared with four options: continue with a person, support the person with a limited AI tool, use conventional automation, or remove or simplify the task.
Full automation should win because it produces the best operating result, not because it appears the most advanced.
Measure the work after human review
AI pilots are often judged through speed, output volume or user adoption. These figures are easy to report and weak indicators of business value.
A communications team can generate 50 campaign ideas in an afternoon. The organisation may spend two days reviewing them. A client-service team can create personalised replies at scale, but low-quality messages may increase complaints and correction work. A research system can process thousands of pages while failing to identify the two facts management needs.
Useful measurements depend on the assignment. They may include total cost per approved output, employee time from request to completion, correction rate, the percentage of outputs accepted without substantial rewriting, the number and severity of factual errors, approval time, client satisfaction, reductions in supplier costs or measurable improvements in revenue and service.
The baseline must be documented before the pilot begins. Without the current cost, processing time and error rate, the company cannot show whether the new system improved anything.
A 60-second draft is not a productivity gain when it creates 45 minutes of review for a document that previously took 30 minutes to prepare correctly.
Keep the person who understands the context
An AI system can process more information than an employee. It does not automatically know what deserves attention.
A media-monitoring agent may detect every mention of a company without understanding which article could influence regulators, employees or an important client. A proposal generator may include all requested services while missing the commercial concern raised during the meeting. A multilingual tool may produce grammatically correct French that sounds unsuitable for a Geneva-based institutional audience.
These are often the reasons the assignment exists.
Experienced employees carry context from previous decisions, client conversations, internal politics and market expectations. Their value is highest where the work involves interpretation rather than production.
A narrower form of assistance may produce the best result. AI can collect material, compare documents or prepare a first structure. The employee decides what the information means, which recommendation is defensible and what should be communicated.
Replacing that person with a fully autonomous process may save production time while removing the part the client was paying for.
Set limits before an agent starts working
Autonomous agents create a different cost problem from ordinary chat tools. They can perform several steps, use external services and continue working until they believe the assignment is complete.
Companies need operating limits before deployment.
The agent should have a narrow purpose rather than a broad instruction to “support the team”. Usage limits should apply by task, user, department and month. The agent should have a stopping rule so that it cannot repeat searches or revisions indefinitely. Research should draw from defined databases, websites or internal sources where possible. Sensitive, ambiguous or high-value decisions should move to a person. One named owner should remain accountable for quality, cost and continued use.
The company should also review the agent after the pilot. A tool that appeared promising during its first month may become more expensive as usage spreads, or less useful once employees understand its limits.
Without these controls, an agent can become an expensive employee that nobody manages.
Swiss organisations should price risk into the decision
Swiss companies often have a strong financial case for automation because labour is expensive. The potential saving can disappear quickly in regulated or multilingual work.
A bank may generate an investment communication faster and then send it through legal, compliance and senior management because the text contains claims that require verification. A pharmaceutical company may automate a medical-information response but retain specialist review for every answer. A public institution may produce German, French and Italian versions, each requiring local adjustment.
The comparison should include the cost of those controls.
Confidentiality also affects the choice of system. Client files, employee cases, unpublished results and strategic documents cannot be entered casually into external tools. A cheaper public model may become more expensive once the company adds protected infrastructure, access controls and monitoring.
Supplier dependence has a commercial cost as well. Prices, usage allowances and model availability can change. A workflow built around one provider may become uneconomic or unavailable without warning.
Swiss companies should retain the ability to move routine work between providers and keep sensitive processes inside controlled environments. The organisation should own its prompts, approval rules, data and process documentation rather than embed them so deeply in one platform that switching becomes impractical.
Ask agencies to show where the efficiency appears
Clients should expect agencies to use AI where it improves the mandate. They should also expect a clear explanation of the commercial effect.
An agency proposal should state which activities use AI, whether third-party technology costs are included or charged separately, which information enters external systems, where human review takes place, whether AI reduces production hours, what happens to the time saved and who remains accountable for the final work.
The saving may appear as a lower fee, faster delivery, more thorough research or greater involvement from senior advisers. Any of these can be reasonable. A client should be able to identify the benefit.
A larger volume of material does not qualify automatically. More articles, concepts and campaign versions can increase the client’s approval burden. An agency that uses AI to narrow ten possibilities to three credible recommendations may provide more value than one that delivers 30 unfiltered options.
Service businesses should be assessed on the quality of the decision and the effort required from the client, not on how much material they can now generate.
Run a proper make-or-buy test
Before approving an AI use case, management can ask a short series of questions.
What problem are we solving? Describe the delay, cost, error or service weakness in operational terms.
What does the current process cost? Include employee time, supplier fees, approval and correction.
Which part will AI remove? “Support productivity” is not enough.
What work remains with people? Identify review, judgement, escalation and accountability.
What is the acceptable error rate? The answer will differ for an internal summary and a regulated client communication.
Can a smaller model or conventional tool do the job? Test the least expensive adequate option.
Can we stop the activity instead? Automating an unused report preserves waste.
How will we measure the result? Set the cost, quality and service metrics before the pilot.
When will we stop? Define the conditions under which the use case will be reduced or withdrawn.
This process should not require months of governance. It does require someone from the business, finance, IT and risk functions to examine the same task together.
An AI pilot should be allowed to fail
Companies often continue weak AI projects because senior management announced them publicly, employees invested time in them or the organisation wants to appear committed to innovation.
A pilot has done its job when it produces a reliable decision, including a decision not to proceed.
Stopping an expensive agent does not mean returning to manual work forever. The company may revisit the task when models improve, prices fall or the underlying process is redesigned. It may also retain one useful element while removing the rest.
A communications team might keep automated transcription and abandon automated recommendations. A professional-services firm might use AI for document comparison while leaving the client report with an analyst. A marketing department might automate approved adaptations but not original campaign development.
Selective use is usually more credible than an “AI-first” rule applied to every process.
AI should reduce work, not redistribute it
A business case is convincing when AI removes cost, improves quality or creates a service that could not be delivered economically before.
It is weak when the technology reduces drafting time but increases review, produces additional material that nobody needs or replaces a clear human process with an opaque variable-cost system.
The practical question is not whether AI can complete the assignment. In many cases, it can. Management needs to know whether the entire process is better after the tool has finished and the people around it have completed their work.
Sometimes the answer will favour an advanced model. Sometimes a smaller system or conventional automation will be enough. A trained employee may still provide the better service at the lower total cost. That is ordinary business discipline.
