At a recent trade fair, I came across a company selling enterprise AI agents as a subscription. The pitch was deliberately broad: give employees access to digital workers that could research, draft, analyse, retrieve information, and carry out tasks across different teams and subject areas.
Curious about how other companies were deploying AI, I asked their team how they dealt with the fact that such a flexible product would produce very different results depending on who used it. An employee who knew how to frame a task, provide the right information, and check the answer might get a great deal from the agent. Someone else might never move beyond asking it basic questions.
They didn’t dispute the point. As a SaaS company, they sold the subscription, and what happened afterwards was largely up to the customer. The agent could be purchased for every employee and still produce wildly uneven results; closing that gap wasn’t something they considered their responsibility.
What struck me was not the lack of an answer, but that they had deliberately placed the problem outside the product. That allowed them to sell the possibility of better work without taking responsibility for whether the work actually improved. The customer wasn’t buying an outcome. They were buying a tool and inheriting the work required to turn it into one.
A Product That Can Be Anything Doesn’t Have to Be Much in Particular
Open-ended agents are attractive products because they can be demonstrated against almost any problem.
A prospect mentions that salespeople spend too long preparing proposals, so the agent drafts one. Another buyer complains about slow research, so it produces a market summary. This makes for a persuasive sales experience, as the prospect sees the product respond to a recognisable situation on the fly. There is never the issue of the software solving a problem the customer doesn’t have, which can happen when a product is built around an assumed pain point.
The flexibility has real value beyond the demonstration. A broadly capable agent can follow an employee across work that would otherwise require several specialist tools, and it remains available when an unexpected task falls outside a predefined workflow. A general-purpose tool will naturally find more uses than a limited specialist toolbox.
The trade-off is that the same product can be extremely useful or almost useless depending on the person using it. The user still has to choose a suitable task, provide the relevant material, explain what they need, and know enough to reject a weak answer.
During a demonstration, the vendor does most of this work. After the purchase, every employee is expected to do it for themselves.
The Real Design Work Begins After the Demo
With the software subscription signed, responsibility for making the agent useful now lands awkwardly inside the company. A larger organisation may have an internal transformation team to work out how it should be used, but smaller businesses are less likely to have that luxury.
Suppose the company rolls the agent out to its sales team. One of the use cases shown during the sales process was proposal preparation, so the salespeople are told to start there. The task sounds straightforward until they try to use it on a live opportunity.
A credible proposal has to reflect what the prospect asked for, what was promised in earlier conversations, how pricing should be calculated, which case studies are relevant, and what requires approval. This information may sit in internal systems or be scattered across email threads, meeting notes, and the memories of experienced employees. The salesperson still has to decide which notes reflect the final position and whether an emailed concession overrides the standard price list, because a generic agent will not know these things out of the box.
For an employee trying to get a proposal out, this can feel like reconstructing the entire process inside the chat box before any useful work begins. Unless the benefit is immediate and obvious, many people will find it easier to reuse an old proposal or complete the task as they always have. This is how a highly flexible tool becomes classic shelfware.
The Best Users Improve Faster Than the Business
Agents of this kind tend to favour employees who are already curious, capable, and willing to experiment. These users can break a problem into parts, recognise what information is missing, and adjust an instruction when the first answer falls short. They also know enough about the subject to notice when the agent is confidently wrong. Put a strong tool in their hands, and they may become much more productive.
This raises the ceiling for employees willing and able to master the tool. A business, however, cannot assume that everyone will invest that effort.
Management can easily mistake stories from star employees for broader adoption. A handful of enthusiastic users produce striking results, present them internally, and create the impression that a new way of working is spreading through the company.
Elsewhere, the picture may be quite different. Some employees won’t know when the agent is suitable for the task, while others may simply have no interest in designing a process through trial and error. Users may expect too much from the agent without first giving it what it needs, then attribute the failure to the limitations of the agent itself and abandon the product after a disappointing attempt. Even those who develop useful methods may keep them as personal habits rather than turning them into something their colleagues can follow.
The company ends up with a collection of individual practices rather than a shared way of working. Its strongest users improve, but the quality of routine work may change very little.
Businesses Have SOPs for a Reason
Companies create standard operating procedures and templates because employees shouldn’t have to rediscover the best way to perform the same work every time it arises. The same principle should apply to business AI.
If a factory buys a new production line, management doesn’t leave the machinery crated on the floor and expect operators to work out how best to use it. Employees should help shape the process, but someone still has to install the equipment, define how the work will run, and prepare the team to operate it. Handing employees an agent and leaving each of them to design their own way of using it is not empowerment; it is unfinished implementation.
Similarly, simply deploying a general-purpose agent often leaves too much ambiguity with the user. The employee must notice that the tool could help, collect the relevant information, explain the task, and decide whether the answer can be trusted. The agent may reduce the effort involved in producing the work, but the employee is still responsible for inventing much of the method.
A good workflow captures what the organisation has already learnt about how the job should be done: which parts should be standardised, where judgement is required, and what should happen when something goes wrong. Without that surrounding design, every employee starts from an empty chat box and has to reconstruct the company’s working knowledge for themselves.
The Subscription Model Rewards Access, Not Results
A subscription business works best when the same product can be deployed across many customers with limited tailoring. Studying each customer’s work in depth, deciding where the agent should fit, connecting the right sources, and accepting some responsibility for the outcome would make sales slower and delivery more expensive.
If usage remains low, the company may be told that it needs better change management. If the answers are poor, users may need to improve their prompts or provide more information. If the agent doesn’t fit the existing process, the customer may need to redesign that process around it.
Any of those diagnoses may be correct, but together they show how much of the risk sits with the buyer. The vendor has demonstrated that the agent can perform a task under certain conditions; the customer must create those conditions across the company and maintain them over time.
Businesses do not invest in technology merely to possess it. A manufacturer buys machinery to increase throughput, reduce defects, or lower costs. A company buys business software because it expects invoices to be processed faster, sales opportunities to be managed more reliably, or customers to receive better service.
AI should be held to the same standard. A vendor may count the number of subscriptions sold or agents provisioned, but what matters to the customer is whether the work itself has improved.
The Harder Product Is a Working System
Giving a company access to a strong model has become relatively easy—the LLM providers have made sure of that. Turning that model into a defined way of working is something much harder. Someone has to decide what triggers the work, what information the software needs, which rules it must follow, where human judgement remains necessary, and how the result should move into the next step.
If the vendor doesn’t take responsibility for those decisions, they fall to the customer or, worse, to each employee using the agent.
This is the work we believe an AI systems integrator should be responsible for. At Mesograde, we start with the result the business needs and examine how the work producing that result is currently done, before deciding what must be put around the AI for it to work reliably. Only then does it make sense to decide what role an agent should play.
Agentic development has made custom software much faster and cheaper to build. A small team can now deliver systems that would previously have required more developers and a much longer project, making it practical to build around the customer’s work rather than forcing the customer to adapt to off-the-shelf products.
As the cost of writing software falls, understanding what to build becomes more valuable. Traditional SaaS came with fairly rigid boundaries, and much of the way of working was already embedded in the product. A consultant could therefore often remain near the surface: configure the fields, map a few approvals, train the users, and help the customer adapt.
AI-native software is far more malleable and flexible. It can be shaped around many different ways of doing the same job, which means surface-level consulting no longer cuts it. The provider has to get into the guts of the work: the informal practices, competing sources, exceptions, judgement calls, and handovers that rarely appear on a high-level process map. Without this, customisation does not solve the problem. It produces a highly tailored system built around an incomplete understanding of how the work should be done.
The more customisable the technology becomes, the less room there is for shallow consulting. Agentic development makes tailored systems economically practical, but it also allows the wrong system to be built and deployed much faster.
The Best AI System May Give the User Less to Figure Out
An open-ended agent performs well on stage because the audience can watch it respond in real time. Its range is visible, and the empty chat box encourages people to imagine how many different tasks it might handle.
A system built for a specific piece of work looks much less dramatic. The user sees a prepared brief, a flagged exception, a draft that already follows company rules, or a suggested next step based on information gathered in the background.
What looks like less flexibility is actually good design. Before the employee opens the screen, the company has already decided which parts of the work should be standardised and where individual judgement is still needed. The software doesn’t ask the employee to choose from everything the model could possibly do because its role in this particular job has already been defined.
A company that needs consistent results from recurring work cannot stop at access to an agent. Access gives employees possibilities; someone still has to turn those possibilities into the company’s normal way of working.
Selling the agent is easier.
The harder work is deciding what should happen after the employee opens it.