Cover image: Concept and art direction by Javad Seif; generated with OpenAI.
The case for Specific AI Services (SAIS)
When large language models entered the mainstream, many institutions worried about falling behind. Their first response was: buy subscriptions, provide broad access, and encourage employees to use AI to become more productive and innovative.
The California State University, for example, made ChatGPT Edu available to students, faculty, and staff. Cal Poly Pomona has also offered faculty $2,000 grants to integrate generative AI into courses.
These initiatives reduce barriers to experimentation. But access alone does not create value.
A general-purpose LLM opens a door. It does not determine which organizational processes should change, how those processes should be redesigned, what safeguards they require, or how success should be measured.
The first wave: Buy access
During the first wave of institutional AI adoption, organizations purchased licenses and encouraged broad usage. Employees were told to use AI however they could: write better code, conduct deeper research, create more content, and work more efficiently.
But without defined use cases and performance measures, an important question remained unanswered:
Are employees using these services to create measurable value, or simply experimenting with an interesting new technology?
Experimentation is necessary, especially during the early stages of adoption. But an organization cannot treat experimentation itself as evidence of productivity.
Before providing AI access at scale, it should know:
- Which process is expected to improve?
- What is the current performance of that process?
- How will AI change the workflow?
- What risks and limitations must be managed?
- Which outcome will demonstrate that the investment worked?
Without answers to these questions, an AI subscription is simply another piece of infrastructure.
The second wave: Demand a return
The second wave began when the bills arrived and leaders started asking what they were receiving in return.
Organizations created AI initiatives, invited employees to propose projects, and offered incentives for integrating AI into existing work. These programs can uncover useful ideas, but they introduce another problem: a proposal is not evidence of improvement, and increased usage is not the same as increased value.
Someone must evaluate the proposals, support the pilots, review the results, monitor compliance, and determine whether each initiative should continue. Those activities create their own administrative costs.
Meanwhile, AI expenses are becoming more dependent on actual consumption. As J.P. Morgan recently observed, enterprise AI pricing increasingly resembles a utility bill rather than a conventional software subscription. Token-heavy applications can cost far more than ordinary chatbot interactions.
The right question, therefore, is not:
How can we encourage more people to use AI and how can we ensure they are using it properly?!
It is:
Where can AI produce an improvement large enough to justify its financial, operational, and managerial costs?
AI cannot replace process improvement
OpenAI, Anthropic, Google, xAI, DeepSeek, and other model providers are competing to create increasingly capable tools. But it remains the institution’s responsibility to understand and improve its own processes.
That responsibility cannot be outsourced to an LLM.
Organizations already struggle to improve processes while keeping daily operations running. AI makes the challenge harder because its capabilities and costs change so quickly. Keeping up with AI can feel like a full-time job.
This is why organizations should begin with a more fundamental question:
Do we already have an effective continuous-improvement system?
If the answer is yes, that system (and the people who understand the work) should determine where, when, and how AI is introduced.
In education, for example, we should identify professors and researchers who are already effective and who use AI systematically. We should study what they are doing, determine which practices are repeatable, and redesign teaching and research processes around the methods that genuinely improve outcomes.
The people closest to the work must help design the solution. A general AI committee cannot do this alone.
From general-purpose models to Specific AI Services
I use the term Specific AI Service (SAIS) for an application designed around a defined user, workflow, and measurable outcome.
A SAIS acts as an interface between an organization’s processes and the foundation models developed by AI providers. Instead of giving users an empty chat window and asking them to invent useful applications, it embeds AI inside a structured process.
A credible SAIS should:
- Address a clearly defined use case.
- Incorporate relevant domain knowledge and organizational rules.
- Preserve human review where judgment or accountability is required.
- Measure improvements in time, cost, quality, risk, or another meaningful outcome.
- Track usage and costs at the level of a task, case, or completed service.
- Use the most appropriate model for each step rather than depending unnecessarily on one provider.
Token counts matter, but they are only a cost input. Organizations ultimately need to understand the cost and value of completing the work.
Examples of Specific AI Services
In software engineering, tools such as Cursor embed AI within the development workflow instead of offering only a general chatbot.
In supply-chain management, a SAIS might estimate material-shortage risks, anticipate equipment downtime, support procurement decisions, and recommend mitigation actions.
In education, a classroom-specific service could connect AI use to learning objectives, assignments, feedback, and assessment while allowing professors to maintain control over the learning process.
The same approach applies to recruitment, career coaching, scheduling, maintenance planning, and other repeatable processes that involve uncertainty or professional judgment.
Two possible paths
Organizations will increasingly follow one of two paths.
Some will identify employees who understand both their work and the effective use of AI. They will listen to these employees and involve them in redesigning internal processes.
Others will rely on specialized companies that turn this combined process and AI expertise into products for particular industries and use cases.
Most organizations will probably use a combination of both approaches: internal experts will define the problem, while external platforms will provide some of the technology.
In either case, the value will move away from raw access to an LLM and toward the application layer that connects AI to real work.
RecomAid: What a Specific AI Service Looks Like
To make this idea concrete, consider the recommendation-letter process.
Traditionally, a student emails a professor with a résumé and deadline. The professor searches through old messages, tries to reconstruct the student’s accomplishments, writes the letter in a separate document, and manages revisions and submissions manually. The admissions committee later receives a collection of PDFs that must be reviewed individually.
A general-purpose LLM can help draft a letter, but it does not improve this fragmented workflow.
RecomAid, which I launched in 2025, redesigns the process from beginning to end.
Students build structured portfolios documenting their courses, research, projects, and other relevant experiences. When requesting a letter, they select the evidence most relevant to that opportunity, provide context and a deadline, and track the request without repeatedly emailing the professor.
Professors receive the request and supporting evidence in one place. They can ask for clarification, provide feedback, or certify experiences they can personally verify. RecomAid then generates a draft based on the student’s structured evidence, the professor’s feedback, and the professor’s writing preferences. The professor can edit the draft, manage versions, and decide what becomes the final letter.
AI assists with drafting, but the professor retains authorship and control.
RecomAid also extends the workflow to admission offices. Committees can examine recommendation letters across a selected group of applicants and ask questions related to their program’s priorities, such as research independence, communication, or perseverance. Responses remain grounded in the recommenders’ letters, with supporting evidence that reviewers can inspect directly. The platform complements existing application systems rather than attempting to replace them, and admission decisions remain with the committee.
This is what distinguishes a Specific AI Service from a general AI subscription. RecomAid does not merely provide access to a model. It organizes evidence, coordinates participants, enforces required information, supports human review, and applies AI at particular points where it can reduce repetitive work.
The objective is not to generate more text. It is to produce a better recommendation process.
Applying the Same Principle Elsewhere
The same philosophy guides my other projects:
- Alongwise.com applies AI to a structured career-coaching workflow, helping users identify aligned jobs, prepare tailored application materials, and manage their job searches.
- ArcAuthor.com, now in final testing, applies AI within a structured classroom environment designed around teaching and learning objectives.
- My industrial research platforms apply optimization and AI to asset replacement, maintenance and flight planning, production planning, and supply-chain reliability.
These applications differ substantially, but they share one principle: AI creates the most value when it is embedded in a process designed for a specific decision or outcome.
The Real Transition
An institutional LLM subscription can be useful infrastructure. But it is not an AI strategy.
Organizations do not need AI usage for its own sake. They need better teaching, better research, better decisions, more reliable operations, and more effective services.
The institutions that benefit most from AI will not necessarily be those that purchase the most access. They will be those that connect specific AI capabilities to specific processes and measurable outcomes.
If you request, write, or review recommendation letters, explore how RecomAid works. Admission leaders can also review the committee-focused workflow.
The principles of process improvement have not changed. Organizations must still measure throughput, cycle time, cost, quality, and efficiency—perhaps more rigorously than before. AI is a tool, not a guarantee of productivity. Its value must be designed into the process, measured, and demonstrated.
Javad Seif
August 14, 2026
Pasadena, California
Editorial note: The ideas, examples, and conclusions in this article are my own. OpenAI tools have been used to assist with the organization of my writings, drafting, and editing. I have reviewed and approved the final version and take responsibility for its content.
Comments
Loading comments…
Leave a comment