How to Start Smart With a Plug and
Play AI Model
There
is a better approach to AI for businesses that don’t want to shell out millions
of dollars to Google or OpenAI to build their own custom version of ChatGPT.
EXPERT OPINION BY HOWARD TULLMAN, GENERAL MANAGING PARTNER, G2T3V AND CHICAGO HIGH TECH INVESTORS @HOWARDTULLMAN1
Oct 7,
2025
It’s hard to say whether
there are more columns arguing that, if you haven’t already infused your
business processes with AI you may be too late and
consequently doomed, or that the vast majority of corporate AI projects to date
have been miserable and costly failures. The latter parade of horribles is led
right now by recent M.I.T pronouncements asserting that
something like 95 percent of the serious large organization efforts have
failed.
There are plenty of
reasons offered for the failures of these initiatives, but they sound
unsurprisingly similar to the explanations offered when the pace of adoption
and implementation of any new tool or technology is slower than the hyped
expectations—and when the promised payoffs in savings and enhanced productivity among others fail
to materialize immediately. The real wonder is why resistance to change comes
as any surprise to anyone at this late date.
In the case of AI, there
is another debilitating and discouraging aspect of the process, which is that
even after all the time and analysis and reviews aimed at attempting to
determine why some given project hasn’t worked, there’s almost nothing to be gathered
or learned because the “black box” that you’re attempting to interrogate and
explore is just that—a black box which performs its function in somewhat
mysterious and unclear ways that no honest auditor would even pretend to fully
understand.
Needless to say, it’s
hard to extract lessons or process improvements in order to achieve better
future results when the underlying engine driving the outputs (a) has no
memory, (b) doesn’t really “learn,” and (c) doesn’t accept or incorporate
input, instructions, and corrections in any consistent or measurable fashion.
It just is what it is and you’re offered the opportunity to take it or leave
it. This isn’t the most appealing context or foundation to try to build the
next phase of your business upon.
To be clear, the
opaqueness and obstinance of the primary AI foundational platforms aren’t even
the most pressing concerns for most smaller companies and businesses which are
trying to incorporate AI into their organizations. These firms — which is to say the
vast majority of all the entities in the world except the
mega-corporations — can’t afford to spend their scarce and critical resources on
broad AI experiments which may or may not ever come to fruition. Even more
importantly, they are facing a much more pressing set of risks and concerns
within their own companies presented by the actions of their own
people.
Close to 60 percent of
employees are using AI tools not approved by
their employers, according to a recent Cybernews survey.
Worse yet, more than half admit that their direct supervisors or managers know
about their use and don’t object, while another sizeable group reports that
their managers don’t care. In fact, the survey found that it was executives and
senior managers who were the most likely to be using these tools. But the
critical finding from the survey is that around 75 percent of the respondents
using unauthorized tools admitted to sharing sensitive data with them. And the
companies themselves have no control or even information about which team
members are doing what. Once the data enters any of these AI platforms, the
user has no control over whether the information is stored, reused or exposed
to third parties.
There is a much better
and smarter approach for SMBs and, frankly, for just about any business not
willing to shell out millions of dollars to Google or tens of millions of
dollars to OpenAI to build their own custom version of ChatGPT. That’s what I
call the “plug and play” solution. Simply stated, this is a “portable” query
engine that any business can install on their own premises and in their own
machines and networks.
Typically, there is an
initial fee for installation and implementation of the system—roughly $50,000
or less—and then a monthly or annual recurring cost which may vary with the
volume of usage but which isn’t likely to be a major cost and can be cancelled
at any time. This puts a known cap and realistic limitation on the entire cost
of the experiment.
The firm designates and
assigns an underlying selection of data for the engine (the corpus) to operate
on, interrogate, and extract answers from. All employees have access to the
system and can simply ask whatever questions about the data which they need
further information about, request analytics or compilations of segments or
selections of the data, or request the creation of documents, reports, or
presentations of parts of the data which are relevant to areas or questions
they are dealing with. One key consideration is that by limiting the dataset to
material relevant to the specific business and activities of each company, the
fit of the engine to the likely inquiry scope is better and the overall costs
of operation as well as response time are reduced since there is no need to
“search worldwide” or attempt to “boil the ocean” in order to develop and
respond with timely and accurate information.
Note here that the
system has two critically important guardrails in place which govern every
query. First, it knows what it knows, so it’s designed to report back that it
can’t answer certain questions. Second, it knows the scope and limitations of
what queries are proper topics for inquiry and rejects irrelevant or immaterial
prompts. For example, the system doesn’t know the meaning of life. In addition,
the system creates a log file of every query and every answer which can be
reviewed by management to track activity and also readily determine how often
and how effectively the system is being used.
Other crucial elements
of these systems are that there are evaluation, review and edit functions which
in essence let the system “learn” and improve its answers and responses on an
ongoing basis as well as allowing for corrections and updates—all within the
control of the company itself and without ever allowing any such proprietary
information to leave the premises. Security and control as well as demonstrable
growth in efficacy are all central to the system. Because the system also
permits any user to evaluate and rank the value and accuracy of any answer,
there is a real time feedback loop that encourages and rewards employee use and
engagement. This prospect makes it far more likely that team members will
adopt and regularly employ the system especially as they see their own input
and roles in the process incorporated into the system.
Bottom line: you can put
a toe in the water at a reasonable cost and try to expose your people to the
upsides of these next-generation tools without incurring substantial expenses,
without exposing your proprietary data to the outside world in an uncontrolled
and unaccountable manner, and without encouraging or permitting your own team
members to go outside the walls and waste time and energy randomly exploring
the large LLMs without any particular guidance, support or benefit to your
firm.
I have installed two
sample systems on my own website in the lower
right-hand corner. These are free to use and to experiment with for all
readers. The white link has a corpus which consists solely of my hundreds of Inc. magazine columns published over the
last decade. The black link is a much broader corpus of all my books, columns,
speeches and presentations, etc. which will attempt to answer a far broader
range of questions. Give it a try before you decide to buy. Test before you
invest.
