Don't Fall for
A.I.'s False Promises
Hot applications such as
ChatGPT are prompting many businesses to dump money into A.I. way too soon.
Most companies would be better advised to develop analytic systems that may be
less leading edge, but more productive.
BY HOWARD TULLMAN, GENERAL MANAGING PARTNER, G2T3V AND CHICAGO HIGH TECH INVESTORS@HOWARDTULLMAN1
We're
in the midst of another FOMO frenzy over artificial intelligence (A.I.),
and especially large language model-based systems like ChatGPT. Seems like
millions of companies and dozens of VCs are rushing to invest billions in what
they hope is going to be the next best thing.
The
term A.I. has become sloppy shorthand and an overused, misapplied buzzword
(like crypto, blockchain and Web 3.0) that's so broad as to cease to be
descriptive. A.I. is so widely misunderstood that the only thing most of these
folks actually know about the subject is that they certainly don't want to miss
out on it. Even if they don't really know what the "it" is that
they'd be missing.
But
what's even more discouraging is that this is one of those semantic swamps and
media-driven hypes that you wade into and, as you're slowly sinking your
precious capital and tech resources into the bog, and your team is finally
learning something concrete about the relevant applications of these new tools,
you sadly discover that the money you've spent and the "tools" you
built or licensed have little or nothing to do with the actual operations of
your particular company.
Most
businesses won't need or find practical, cost-effective uses for actual A.I.
tools for years to come, if ever. Because their day-to-day information
requirements and the characteristics and attributes of their products,
services, markets and customers simply don't line up enough with the actual
information outputs of these systems. It's like asking for tax advice from a
philosopher rather than an accountant. You might get an answer - and one that
might be directionally correct--or, in the current ChatGPT world, one that
might be completely made up. But that advice is nothing that any sane
person would rely on.
If
you're really trying to solve certain problems in better and faster ways,
including scheduling and routing, just-in-term supply chain projections, cost
and damage estimates or pricing matrices, A.I. might be the wrong approach. You
don't need to spend the time and money training your people on new systems to
build precise and iterative prompts and inquiries to interrogate huge,
generalized data aggregations to get responses based on tons of information
that have nothing to do with the specifics of your business; or the marketplaces,
geographies and regulatory environments in which you operate. It's
overkill and very much like using a hammer to put out the flames when your hair
is on fire. Painful, costly, and not particularly helpful.
On the
other hand, spending your time and energies on creating analytical and
heuristic systems that are practical, cost-effective, relatively rapid, and
readily accessible -- and built from your own datasets, related third-party and
adjacent others, shared and documented experiences, and archived knowledge
bases-- is the smartest way to enter this new generative world. Let's just say
that the simple idea of machine learning is a lot more useful and
understandable than all the gobbledygook about A.I.
The
premise couldn't be simpler: you don't need an oracle to predict next month's
likely demand for specific products if you already have a decade of prior sales
data, an experienced sales force, good info about your competitors' pricing
strategies and a relatively stable marketplace in terms of regulation or other
external factors. Every decent sales organization, every restaurant, and any
smart entertainment venue has its own version of a "beat yesterday"
book, which helps them look backwards and plan ahead. Most businesses have been
doing their own variations of these kinds of inquiries for years in some
combination of manual and mechanical approaches, so there's virtually no new
training involved.
The
trick is that any decent machine can manage, absorb, manipulate and display
results, variations and projections for thousands of different products and
scenarios at multiple price points, in minutes, far more accurately than even
your best salespeople. The critical change is that the speed and abilities of
even the base level computing machines have grown exponentially while the costs
of accessing and employing the processing power are now close to zero.
The
use of machine learning isn't limited to applications like sales projections.
Millions of other business interactions occur every day that are subject to
known rules, processes, regulation, and limitations that are also capable of
being accumulated, archived, analyzed and converted to real-time tactical
instructions and directions, to be played back to customer-facing team members
in hundreds of different roles and positions.
Balto deploys one
shining example of such a system, which equips customer service agents and
their supervisors and managers with immediate, in-stream, customer history,
transaction data, appropriate responses and escalation directions, all drawn
from current interactions, company policies, and historical
experiences in similar cases. If
you have to have a name, I'd suggest that you more properly call this type of
assistance "augmented intelligence" as long as you understand that
it's the intelligence of the human end user that's being enhanced and extended
rather than some novel output being created by a miraculous black box that is
answering questions that no one needs to ask.
The
improvements in customer engagement, and in customer and CSR satisfaction, and
the gains in time, accuracy and productivity aren't the product of new
discoveries. They're simply the result of better and more quickly equipping
team members with the data and tools they need to do the best job possible in
the shortest amount of time. Slicing and dicing at scale isn't some new magic--
it's a case where we're more focused on the power and value of such analysis
and also have the ability to convert massive amounts of data into useful and
actionable information.
Snapsheet creates claims management software for the insurance
industry; its customers provide access to millions of interactions among claims
adjusters, repair shops and consumers who may be insureds or claimants. As
unique as each and every crash may seem to the parties, the nature of the
damage to the vehicles, costs of repairs and time required, are remarkably
consistent when similar cars and circumstances are present. In addition, the
claim documentation and submission processes have also been streamlined and
standardized by major insurers, which means that the vast majority of claims
being processed on any given day for comparable vehicles are virtually
identical. Likewise, the entire repair process is fully understood. And
because virtually all the descriptive language regarding damaged parts and
systems is also available in commonly employed digital designators, all claim
submission material can be captured instantly by the entry systems and flowed
directly into the analytical engines.
Here
again, as Snapsheet continues to demonstrate, circumstances are ideal for the
expanded application of machine learning. By relying on millions of
accumulated prior damage examples, the system can analyze, document and process
claims as they are submitted and automatically create initial estimates without
any human involvement. These initial estimates can then be quickly reviewed by
adjusters, edited, or corrected where necessary, and returned in real time to
their insureds or claimants, along with directions and authorizations to the
body shops to get the damaged vehicles repaired and returned to their owners
quickly.
If
this process seems remarkably mechanical and straightforward, that's because it
is. There's no magic. There's no novel "intelligence" or native
thought. There are constantly improving ways to use increasing computer power
to manage massive amounts of data. This is precisely the value and virtue of
using computers and other machine learning tools to replace endlessly
repetitive human actions, including data entry, with systems that can
immediately assemble, process, and evaluate materials to increase productivity,
avoid entry and calculation errors, assure consistent legal and regulatory
compliance, save time and improve everyone's satisfaction.
The
most intelligent thing that any business owner can do today is to take stock of
their own operations, determine which parts of the workflow have the essential
attributes that can be optimized and enhanced through the application of
machine learning, and get started implementing those kinds of improvements.
There's nothing artificial about the results and savings you'll see in no time
at all.