Bumper to Bumper: Your Competition Is Just a Browser Tab Away

A while back, a client of ours was getting ready to pitch an AI agent product based on their firm’s expertise to Fortune 500 buyers (for the sake of anonymity, let’s say they advised hospital quality teams on investigating serious patient safety incidents: they don’t, but their work is comparably technical and high-stakes.) 

Among other things, the agent coached professionals through writing a root cause analysis report. We’d built in a structured interview workflow to make sure users covered every critical point before anything got written. However, in the original version of the AI agent, that’s where the conversation stopped.  Basically, we were having the AI agent help the user think through the whole investigation only to have it say “Great. Now go write your report.”

This led our team to ask an obvious question: “Why don’t we let the AI agent produce a draft report, itself?”  

After all, the user had already supplied the facts. The agent had already helped shape the analysis. Writing a draft that followed the firm’s best practices was the natural next step, and the user would never need to leave.

The client’s answer: “Maybe for version two… Let’s walk before we run.”

But consider the interaction from the end user’s perspective: sitting there with their notes from our AI agent, starting down a deadline to get the report written. There’s maybe a 70-80% chance that the user is going to open ChatGPT or Microsoft Copilot, paste everything in, and type “Write this up for me.” 

In other words, we were about to do all the expensive work of gathering the facts, applying the firm’s expertise, and structuring the analysis, then hand the final, most visible part of the customer experience to a generic chatbot. Basically let ChatGPT take the victory lap simply because we decided to pull over at the end of the race. 

We knew our client’s customers wouldn’t be happy with an 85% solution, so we took a risk and built the drafting step anyway, then handed it over with a note that amounted to: “Merry Christmas. Take it for a spin and see if it matches your frameworks, if you don’t like its output we can shut it off.” 

Our client did like the output, once they saw it. And they immediately showed it to the next Global 100 buyer they met with, who was very much impressed.

Now, one impressed buyer isn’t proof of concept, let alone proof of adoption. But it proved something: a feature can be ahead of a company’s initial roadmap without being ahead of its customers’ needs.

Friction and Drag: Contending With Market Forces

The situation we just described should be familiar to anyone with an MBA as a textbook example of “Porter’s Forces” in action. 

In 1979, Harvard professor Michael Porter identified five forces that shape every business, one of them being “threat of substitutes”.  Whenever a customer asks general-purpose chatbot like ChatGPT to provide a strategy recommendation or create a presentation deck instead of hiring a consultant or a graphic designer, that’s the purest substitute any industry has ever faced.

And the threat isn’t hypothetical. Microsoft is already pushing Copilot into HR workflows, while OpenAI has launched a financial services product that can help bankers produce research, valuation models, and client presentations. The companies building the underlying AI technology are increasingly moving up the value chain into the specialized work that traditional software vendors and professional services firms have historically owned.

While the dynamics of competition haven’t changed much since the 1970s, the speed of competition has increased dramatically. With AI models and systems, the substitutes get better every quarter, with generic AI tools able to stand in for an astonishing range of tasks.  The practical implications being:

• Substitution is nearly frictionless. A customer can open ChatGPT and finish the work your specialized product left undone.
• Semicompetitors quickly become direct competitors, as AI makes adjacent capabilities unusually cheap to add. Once an AI system has the data, context, and integrations to execute one workflow, extending it into the next may require relatively little additional engineering.
• As computers become capable of more kinds of work, tech-based rivals can become players in more industries. Your traditional rivals are now joined by startups, traditional software vendors who used to support your work but now threaten to replace it, and customers building their own AI tools.

Plus, you’re not only competing with the AI tools your customer organizations officially purchase, but all the “Shadow AI” tools their employees use on the sly. Microsoft and LinkedIn’s 2024 Work Trend Index found that 78% of AI users brought their own AI tools to work, and 52% were reluctant to admit using AI for their most important tasks. In professional services, the Thomson Reuters Institute’s 2025 survey found 41% of respondents personally used public tools like ChatGPT, while only 22% said their organizations were actively using generative AI.

In other words, your customers and their employees aren’t waiting for your permission or your official product roadmap to start using AI tools within your domain.

The Risk of “Staying In Your Lane”

Another of our clients built an AI agent to teach people how to perform a particular kind of facility inspection (let’s say it was for fire suppression systems a.k.a. “fire sprinklers”: it wasn’t, but it was just as regulated, procedural, and tedious enough to tempt people into cutting corners.)

The first version of the agent was good. It walked new technicians through the inspection sequence, quizzed them on code requirements, and role-played the building manager who wants a passing grade on a system that doesn’t deserve one.

But the client was locked into a mindset in which they viewed the AI agent as merely a “learning tool”. 

We pointed out that the same agent could ride along on the job: pulling up the relevant codes for an oddball installation, prompting technicians through the checklist, helping them document deficiencies in the required format. 

The client’s answer: “We’re a training company. Let’s stay in our lane.”

Traditionally, that boundary made sense: training companies train while consultants consult, and software providers provide software, all to help the end user execute. 

But an AI system that already possesses the contextual information and encoded judgment to teach a task is already well on its way to accelerating or even automating the task. Giving a customer a well-designed and well-grounded AI system for training then taking it away isn’t a strategy: it’s a self-imposed speed limit. An AI system that “understands” the task but doesn’t see it through to completion isn’t disciplined, it’s just an unfinished product.

Don’t Clear the Track for Your Competition

Yet another client had us connect an AI agent to a particular type of customer data and used it to obliterate a workflow bottleneck that had plagued their customers for years. Let’s say they served corporate shipping departments (they don’t, but their data is just as ugly.) 

Phase 1 of their customers’ process was reconciling thousands of carrier invoices against contracted rates, shipment records, and surprise surcharges. Miserable, error-prone work that ate whole teams alive. The agent cut it to a fraction of the time.

Customers loved it, and told our client exactly why: “Now we can spend more time on Phase 2!” Phase 2 being: use what the audit revealed to renegotiate carrier contracts and rethink routing.

We pointed out that the agent could help with Phase 2, too. It already had the data, the integration, and the customers’ trust. 

Once again the client responded with caution: “Phase 1 is enough for now. Let’s put it out there for a year and see how it does.”

This made us nervous on the client’s behalf, especially since they had a semicompetitor whose software covered Phase 2, and who were aggressively incorporating AI into their software platform, not as effectively as our clients just did, but far more extensively. And now our clients had essentially built a very expensive demo for their competitors, providing that AI could handle a painful chunk of an adjacent workflow.

The whole situation triggered my own business PTSD: twenty years ago, a major record label I worked with decided to launch a highly innovative digital music platform in just one market (sort of a “Spotify before Spotify), but they limited the release to just one geographic market and were slow to expand, allowing iTunes and others to copy all their best features. 

Hence, where our client saw the success of their Phase 1 solution as the finish line, we pointed out how they were about to draw a new set of battle lines with their rival.

A Head Start Is Only Valuable If You Keep Moving

Let’s make something clear: in all of the cases mentioned above, our clients were actually among the first movers within their respective fields when it came to developing and marketing AI solutions.  But that fact actually made their decision to slow or pause after an initial MVP more risky, not less.  

Early movers don’t just get a head start in product development: they also get a head start in learning, integration, customer access, and the first-mover advantage of defining what customers expect from an AI solution within their product category.  And those advantages compound with each sale and implementation.

Before a single user touches your product, you’ll also run a gauntlet of procurement, legal, IT, cybersecurity, data governance, and – depending on the industry – regulatory compliance officers. Andreessen Horowitz’s 2025 survey of 100 enterprise CIOs found that AI buying now looks like traditional software procurement, but with more rigorous evaluations.

Having an approved AI system deployed in production within a client’s walls is a rare and valuable asset, yet I routinely watch companies spend months clearing that hurdle then slow down, as if the Allied armies at Normandy decided to stop and take a beach vacation instead of pressing on. 

A more strategic question is: “Now that we have an approved relationship, trusted infrastructure, access to the right systems, and a working AI capability, what’s the maximum additional value we can responsibly deliver from here?” Then expand like your business depends on it, because it does. Secure your position and keep advancing, until you own every use case your company can reasonably serve before a competitor slips through the gaps.

Narrow Lanes, Crowded Track: Surviving Consolidation and Competition

Every major technology shift in recent years has brought a wave of consolidation. And while AI might buck that trend by making specialized software cheap to build, so far procurement gauntlets, compliance barriers, integration costs, and rising switching costs are shrinking the number of actual enterprise relationships. Buyers would much rather have a few approved AI systems (or one big monster AI system) handle many connected workflows than forty point solutions that each needed a security review.

For an incumbent sitting on a narrow AI product, that’s a nasty combination of Porter’s forces: hordes of “new entrants” coming after your market, while buyers flex their negotiating power to cut vendors.

The only defensive logic for waiting is, “Our competitors are just as hesitant, or even less imaginative, than we are.” But once you’ve imagined the next move in your product’s natural expansion, you can assume somebody else has. 

We actually demoed an AI simulation to train newly promoted military commanders (in Ukraine) for a European government buyer, and that required us to not only study reliably successful battlefield tactics but also reliably ineffective tactics. One of the mistakes a commander can make is “attempting to hold an indefensible position” (as opposed to keeping their squad moving, searching for a better position). Applied to AI product development, standing still becomes a bet against your current competitors, startups, adjacent software vendors, the frontier labs, customers building their own tools, and rivals you haven’t even spotted yet.

From Minimum Viable to Maximum Responsible

While I absolutely agree that companies looking to market an AI product should start with a “minimum viable” product, that should not be the end. So let me propose the Maximum Responsible Product (MRP) as the immediate next step in a defensible AI strategy.

An MRP isn’t necessarily a bigger product. It’s a more complete solution to the customer’s problem. In the root cause analysis example, one fairly simple drafting feature did more to finish the product than a dozen extra advisory features ever could. And yes, sometimes finishing the job means cannibalizing a profitable piece of your own business. But if an AI offer can substitute for your premium service in your clients’ minds, that weakness already exists. Better you own the replacement than someone else.

The word responsible carries a lot of weight in MRP, and it isn’t necessarily about slowing down so much as developing the competence to expand safely at speed. When Bill Elliott said his response to danger was “Keep my foot on the gas [and] try to drive around it” that might sound like bravado, but in reality it was the reasonable answer of a professional who’d spent a massive amount of time learning exactly what his car could do, what the track would allow, and how to react when something went wrong at 185 mph. 

A “maximum responsible” approach is as much about competence as guardrails. Start by asking where do you have, or can you quickly acquire, enough expertise to build an AI system that does the job well? And where do AI’s capabilities overlap with that expertise in a way that genuinely benefits customers?

Picture three circles: what your firm knows deeply, what AI can reliably do today, and what your customers need done.

• Where all three overlap, build now, and build all the way to the edge of that overlap. That’s your MRP.
• Where AI and customer needs overlap but your expertise runs thin (you’ve mastered Phase 1, you’re shaky on Phase 2), don’t build something mediocre, and don’t stop. Close the gap fast: hire, partner, license, or acquire, the way a race team buys a better engine rather than forging one. Leave the gap open and somebody else will happily drive through it.
• Where your expertise runs deep but AI can’t do the core work yet (tacit physical judgment, a professional’s signature, a call where being wrong ends something), keep the human at the center and build everything around them: intelligence gathering beforehand, decision support afterward, documentation in between.

Once you’ve drawn that map, responsibility becomes a matter of building well. The genuinely irresponsible AI products I’ve seen weren’t necessarily overly ambitious in what they tried to do: rather, more often they were either fundamentally ill conceived (i.e. in the wrong circle above) or just executed sloppily, skipping necessary human checks or trying to assign highly consequential work to cheap models to save a few dimes per message on their cloud bill (while risking far greater losses and lawsuits due to poor quality output.)

You Don’t Have to Build the Engine to Win the Race

A skeptic might ask: “Why bother trying to play in the AI space at all, building solutions for a given niche when OpenAI, Anthropic, Google, or Microsoft might hand everyone for free in six months?”

It’s a fair question, and it’s true that nobody else is going to build a better model than the frontier AI labs (Bloomberg learned this the expensive way when the model it spent millions training on financial data got outrun on financial tasks by OpenAI’s next general-purpose release).  However, the model isn’t the entire system any more so than an engine is an entire car: the value add for domain expert firms comes from everything you build around the model to enable it to do professional-caliber work, reliably at scale. 

The drafting feature in my first story wasn’t valuable because it could write; ChatGPT can write. It was valuable because it wrote by the firm’s own frameworks, inside a workflow that had already done the hard thinking. 

Put another way, I had multiple plumbers visit my house this week to fix various leaky pipes and blocked drains, and I look forward to the day those plumbers show up with a robotic assistant that does 70% of the work in 30% of the time allowing them to charge 50% of the price. 

But while I wouldn’t be surprised to see a Deepseek, Google Gemini, and/or Zhiyuan Robotics logo on the side of that robot, I’d be shocked if it was OpenAI’s name emblazoned on the robot’s torso. Instead, I’d expect to see either some established name in plumbing equipment (like Moen or Rigid) or some plumbing-specific tech company nobody’s heard of yet that knows robotics, AI, and plumbing, but builds the thing mostly from other people’s parts. 

That’s the market position I’d recommend any domain-expert firm to aim for: not a consulting firm with AI features or a software vendor with a chatbot, but a domain-native AI company, the brand in its niche that delivers outcomes using whatever combination of humans, AI, software, and hardware works best, with competitive edge coming from domain knowledge, systems engineering, operating discipline, and risk competence.

Conclusion: When You’re Out Front, Don’t Hit the Brakes

That’s the choice facing every expert firm and software company whose work AI could do. The danger usually isn’t missing the next move: your customers already told you about Phase 2, and your users are already opening ChatGPT to finish the job your product started. The danger is seeing it, then pumping the brakes for a year to see how things go.

So flip the burden of proof. Stop asking “Why should we build this?” and start asking “We can deliver this value. What’s our defensible reason for withholding it?”

Build the competence to understand the risks. Figure out the most value you can responsibly deliver, build it with the gates, checks, and knowledge that make it trustworthy, and put it in front of customers before they think to ask. If you’ve already earned a spot inside their organization, use every inch of it. Because if you manage to get out front, that’s not the time to slow down: it’s time to floor it, and see if you can lap the competition.