The pitch is good. One tool, one job, a clean demo. It reads your drawing sets, finds what your reviewers missed, connects to your files and learns your standards.
The last slide says it will transform your firm.
The tool probably works. That isn't the problem.
The problem starts when a firm buys a handful of these and calls the stack its AI strategy, instead of building the layer that spans those tools.
Every one of those tools is a bet against what Rich Sutton called the biggest lesson of 70 years of AI research. Make that bet your AI strategy and you will learn what researchers call the bitter lesson.
What is the bitter lesson?
In 2019 Sutton, a computer scientist, looked back over those 70 years and drew one conclusion. General methods that get better with more computing power, meaning search and learning, have ended up beating methods built on hand-coded human knowledge, and by a large margin.
The hand-built approach wins at first, then plateaus, then gets in the way.
Sutton called the lesson bitter because the approach that wins beats the one researchers favor, the one built on their own understanding.
He was writing about research, not about buying software, so I'm applying his argument rather than quoting his view. But it maps cleanly onto the AI tools being sold to firms right now.
Why is an AI point tool a bet against it?
Most AI point tools are a general model with a hand-built layer around it: a parser for your documents, a search index, connectors to your files, a set of checks built with experts.
That layer is the product's edge. It is also exactly the kind of hand-built knowledge Sutton says gets overtaken.
The vendors will say they ride every new model too, and they do. What improves is the model you could license directly. What they add on top is the part that shrinks.
Look at what is being sold to engineering and construction firms this year. One tool reviews drawing sets and says it runs on whichever leading models are available, plus its own code built with guidance from practitioners.
Another parses drawing sets because, it argues, general models misread them.
Both may be worth buying today. The question is what their advantage is built from, and how long it lasts. (And to be clear, the pitch at the top of this post is a composite of what many vendors promise, not the words of either one.)
Meanwhile the general models keep moving toward the territory software vendors say makes them unique.
Anthropic's documentation describes its models reading PDFs as both text and images, up to 600 pages in one request. Claude connects to SharePoint and OneDrive, and ChatGPT to Google Drive.
None of that means the point tools are finished. It just means their head start is on a clock, and each new model release shortens it.
This has happened before.
In March 2023 Bloomberg introduced BloombergGPT, a 50-billion-parameter model trained for finance. Later that year, a study by Li and colleagues, published at EMNLP, reported ChatGPT ahead of it on a financial question-answering benchmark, 59.9 to 43.4.
On May 1, 2023, Chegg said ChatGPT was hurting its new-customer growth. Then CNBC reported the next day that Chegg's stock price fell 48 percent.
What is the pitch selling?
The test comes from the Linear Digressions write-up of Sutton's essay: "Would a perfect model make this unnecessary, or make this better?"
Run the pitch through it, one line at a time, and sort each line into two columns: unnecessary, or better.
"Reads every sheet and spec." Unnecessary. Reading documents is where general models improve with every release. You are paying for a head start.
"Connects to your files." Unnecessary. The connectors now ship with the general assistant your firm may already license.
"Finds what your reviewers miss." Split it. The finding is model work, and it improves on its own.
The expert who checks the finding and stands behind it is worth more with every model. Pay for the expert, not for the finding.
"Learns your standards." This is the line to read twice. The answer depends on where the learning lives.
If your standards, decisions and project history live inside the vendor's system, a better model makes the vendor better and leaves you renting your own record. If they live in a record your firm owns, every new model gets better at your firm.
"Transforms your firm." Neither column. A tool built for one job changes one job. Change across a firm comes from what connects the jobs.
Most of what the pitch charges for sits in the first column. The second column is the part worth owning, and it is usually the part the pitch is vaguest about.
What does a point-tool strategy cost?
Two things, and neither shows up in the contract.
When you start, you pay for the first column for the length of the contract while its value runs down with every release.
After that, the record that makes AI useful at your firm ends up scattered across every vendor: why you made a call, what the last job taught, which detail failed. Each vendor holds a piece. None of them holds the firm.
It is the fragmentation personal AI tools already created inside firms, now bought on purpose.
That record is the one thing the bitter lesson doesn't erase.
General methods need something to search and learn from, and your firm's history is the one supply of that nobody else has.
It is the firm's organizational intelligence, what it keeps when a tool, a model or a person moves on, and what gets every new hire up to speed.
Tome learned this directly. In March 2025 it announced it was shutting down its AI presentation product, and its co-founder wrote that the models lacked the context needed to finish the work. What the company rebuilt around was context.
What should you decide this week?
List every AI tool your firm pays for. For each one, write the pitch's promises in two columns: what a perfect model would make unnecessary, and what it would make better.
For everything in the first column, buy short: one-year terms, and the right to take your data out in a form you can use.
Treat these tools as components you expect to replace, because the accumulated work, not the model, is where lock-in sits.
For everything in the second column, own it. How your firm works and what it has learned belong in a record your firm controls, where whatever model you use next year can reach them.
Ask every tool one more question: which models can it run on?
The best model for your drawings, specs or contracts next year may come from a different lab than the best one today, and the lead changes by task.
A tool, or a record, that only works with one provider turns that switch into a migration project. Set both up to work with many models, so the one that fits the job is the one that does it.
Point tools can earn a place in your firm. They can't be your strategy. The firms that come out ahead in 2027 will have bought less AI that expires and kept more of what compounds.
Where to go next: a Source of Truth engagement takes six to ten weeks and consolidates scattered documents, systems and tribal knowledge into one governed record your firm owns, structured so both people and AI can use it, whichever model comes next.
Sources
- Rich Sutton, The Bitter Lesson, March 13, 2019. incompleteideas.net
- Anthropic, PDF support, Claude platform documentation, read November 12, 2026. platform.claude.com
- Anthropic, Productivity platforms, Claude resources, read November 12, 2026; OpenAI, Connected apps in ChatGPT, OpenAI Help Center, read November 12, 2026. claude.com, help.openai.com
- Xianzhi Li and colleagues, Are ChatGPT and GPT-4 General-Purpose Solvers for Financial Text Analytics?, EMNLP 2023 Industry Track. The 59.9 and 43.4 scores are zero-shot accuracy on the ConvFinQA benchmark; a smaller fine-tuned model scored higher on the same benchmark, and fine-tuned models still led on extraction tasks. aclanthology.org
- CNBC, Chegg drops more than 40% after saying ChatGPT is killing its business, May 2, 2023. The remark was made on the company's May 1 earnings call. cnbc.com
- Linear Digressions, The Bitter Lesson, March 16, 2026. lineardigressions.substack.com
- Keith Peiris, co-founder of Tome, LinkedIn post, March 20, 2025. linkedin.com
