Between early 2024 and mid-2026, the way businesses can use AI changed in two distinct steps. Large enterprises with technology budgets, engineering teams, and custom ERP systems could integrate the models through their APIs from the start. Lower middle market companies, defined here as privately held businesses with enterprise values below $100 million, generally could not. Most have small teams, little or no internal IT, limited capital for custom software, and operations in sectors that are not heavily digitized. For them, what a model could do mattered less than whether someone without technical training could use it.
Our thesis is that this access did not improve gradually. It arrived in two steps about 20 months apart, and each step made AI usable for a broader range of work by people without technical training.
For companies without technical staff, AI became usable in two steps about 20 months apart: first through conversation, then through delegated multi-step work.
Act 1 began with OpenAI’s release of GPT-4o in May 2024. GPT-4o was a capable, natively multimodal model, and OpenAI made it available at no cost in ChatGPT.3 Its reasoning gains over earlier models were incremental. What changed was that a person without technical training could use it productively through plain conversation.
Act 2 began with Anthropic’s research preview of Claude Cowork in January 2026 and the release of Claude Opus 4.6 on February 5, 2026.67 Cowork moved AI from the chat window to the user’s own files and folders, so an operator could assign a multi-step task and review the result instead of carrying information back and forth by hand. Agentic AI already existed in developer tools; Act 2 put it within reach of non-technical users.
The distinction matters most in the lower middle market, because these companies are the least able to build integrations themselves. When AI requires no technical skill to use, a company’s readiness to use it becomes a factor in its operating performance and, in our view, in how buyers assess it.
Who can use a tool determines adoption
Through the end of 2023, large language models needed prompt engineering, API connections, data pipelines, and middleware to work inside a business process.1 A large financial institution could fund that work as a normal capital project. For a $25 million regional HVAC company or a specialty environmental remediation firm, the need for a technical intermediary was enough to prevent adoption.
In the businesses we review, lower middle market companies often depend on owner-operators, older on-premise systems, and knowledge held by long-tenured employees rather than written down. For AI to reach them, it had to work with the tools and skills those employees already have. If a deployment required a systems architect, it was out of reach for most of these companies.
Act 1: GPT-4o and conversational use (May 2024)
Before GPT-4o, a voice conversation with ChatGPT passed through separate models: one transcribed speech to text, a language model processed it, and a third converted the response back to speech, which added delay and lost context. GPT-4o processed text, images, and audio in a single model.3 An operations manager could photograph a damaged shipping document, explain the problem aloud, and receive a written summary with suggested next steps.
The usage data that followed show who adopted the tool. A September 2025 study of ChatGPT usage, published as an NBER working paper by researchers at OpenAI and Harvard University, found that by July 2025 ChatGPT had been adopted by about 10% of the world’s adult population, and that non-work use had grown from 53% to more than 70% of messages.4 Over the same period, the share of messages seeking technical help such as coding and data analysis fell from 18% to 10%, while requests involving images and other media rose from 2% to 7%.4
Gains in raw reasoning were modest during this period, and some observers argued that adding compute was producing smaller improvements in complex reasoning.2 For operators, the change was usability: the chat interface became the main way people without technical training used AI for writing, summarizing, and interpreting information.
The conversational plateau (mid-2024 to early 2026)
GPT-4o made AI usable, but the chat window limited what it could do inside a business. Between mid-2024 and the end of 2025, the person using the tool had to connect it to the rest of the company’s systems by hand. A logistics coordinator would locate a shipping manifest, copy the text into ChatGPT, explain what to extract, copy the result, reformat it, and paste it into a spreadsheet or an older database. Each transfer risked lost context and transcription errors.
Repetition was the second limit. Much of a company’s operating work consists of known tasks repeated on a schedule, such as reconciling payroll hours against contractor invoices.5 Running those tasks through a new conversation each time meant restating the instructions and the context each time. In companies whose processes live in employees’ experience rather than in written procedures, that restatement took significant time, and context limits made long tasks hard to sustain. Individual tasks such as drafting an email became faster, while end-to-end workflows changed little.
Act 2: Agentic work on the user’s own files (early 2026)
In January 2026, Anthropic released Claude Cowork as a research preview, and on February 5, 2026, it released Claude Opus 4.6.67 Cowork lets a user assign an outcome, for example “review the competitor price sheets in this folder, remove duplicates, and build a comparison spreadsheet,” and the agent plans and carries out the steps, reading and writing files only in folders the user has connected.8
Several features of the release addressed the plateau directly. Opus 4.6 offered a context window of up to 1 million tokens, in beta on Anthropic’s developer platform; context compaction, which summarizes older context so long tasks can continue; adaptive thinking, which lets the model decide when deeper reasoning helps; and outputs of up to 128,000 tokens.6 Cowork runs its work in an isolated environment, works with Claude in Excel and PowerPoint, and offers permission settings that require a person’s approval before consequential actions.8
| Benchmark | Prior model | Claude Opus 4.6 | What it measures |
|---|---|---|---|
| GDPval-AA | Claude Opus 4.5 | 190 Elo points higher | Economically valuable knowledge-work tasks6 |
| MRCR v2 (8-needle, 1M tokens) | 18.5% (Claude Sonnet 4.5) | 76% | Retrieving specific information from very long inputs6 |
| BigLaw Bench (Harvey) | Earlier Claude models | 90.2% | Legal reasoning tasks; the highest score of any Claude model6 |
| Humanity’s Last Exam (with tools) | n/a | 53.0% | Multidisciplinary reasoning6 |
Benchmarks measure isolated tasks. The practical change was that an operator without technical training could ask for multi-step work across their own documents, such as organizing files, combining research from several sources, or producing a formatted presentation, and then review the result.7
Why the lower middle market is affected most
Large enterprises did not need Act 2 to use agentic AI. Through 2024 and 2025, many had the capital, infrastructure, and engineering staff to build their own multi-agent systems through direct API connections. Lower middle market companies generally did not.
Several features of these companies make packaged tools important. In the businesses we review, most are founder-owned and rely on a small group of key people, and IT is often one generalist or an outside managed service provider. Leverage is modest: platform debt in lower middle market buyouts averaged 3.2 times EBITDA through the third quarter of 2025, according to GF Data figures compiled by CapitalPad,10 which puts more of the return on operating improvement. For a technology to be useful in this segment, it has to work without custom development, and Cowork was among the first agentic tools packaged that way for non-technical users.
Applications in lower middle market operations
Financial diligence
A buyer reviewing a $15 million family-owned HVAC business may receive five years of financial records as scattered PDFs, inconsistent spreadsheets, and scanned receipts. Before 2026, an analyst would extract the data by hand, use a chat tool to resolve categorization questions, and rebuild the income statement line by line. With Cowork, an analyst can point the agent at the folder, ask it to deduplicate records and match vendor invoices against bank statements, and have it produce a spreadsheet model with working formulas for the analyst to check.7
Freight documents and rate reconciliation
A freight business that receives manifests and rate confirmations as email attachments in many formats has relied on clerks to enter routing codes, weights, and rates into its system. A Cowork project with standing instructions can extract those fields from incoming documents into a tracking spreadsheet and flag where a billed rate differs from the contracted rate, for a person to review before anything is paid or disputed.8
Compliance review
For a company with heavy regulatory requirements, an agent can compare a municipal code against the company’s field procedures and draft a gap analysis for a compliance manager to review. Opus 4.6’s 90.2% BigLaw Bench result indicates that the model handles legal-reasoning tasks well, and a qualified person still has to approve the conclusions.6
Counterarguments and risks
Security and prompt injection
Moving from a chat window to an agent that works in a user’s files changes the security exposure. Cowork runs its work in an isolated environment, but that isolation does not limit what the agent can read or do within the folders and connectors a user has granted it.8 If an employee directs the agent to process outside vendor emails, text hidden in an email could attempt to redirect it, an attack known as prompt injection, and Anthropic’s guidance identifies agentic use with internet access as a source of that risk.8 A company using these tools needs written rules for which folders and connectors an agent may use and which actions require a person’s approval.
The limits of autonomy
Strong benchmark results do not mean an agent can run a business function unsupervised. Anthropic’s system card for Opus 4.6 states that the model lacks the broad, coherent problem-solving needed to fully automate complete roles, such as a remote software engineer, without substantial human support.9 Errors over long tasks still occur. A company that stops reviewing back-office work run by an agent can accumulate errors that are expensive to find later.
Vendor dependence
As more of a company’s operating knowledge is written into instructions for one provider’s tools, switching becomes harder, and the company is exposed to that provider’s pricing and model changes. Anthropic, for example, charges a higher rate for Opus 4.6 prompts above 200,000 tokens: $10 per million input tokens and $37.50 per million output tokens.6 Keeping the process documentation in the company’s own files, in a form any tool can use, reduces that exposure.
Implications for operators and acquirers
Commentary on lower middle market private equity over this period describes a shift in how technology is used to create value, from cost reduction toward growth: handling more volume without a matching increase in administrative cost.12 Technology due diligence on targets now looks beyond servers, cyber insurance, and software licenses to how well a company can organize and use its own data.11
A business that has done this work should be worth more to a buyer, and the case rests on valuation arithmetic rather than on a market survey. A business is worth the present value of the free cash flow it will produce, and a proper AI implementation improves each input to that calculation. The same team handles more work, so earnings grow faster. Faster billing and collection convert more of each dollar of EBITDA into cash. Operating knowledge written into the company’s systems makes the business less dependent on a few long-tenured employees, which lowers the risk premium a buyer applies. A buyer who prices on a multiple of EBITDA reaches the same result, because the multiple rises with growth and cash conversion and falls as risk rises.
The premium depends on evidence. A buyer models what a business has demonstrated: an implementation that has run for two or three years supports a higher growth rate and a lower risk premium in a buyer’s model, and one completed in the months before a sale supports neither. The measurable items in diligence are how many recurring processes are written down, how many run with AI support and human review, and how much rework the financial records need.
For an owner, Act 2 means AI adoption no longer depends on hiring technical staff. It depends on documenting how the company’s recurring work is done and setting rules for what an agent may access, and both can start with one process this quarter. For an acquirer, the number of documented, AI-supported processes is now a diligence item with a measurable answer.
Works Cited
- ChatGPT’s Expanding Horizons and Transformative Impact Across Domains: A Critical Review of Capabilities, Challenges, and Future Directions | MDPI, Computers
- So AI Won’t Scale, Now What? | Brainsteam, November 12, 2024
- GPT-4o | Wikipedia
- How People Use ChatGPT | Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan and Kevin Wadman, NBER Working Paper 34255, September 2025
- Why Chatbots Are Not the Future of AI | Srinivas Rao, Medium, April 2026
- Claude Opus 4.6 | Anthropic, February 5, 2026
- Claude Cowork | Anthropic
- Get Started with Claude Cowork | Claude Help Center
- Claude Opus 4.6 System Card | Anthropic
- Lower Middle Market Private Equity Statistics (2026) | CapitalPad
- M&A Advisory: Tech Due Diligences | BearingPoint
- Beyond Cost Cutting: Shifting from Efficiency to Growth Through Technology in LMM PE | Itally