When does an AI experiment become a business tool? Through 2022, most business use of ChatGPT was personal and informal — people testing a novelty against their own inboxes. During 2023 that began to shift, and the clearest marker of the shift was a model launch: on 14 March 2023, OpenAI announced GPT-4, describing it as its "latest milestone" and highlighting stronger reasoning, broader general knowledge and the ability to follow far more nuanced instructions than GPT-3.5 (OpenAI's GPT-4 announcement). The date and the capabilities are record. Our reading of what they meant for ordinary businesses — the interpretation that follows — is analysis, and we keep the two apart.

What improved, concretely

Three capabilities mattered most for knowledge work. First, longer, more coherent documents: GPT-4 handled substantially more context, so a full contract section or a forty-page report could be summarised or reworked in one pass rather than in fragments. Second, more reliable instruction-following: prompts with several conditions — tone, length, audience, exclusions — survived contact with the model far more often than they had in the ChatGPT of late 2022. Third, stronger performance on structured reasoning tasks, which is why coding assistance matured quickly in this period.

The practical effect was that AI stopped being only a drafting aid and started becoming a working tool across three areas. Assisted writing moved from email polishing to first-draft generation for proposals, reports and customer communications. Assisted coding matured into tools that suggested whole functions and explained unfamiliar code — useful leverage for small teams without deep specialist coverage. Assisted knowledge work — extraction, comparison, summarisation — began appearing in research, compliance and bid preparation.

Illustrative example: a 60-person engineering firm bidding for public contracts in 2023. Bid preparation previously consumed senior engineers' evenings: assembling method statements, tailoring case studies, checking submissions against compliance schedules. With GPT-4-era tooling, a coordinator drafts a compliant first pass from the firm's library of past bids, and an engineer verifies technical claims before anything is submitted. The senior hours move from drafting to checking. The bottleneck moves — it does not disappear.

The verification burden

That last point is the constraint that defined 2023, and it is worth dwelling on. GPT-4 reduced error rates; OpenAI itself reported the model was less likely to produce disallowed or factually wrong content than its predecessor. It did not eliminate them. Fluent, confident, incorrect output remained possible, and in business contexts the cost of such output lands on whoever signs the document — not on the model.

This creates a first-order operational question: who checks, and how? An AI-assisted bid paragraph that invents a client reference is worse than no bid. A code suggestion that compiles but mishandles an edge case ships the same as hand-written code. Organisations that answered this question well in 2023 designed verification into the workflow — named reviewers, checklists, source-of-truth documents the AI was confined to — rather than bolting it on after a mistake.

Data handling as a gating concern

The second constraint was data. Free consumer chat interfaces were never designed as business processing environments, and in 2023 the gap between casual use and business data became a genuine risk surface. Several large companies publicly restricted staff use of consumer chatbots in this period over concerns about confidential information entering third-party systems — a reminder that the constraint was governance, not capability.

The trade-off was real and worth stating without drama: the easiest tools to access were the least appropriate for sensitive work. Businesses that wanted the productivity without the exposure had to choose between behavioural policy alone ("don't paste client data") and properly governed routes — enterprise accounts with contractual data protections, or self-hosted and API-based arrangements with controlled data flows. The second route costs more and takes longer. It is also the only route that survives contact with a serious client contract or a data protection officer.

The other 2023 story: open weights arrive

GPT-4 was not the only 2023 milestone, and for many businesses it was not the most consequential one. In February 2023 Meta released LLaMA — a capable model whose weights circulated far beyond its intended research audience, demonstrating enormous demand for models businesses could run themselves. In July, Meta's Llama 2 went further with a licence that permitted most commercial use, and by the end of the year Mistral was publishing strong open-weights models from Europe. The significance was not that open models matched GPT-4 — they did not, yet. It was that "the model" stopped being a single vendor's product and became a market. From 2023 onwards, a business with sensitive data or serious volume had a real alternative to sending everything to one provider's API: run an open-weights model on infrastructure it controlled. That option matured quietly through the rest of this series — and it is the foundation of the private AI deployments we build today.

The line between experiment and tool

Looking back from 2026, 2023 was the year the question matured from "is this impressive?" to "is this repeatable?" Our interpretation of the trend: an AI experiment becomes a business tool at the moment you can name its inputs, its checker, its failure modes and its cost per use — not the moment the output looks good. Firms that crossed that line in 2023 built durable advantage quietly. Firms that stayed in demo mode accumulated impressive screenshots and unowned risk in equal measure.

Action checklist

  • Pick one documented, repeated knowledge-work task as your first governed use case.
  • Name a human verifier for every AI-assisted output that leaves the business.
  • Separate drafting assistance from data-sensitive work; use governed tools for the latter.
  • Write down where your AI tools send data, and check that against client contracts.
  • Keep a source-of-truth library (past bids, policies, specifications) that assistants draw from.
  • Measure time saved per task, so the case for wider rollout is evidence, not enthusiasm.
  • Review failures as process findings, not individual blame.

This is the second article in From ChatGPT to operational AI: 2022–2026. If you have a drafting, coding or knowledge workflow that deserves a governed first step, Cooper Agile's AI-assisted improvements service delivers it as a measurable FORT cycle: FRAME, ORCHESTRATE, REALISE, TRANSFER.