What changes when AI can read what your business actually produces? Until 2024, most business use of generative AI dealt in one input: typed text. Yet the information a real company generates every day is rarely typed text. It is purchase orders scanned as PDFs, site photographs annotated in pencil, voicemail left by a supplier, a whiteboard from Friday's production meeting. On 13 May 2024, OpenAI announced GPT-4o, its new flagship model, and made multimodal interaction — handling text, vision and audio natively, and rolling improved tools out to free ChatGPT users — the headline of the launch (OpenAI's GPT-4o announcement). The announcement and its date are on the record. What follows is our interpretation of what this meant in practice — and one distinction about timing that businesses at the time often missed.

Announcement versus rollout

A point worth separating from the start: the May 2024 event was an announcement of capability, and the availability of each capability arrived gradually, feature by feature and tier by tier, over the weeks and months that followed. OpenAI said as much in its own materials. Businesses reading press coverage in May could reasonably conclude "we can now talk to our documents"; businesses that planned against the announcement rather than the rollout often found specific features gated, rate-limited or not yet present in their region or plan. The durable lesson for any technology adoption: plan against availability, not against the keynote.

What multimodal actually unlocked

Three practical doors opened in 2024. Documents: long, awkward files — scanned invoices, supplier terms, archived contracts — could be ingested and interrogated in natural language, turning a retrieval exercise into a conversation. Images: a photograph could become structured input — identifying damage, reading a nameplate, comparing an installed part against a specification. Voice: spoken reporting and dictation moved closer to real-time, which mattered for people whose hands were occupied and whose keyboards were not.

Illustrative example: a facilities maintenance provider with twenty field engineers. Historically, a completed job produced a handwritten checklist, some phone photographs and a verbal summary that an administrator transcribed into the job system on Monday. In 2024, the firm pilots a workflow in which the engineer photographs the completed work, dictates a thirty-second summary on site, and the system drafts the structured job report — photographs attached, observations categorised, follow-up flagged. The administrator moves from transcription to exception-handling. Crucially, the engineer still confirms the draft: a misread serial number on a nameplate is exactly the kind of small, confident error that verification exists to catch.

Why this mattered more than the demo suggested

The business significance of multimodality was not novelty; it was coverage. Text-only AI touched the parts of a business that were already digitised and well-described — typically a minority of operational information in an established SME. Document intake and image understanding reached the messy majority: the paper, the photos, the verbal reports. That is why 2024 is the point in this series where AI stops being a tool for the marketing department and starts appearing on the shop floor, in the van and on site.

It also explains why integration moved up the agenda. Once the inputs are photographs and dictated notes, the valuable place for AI to sit is inside the workflow that already receives those inputs — the job management system, the CRM, the document store — not in a separate chat window. A drafted report that a human retypes into another system saves half a step; the same draft flowing into the system of record saves the whole step.

Meanwhile, open weights closed the gap

While the multimodal headlines ran, the open-weights market caught up at remarkable speed. Meta's Llama 3 (April 2024) raised the quality bar for self-hosted models; Mistral continued publishing open models competitive with far larger systems; and in December 2024 DeepSeek's V3 showed that an open-weights model could train efficiently enough to challenge the frontier labs on benchmark and price. For business use the practical consequence arrived in 2024: for a growing set of tasks — summarisation, extraction, classification, first-draft generation — a model you run yourself became genuinely good enough. That mattered most exactly where the gating concerns of 2023 lived: regulated data, client-contract restrictions, and workloads where per-token API pricing made the business case wobble. Self-hosting was no longer a research hobby; it was a procurement option — one that needed the same verification discipline as any other, on infrastructure someone had to own.

Constraints and trade-offs

The verification burden from 2023 did not lift; it changed shape. When the input is a photograph, errors concentrate at the edges — glare on a label, a worn plate, handwriting — precisely where business consequences live. A model that reads a part number as correct when it is one digit off is not making a stylistic mistake; it is corrupting a record. Human confirmation at the point of capture remains cheap; correcting a corrupted record after a month of reporting is not.

Data handling, too, became more demanding rather than less. Voice notes and site photographs routinely contain personal data — a bystander, a vehicle registration, a client signature — and uploading them to a consumer tool raises questions that a typed paragraph did not. Organisations that handled this well in 2024 asked the same gating questions as in 2023, but asked them about every new input type: where does it go, who can access it, what do our client contracts and our regulator say, and can we delete it on request?

Our interpretation of the 2024 trend: capability expanded faster than governance in most firms, and the gap became the main risk. Closing it did not require heavy machinery — it required deciding, per workflow, which inputs were allowed, what verification applied, and where the outputs landed.

Action checklist

  • Map where operational information enters the business: paper, photos, calls, forms.
  • Choose one intake workflow — document processing or field reporting — as a pilot.
  • Plan against feature availability, not announcement dates.
  • Require confirmation at the point of capture for anything read from an image or recording.
  • Extend your data policy to cover photographs, voice notes and the personal data they contain.
  • Integrate AI output into the system of record rather than a parallel chat window.
  • Baseline the manual handling time before you start, so the improvement is provable.

This is the third article in From ChatGPT to operational AI: 2022–2026. If your operational data lives in paper, photographs and phone calls rather than tidy databases, that is precisely the territory Cooper Agile works in — through CRM and workflow automation or a technical review to FRAME where AI can safely meet your real information.