Panasonic Connect's ConnectAI: 788,000 Hours Saved and How the Agents Are Built

Who this case is about

Panasonic Connect is the B2B solutions arm of the Panasonic Group. Its July 29, 2026 release puts the group at roughly 29,100 employees with FY2025 revenue of 1,380.3 billion yen, serving supply chain, public services and living infrastructure customers — what the company calls "the front line".

Since February 2023 it has run ConnectAI, an in-house AI assistant service built on large language models. Every figure and component described here comes from the company's own press releases. Nothing is filled in by inference.

The problem they set out to solve

ConnectAI launched with three stated goals: (1) improve business productivity, (2) raise employees' AI skills, and (3) reduce the risk of shadow AI use.

The third goal is the interesting one. Rather than banning external AI tools and hoping people comply, the company reduced the risk of employees pasting work data into consumer services by handing everyone a sanctioned environment instead. For any organisation planning a company-wide rollout, that is the first policy decision to get right.

How they used AI: three years of published numbers

What makes this case unusually useful is that the results are published every year in the same format. Putting the July 2025 and July 2026 releases side by side shows exactly how usage grew.

• ▸2024 results (announced July 7, 2025): 448,000 hours saved (2.4x year on year), 2.4 million uses (about 1.7x), 28 minutes saved per use (36 minutes when images were involved), 49.1% monthly unique user rate. Rolled out to roughly 11,600 employees in Japan.
• ▸FY2025 results (announced July 29, 2026): 788,000 hours saved (2x year on year), equal to 3.4% of total annual working hours. 3.61 million uses (1.6x), 33 minutes saved per use (1.2x), 61% monthly unique user rate (up 12 points). Rolled out to roughly 11,800 employees in Japan.
• ▸How usage shifted: in 2024 the company described the shift as moving from "asking" to "delegating". For FY2025 it attributes the growth to individual-task uses expanding — minutes from meeting files, code generation, log analysis, proposal drafting, and writing design documents and specifications.
• ▸Models behind it: the July 2025 release states that ConnectAI was developed using large language models from three major providers — OpenAI, Google LLC and Anthropic.

The monthly unique user rate climbing from 49.1% to 61% is the number to watch, because it speaks to whether a rollout actually gets used rather than quietly gathering dust. Time saved per use also rose from 28 to 33 minutes, so it is not just more people — each session is doing heavier work.

Architecture

The following is the configuration as described in the official press releases. Nothing here is inferred.

• ▸Assistant layer: ConnectAI, an in-house service built on LLMs from OpenAI, Google LLC and Anthropic, deployed to roughly 11,800 employees in Japan.
• ▸Data platform: the "Panasonic Connect Corpus", an AI/Data foundation that consolidates company data in one place. It covers structured data such as sales and production results alongside unstructured data such as drawings, specifications and meeting minutes, and connects to structured data held by 50 internal business systems. Built out from 2023 in parallel with the AI rollout.
• ▸Unstructured data access: RAG (Retrieval-Augmented Generation) has been used since 2024 to connect documents, drawings and specifications to the generative AI.
• ▸Structured data access (from FY2026): integration with business systems such as SFA and ERP, so that AI moves beyond answering questions into gathering information, analysing it and supporting decisions — what the company calls business-data-linked AI agents.
• ▸A concrete agent: the "Manufacturing AI Agent" for comparing drawings and design specifications, developed in-house on Snowflake's data cloud platform using Snowflake's Cortex AI.

What the drawing-comparison agent actually does

Announced on February 19, 2026, the Manufacturing AI Agent is where the rollout moves past "chat with a bot" into a specific production task. The processing flow is public too.

• ▸The task: checking that specifications match across product drawings, component drawings and technical specification documents — work that previously depended on manual, visual checking.
• ▸The processing: text is extracted automatically from multiple PDF drawings, then AI compares items such as material and finish between the product drawing and the component drawing or specification, presenting the results as a list.
• ▸The result: comparison work that took 50 to 340 minutes by eye now takes 10 minutes — an 80% to 97% reduction. Standardising the work also curbs quality variation between individual staff.
• ▸Rollout order: starting with standards comparison and exterior-parts comparison, then extending horizontally to other comparison tasks.

Read the wording carefully: the company calls it "semi-automatic comparison by AI", positioned as support for the person doing the checking. The headline 97% does not mean the human was removed.

Results and what to take from it

• ▸Result: 788,000 hours saved in FY2025, 3.4% of total annual working hours, against a stated target of cutting 10% of total working hours with AI by 2030.
• ▸Worth noting 1: they state where the reclaimed time goes — into customer proposals and creating new solutions. Time saved is not treated as the end of the story.
• ▸Worth noting 2: the data platform came first. Data consolidation ran alongside the AI rollout from 2023, and only once the corpus existed did the company move to the agent implementation phase in 2026. The stated order is explicit: it is not model performance but the readiness of the data the AI reads.
• ▸Caveat 1: the 788,000 hours and 3.4% come from the company's own analysis of its usage data, and the calculation method is not published. Treat them as this company's reported figures, not as a benchmark another organisation can reproduce.
• ▸Caveat 2: the 2024 figures cover a calendar year while the latest cover fiscal 2025. Keep that in mind before drawing a straight year-on-year line.

What other companies can borrow

The lesson here is sequence, not model choice. Give everyone access so the barrier to use drops; reduce shadow AI by substitution rather than prohibition; build the data foundation underneath; then move to agents for specific jobs. That the first headline win is drawing comparison — an unglamorous manufacturing chore — makes it more credible, not less.

If you are weighing this up for your own organisation, the shortcut is not to start by building an agent. It is to ask whether the data you want the AI to read is currently in a form it can read at all. That is the part Panasonic Connect spent three years on.

Sources