A German network of tax advisers, auditors and lawyers got 84% of its people using ChatGPT every week. Not signed up. Using it, six months after launch.
That number is the story. Most AI rollouts do not come close.
The usual pattern is familiar enough to be boring. A company buys licences, sends an encouraging email, runs a demo at lunchtime and waits. Three months later somebody opens the dashboard and finds a tenth of the seats have been touched in the last fortnight. The tool gets blamed. Sometimes the vendor gets blamed. Occasionally the staff get blamed.
HSP GRUPPE did something different, and OpenAI’s write-up of it contains one sentence that explains the whole outcome:
“HSP approached AI as an organizational transformation rather than a software rollout.”
That is not a marketing line. It is a description of a method, and the results follow from it.
Who they are matters
HSP GRUPPE is a network of legally independent tax advisory, auditing and law firms in Germany. The figures cover a shared ChatGPT Enterprise workspace run with Kanzleipakt, spanning 81 organisational groups.
Worth pausing on that. These are not technology companies. Tax advisers and auditors work under professional liability, client confidentiality rules and regulatory scrutiny. A partner in a German tax practice is roughly the least likely early adopter you could pick. If it works there, the “that only works for tech firms” objection loses its footing.
What they actually did
Three things, none of them about the software.
They ran it as a change programme. The emphasis went on adoption, governance and continuous learning rather than features. Monthly AI forums gave people somewhere to bring real use cases and learn from each other. That is a standard change-management structure, applied to a tool.
They standardised what worked. Successful individual experiments were turned into shared agents available to everyone. One drafts client communications with a consistent tone. Another helps classify booking questions against the German SKR03 and SKR04 accounting standards. Good practice stopped being a private discovery and became infrastructure.
They kept professional responsibility where it belongs. Final review and accountability stay with the qualified tax, legal or accounting professional in every case. That is what made the governance credible internally, and it is why a regulated firm could move quickly without being reckless.
The precondition nobody wants to hear
Here is the uncomfortable part. HSP had spent more than two decades standardising processes, embedding quality management and building a culture of continuous improvement before generative AI existed.
The rollout succeeded on top of operational foundations that were already there. OpenAI’s own summary of the lessons puts it plainly: strong processes, governance and quality systems accelerate AI adoption.
That inverts a common assumption. AI is often sold as a way to escape messy processes. This case suggests the opposite. The organisations getting the most from it are the ones that had already done the unglamorous work.
The numbers, with the caveats attached
84% weekly active usage, from 755 people using it weekly and 913 individuals in total across six months, February to July 2026. More than 500,000 messages exchanged in that period.
98.6% of surveyed employees reported higher productivity and 84.6% reported better quality of work. 95.9% reported weekly time savings, with 63.5% saving at least two hours a week.
One partner, Magdalene Posnak, cut the evaluation of multiple real estate investments from around nine hours to about two, and put the difference into advising clients.
The headline figure needs the most care. A deliberately conservative internal scenario estimates roughly 40,000 hours of additional annual capacity across the network. That is modelled capacity, not measured saving, and it is best read as an indication of scale rather than a fact about the bank balance.
Two honest caveats. The productivity figures come from a survey of employees, which measures perception rather than output. And this is OpenAI’s own case study about its own customer, so every number is self-reported and published by the vendor. Neither makes it worthless. Both mean the method is the transferable part, not the percentages.
What they did with the time
They did not cut staff. The firms in the network were already running at capacity with substantial backlogs, so recovered time went into more client work, shorter turnaround and more advisory capacity.
This is the decision most organisations skip. Time released by AI does not allocate itself. If nobody decides where it goes, it disperses, and the gain shows up in no measure anyone tracks. HSP had somewhere for it to go and said so in advance.
What to take from it
If your team has the licences and nobody is using them, the problem is not the tool.
Adoption is a management problem, and it responds to the things management problems respond to. A named owner. A forum where people share what actually works. A decision about where recovered time goes. Governance clear enough that cautious professionals feel safe. And the patience to turn individual wins into shared practice.
None of that requires a technology budget. All of it requires somebody to treat the rollout as a change programme rather than a purchase.
That is the difference between 84% and 8%.
Source: How HSP GRUPPE builds AI capabilities for tax advisory, OpenAI, 7 August 2026. All figures are self-reported by the customer and published by the vendor.