Local AI for a small professional firm

The decision in one minute

We tested whether a small firm could run useful AI on its own premises instead of sending confidential work to an external AI provider.

The answer is yes, for the right work.

The EUR 3,680Hardware is a GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB unified memory): EUR 2,960 net, about EUR 3,680 cash outlay including Irish import VAT.
C-HARDWARE-SPEC-001
manifests/MANIFEST.md test machine was good at reading supplied documents, summarising them, extracting information, preparing first drafts, and transcribing meetings.

It was less capable at difficult, open-ended reasoning and did not remove the need for professional review.

Where it makes sense

The strongest case is a firm with confidential documents that would benefit from AI but which staff are reluctant or unable to send to an ordinary cloud AI service.

Good examples include reviewing agreements, extracting dates and obligations, summarising files, preparing routine correspondence, and transcribing meetings.

What staff should expect

Short tasks generally complete in secondsWorkhorse fresh-context generation speed is about 92 tokens per second (baseline).
C-SPEED-WORKHORSE-P1
CI 91.1-93.5 tokens_per_second
results/experiments.md.

Large documents can take minutes rather than secondsFirst-answer latency climbs with document length: about 30 seconds at 16K tokens, 95 seconds at 32K, 3.5 minutes at 48K, and 14 minutes at 80K; follow-up questions return in seconds.
C-LONGCTX-FIRST-ANSWER
results/experiments.md.

A one-hour meeting can be transcribed in roughly two minutesMeeting transcription (Whisper) runs at about 28 times real time - a one-hour meeting in roughly two minutes.
C-CAPABILITY-TRANSCRIPTION-001
results/capability-probes.md.

The output is strongestLocal output is strongest - close to send-ready, facts exact - on grounded tasks like extracting terms and summarising a document, which is precisely the confidential document work the box is for; it is weakest on open-ended drafting, where faithfulness is the limit and human review catches it.
C-QUALITY-GROUNDED-001
docs/FINDINGS.md when the answer must come from material supplied to the AI.

For open-ended drafting, expect a useful first draft rather than finished professional work (about 3 out of 5On a real-work send-readiness test (would a professional send the draft, graded 1-5 by an independent frontier model over ten solicitor/accountant tasks), the best local models rate about 3.0 - a usable first draft that needs a real edit - against the frontier's about 4.4. Getting facts right on a benchmark is not the same as being send-ready.
C-QUALITY-SENDREADY-001
docs/FINDINGS.md on our send-readiness scale).

What it does not solve

It does not match the strongest cloud systems on difficult reasoning.

It does not make professional checking unnecessary.

It does not make security automatic simply because the model is local.

And it is not primarily a way to save money on AI subscriptions.

What you are paying for

At normal small-firm volumes, cheap cloud AI costs lessThree-year total cost of ownership: about EUR 3,683 local (hardware-dominated) against about EUR 85 (EU) to EUR 907 (frontier) for cloud at roughly 5,000 document tasks a year.
C-COST-TCO-3YEAR
tools/cost_model.py.

The reason to own the machine is control of the data path.

In the setup we tested, the model processed the firm's material on the local machine instead of sending it to an external AI model provider.

The machine still needs normal IT security, access control, encryption, backup, and maintenance.

Capacity

One machine is a sensible starting point for a small team of three to five peopleOne box comfortably serves a small team of 3-5 with intermittent use, and about four concurrent users is its best point: aggregate throughput peaks there at roughly 121 tokens per second, giving each user about 39 tokens per second - faster than anyone reads. Beyond that it does not keep improving; at sixteen concurrent users aggregate FALLS to about 109 tokens per second and each user drops to about 8. A single user is unaffected.
C-CONCURRENCY-SMALL-TEAM-P1
docs/FINDINGS.md whose use is intermittent.

It is not designed to give a large number of heavy simultaneous users full performance.

Our recommendation

Do not start by buying hardware.

Choose a handful of confidential, repetitive workflows and run a short pilot.

Measure whether the time saved by the first draft or first analysis is greater than the time required to check it.

If that works on your own documents, local AI has a credible role in your firm.

The opportunity is not "AI without people". It is private AI doing more of the routine work before the person takes over.

Next step: read the full report, or book a Focus Call to pressure-test your shortlist of workflows.

Release 4.6 - meeting-intelligence workstream added; data frozen 2026-07-17.