Start with the problem you need to solve. Explore team training, agents, automation, internal knowledge, custom models and alternatives to costly software.
A service work order should retain the diagnosis scope the customer approved, even if a later quote revision adds repairs and parts. Before commissioning custom software, check whether your CRM can preserve quote revisions, tie approval to a specific version, and carry that version into the work order. Then compare configuration, an extension, other quoting SaaS, and a custom app.
A service partner needs an account just to accept a job and record the repair. You pay for that access, while your coordinator still emails photos and answers the installer's calls about progress. If that sounds familiar, start by finding out what your current system cannot handle. A configuration or plan change may be enough. Custom software becomes worth considering when it can remove recurring workarounds that the existing product cannot address.
A custom supplier portal deserves consideration when access fees grow faster than the value of the features the company uses. Compare the full cost of operating and changing the system, including migration, maintenance and user support.
A scratch on a painted housing can stand out in one photo and almost disappear when the light changes. Before paying for a custom inspection model, look at what your images actually show and how many different parts they represent. Adapting a model is worth considering when the defect is visible and simpler image checks still fall short.
If a system repeatedly assigns readable freight document details to the wrong fields, fine-tuning may be worth evaluating. First identify whether the cause is image quality, text recognition, field meaning, or missing information, then test any improvement on documents reserved for evaluation.
Fine-tuning may help route support tickets when a general model repeatedly confuses company-specific categories. First establish whether employees interpret those categories consistently and whether historical records reflect the decision you want the system to make.
Low AI license usage can reflect a skills gap, but it can also indicate poor tool fit, missing data access or an unresolved workflow. Diagnose what stops employees on a real task before purchasing another course or subscription.
A local AI model may suit confidential proposal work when the company’s data requirements rule out its current tool. The decision should cover output quality, user access, capacity and responsibility for ongoing operation.
Building a case chronology means returning to the same pages: finding a date, checking what it refers to, adding it to a timeline, then finding the document again during review. Local AI, running on a computer or server at the firm, can help prepare a first draft. To be useful, each entry needs to point to the place in the file where someone can check it. Start with one matter and look at how much reading and correction the draft still requires.
A workstation suits local AI trials when one person can schedule the work and accept interruptions. A shared server is worth considering when a team needs access regardless of whether an employee's computer is running. The decision requires testing simultaneous tasks, checking answer quality, and assigning responsibility for keeping the service running.
A small local model can help classify document types before a workflow sends each one to the right review queue. Before buying one, separate text recognition, page grouping, classification, and routing. Running a model on company hardware alone does not establish accuracy, privacy across the full workflow, or performance under peak load.
Supplier statement reconciliation automation can reduce manual comparison between supplier statements and accounting records. Its useful output is a clear set of matches and reviewable differences that the accounts payable team can investigate.
OpenCode documents local llama.cpp. Claude Code can traverse a compatible gateway, but Anthropic does not support routing it to non-Claude models.
Codex can use a custom provider over the Responses wire. This guide shows the minimum profile, support boundary and the Qwen result.
Bielik led one Polish RAG study and PLLuM led another. Here are the numbers, why the order reversed, and how to build an auditable company brain.
A useful coding-agent benchmark needs work requests written like ordinary issues, clean workspaces, hidden checks, retained metrics and a separate review. This method evaluates a patch before acceptance.
A patched vLLM profile won one complete Qwen Code task. llama.cpp remained the simpler fallback. Here is what the evidence covers and what it leaves open.
Wire-level reasoning effort, no fixed max_tokens, preserved thinking and task-scoped sessions removed observed failures in a home-PC experiment.
A practical local Qwen3.8-27B startup pattern: verify the profile, inspect the API, watch GPU memory, switch to a fallback and run a restart test.
Qwen recalled facts from 230,085 tokens but needed 563.4 seconds. A public 146-question Polish RAG study shows what retrieval adds.