Custom AI development: assistants, search and integrations built on your data
Off-the-shelf assistants know nothing about your customers, your prices or your procedures. Custom AI development connects a model to your own data and systems so the answers are about your business.
Custom AI development is building AI features around a company's own data and workflows instead of using a general assistant as it comes. The common builds are an assistant that answers from your documents (retrieval-augmented generation, or RAG), AI steps inside existing workflows, and integrations that let a model read from and write to business systems. The model is rented; the value is in the data, the connections and the checks.
Anthropic
OpenAI
Microsoft 365
Google Workspace
SnowflakeSalesforce
Slack
How to read this chart
The scores are our planning judgement for a typical small or mid-size company: they show where effort or risk tends to concentrate, not measured data. Replace them with your own once you have looked at your company.
What is RAG, and when do you need it?
Retrieval-augmented generation finds the relevant passages in your own content and gives them to the model with the question, so the answer is grounded in your documents and can cite them. You need it when the answer depends on information the model was never trained on.
- It is the standard way to build an assistant on company documents.
- Quality depends more on the documents and how they are split and searched than on the model.
- Answers should show their source so a person can check them.
Should you build custom or buy?
Buy when a product already does the job on your data. Build when the value is in your own data, your own process or a connection between systems that no product covers.
- Check the AI features already included in the software you pay for.
- Build the connective pieces; rent the model.
- Avoid training your own model unless you have a reason no hosted model can meet.
How do you know it is good enough to launch?
By testing it against real questions with known good answers before anyone relies on it, and again after every change.
- Collect real questions and tasks from the team.
- Agree what a correct answer looks like for each.
- Score the system against that set.
- Fix the causes of the failures, usually in the data.
- Keep the set and re-run it after each change.
Questions buyers ask about this
Will our data be used to train a public model?
That depends on the vendor terms for the plan you use. We build on business plans whose terms exclude training on customer data, and we show you the terms.
Do you fine-tune models?
Rarely. Retrieval, good instructions and evaluation solve most business cases at lower cost and risk. We recommend fine-tuning only when those are not enough.
Can it respect who is allowed to see what?
Yes. The assistant should answer only from content the person asking is permitted to see, using the permissions of the source system.
Where does it run?
On your own accounts with the model and cloud vendors you choose, so you hold the contracts and the data.
Vendor documentation to check against
Building effective agents (Anthropic engineering)
Model Context Protocol: connecting AI applications to business systems
Agents guide (OpenAI platform documentation)
Microsoft 365 Copilot overview (Microsoft Learn)
Product details, plans and pricing change. Confirm current details in each vendor's own documentation. Consultants in a Box is a Make partner and otherwise an independent consultancy; this page is not published or endorsed by any vendor named. Names and logos are trademarks of their owners.
Go deeper on our site
Related solutions
AI agent development
An agent is only useful if it can act in your systems and only safe if it is limited, logged and owned. We build agents that do real work in the software you already have, with a person in the loop where it matters.
See the pageMicrosoft Copilot consulting
Copilot works with whatever each person can already open. If sharing in your tenant is loose, Copilot will surface it. Readiness is mostly permissions and content, and adoption is mostly training.
See the pageData warehouse consulting
Every report that disagrees with another report traces back to the same cause: the numbers were assembled by hand from different systems. A warehouse puts them in one place with one set of definitions.
See the pageAll solutions
AI, automation and integration, ERP and CRM, data and analytics, and leadership.
See the hubWant to talk this project through?
Book a call and we will look at your systems, what the work involves and what it should cost. Or start with the free business analysis.
