Custom AI is software built around one company's own data, processes and approval rules, rather than a general tool everyone else also uses. A packaged tool answers anyone's question; custom AI does your company's specific job, such as reading every incoming vendor quote and preparing the comparison for approval. A prototype takes two to four weeks, a production system eight to sixteen, and the cost is driven mostly by how many systems it has to talk to and how much a person needs to check its work.
This guide covers what custom AI is, what drives the cost, a realistic timeline, and the test we use to tell clients when they should buy a tool instead.
What counts as custom AI?
Four kinds of system make up most of the work we do.
Agents carry out multi-step tasks. They read the situation, decide the next step, use tools such as email, the CRM or the ERP to act, and ask a person before anything that matters leaves. Post-sales follow-up, purchase approvals and support triage are typical. See agentic AI explained for worked examples.
Assistants over your documents answer questions from contracts, manuals, policies and tickets, showing the source for every answer. The technique is retrieval-augmented generation, or RAG: the system finds the relevant passages first and the model writes from them, which is what keeps it honest.
Document intelligence turns quotes, invoices, forms and photos into structured data, checks it against rules, and routes it. This is the unglamorous work that saves the most hours.
Private deployments run any of the above inside your own cloud tenancy or on your own servers, for data that cannot leave.
If what you need is none of these, and a tool on the market already does it, you probably do not need custom AI. More on that below.
What drives the cost?
Three things, in this order.
Integrations. Every system the AI has to read from or write to is work: authentication, data mapping, error handling, testing. A system that only reads email is cheap. One that reads email, writes to SAP, updates Salesforce and sends on WhatsApp is four times the integration effort. Count the systems before you ask for a quote.
Data. Clean data in one place is fast. Data spread across spreadsheets with the same vendor spelled five ways needs a cleaning step first. Photographs of handwritten forms need more work than typed PDFs. The assessment should say which of these you have.
Review requirements. Output that a person approves before it takes effect can go live early and improve in use. Output that must be right without review, because nobody will look at it, needs more testing, more guardrails and a longer run beside a person before launch. Decide early how much you need the system to be trusted on its own.
Model usage is rarely the big line. For most small and mid-sized deployments it is a modest monthly cost that we report alongside accuracy.
The cheapest way to learn the real cost is a two-to-four-week prototype on the real process. It tells you whether the full build is worth it before you commit to it.
A realistic timeline
| Phase | What happens | Typical duration |
|---|---|---|
| Discovery | Process, data and people assessed; one-page specification and a written quote | 1 to 2 weeks |
| Prototype | A working version on real data, reviewed by the people who will use it | 2 to 4 weeks |
| Build in sprints | Two-week sprints, scope adjusted on evidence | 4 to 8 weeks |
| Deploy and integrate | Connected to the real systems and logins, monitoring on | 1 to 2 weeks |
| Run beside a person | System proposes, a person approves, accuracy measured | 2 to 4 weeks |
Eight to sixteen weeks to production is the normal range. Single-workflow automations with one integration have gone live in three to four weeks. Anything that involves a core ERP, a regulated dataset or more than three systems tends toward the long end.
The timeline for a broader programme, where custom AI is one piece of a company-wide change, is in what is AI transformation.
Build or buy: a five-question test
Ask these in order. The first "buy" answer usually settles it.
- Does a packaged tool already do this job for companies like us? If yes, and it fits your process without you bending the process to fit it, buy.
- Is the data allowed to leave our environment? If no, you need either a tool with a private deployment option or a custom build inside your walls.
- Does it have to talk to our ERP, CRM or internal systems in a specific way? If the tool's integrations do not cover it, custom is the only route that does not create a new manual step.
- Is the process genuinely specific to us? Approval chains, pricing rules and compliance checks that exist nowhere else are the case for custom.
- Will the output need to be right without a person checking? If yes, custom gives you the guardrails and the audit trail a general tool cannot.
We sell ready-made apps as well as custom builds, so we have no reason to push the expensive option. Roughly a third of the companies that come to us for custom AI leave with a packaged app and a short training programme instead.
Questions to ask any custom AI developer
- Which models will you use, and why those? The answer should name specific models and give a reason per task.
- Where will our data be stored and processed, and will any of it be used to train a model? Get it in writing.
- What does the handover include? Documentation, a named owner on our side, training for our team, and the code or configuration.
- What happens when the model provider changes or raises prices? A good design can swap models without a rebuild.
- How will we know it is working? A number, measured against a baseline, reported monthly.
- What is the support model after go-live, and what does it cost?
If a developer cannot answer these plainly, keep looking.
A worked example
A composite of projects we have run, with the details changed. A distributor with 40 staff received around 200 supplier quotes a month as PDFs and photographs on WhatsApp. Two people spent most of their week re-typing them into a spreadsheet to compare prices, and errors in the re-typing were costing real money.
The build read each quote into structured data, matched items to the product master, flagged prices outside the expected range, and produced a comparison that the purchase manager approved with one tap. Integrations were email, WhatsApp and the existing accounting system. Discovery took one week, the prototype two, and the system was in production in nine weeks including three weeks running beside the two staff members, who then moved to supplier negotiation.
Where Quadrivium fits
We design and ship custom AI systems: agents, assistants over your documents, document intelligence and private deployments, specified with you, built in sprints and handed over with your team trained to run them. The service is on the custom AI page. If the data involved is sensitive, read safe AI for business first; if the job is a multi-step workflow, agentic AI explained shows what an agent can take on.
