What AI integration actually covers
"AI integration" is a broad label, and the price gap between projects that wear it is enormous. The phrase usually means one of four things, and pinning down which one you need is the first — and most expensive to get wrong — decision you make.
- Chatbots and assistants. A conversational layer on your site, app, or internal tools. The cheapest version answers a fixed script; the most useful version reasons over context and takes actions.
- RAG over your documents. Retrieval-augmented generation lets a model answer from your knowledge base — policies, manuals, tickets, contracts — with citations, instead of hallucinating from its training data.
- Workflow automation. AI that triages, classifies, summarises, drafts, or routes work inside the tools your team already uses, removing manual steps rather than adding a chat window.
- Custom LLM features. An AI capability shipped inside your product — a smart compose box, an analysis engine, a recommendation layer — built to run reliably at user scale.
Each of these is a different engineering effort with a different risk profile. Our AI integration services page breaks down what we build under each, but the cost framing below applies whoever you hire.
2026 cost breakdown by project type
The ranges below are realistic industry estimates for agency-built work — design, build, and evaluation included — not raw model spend and not a single freelancer’s day rate. Treat them as a planning anchor, not a quote: the actual figure moves with data readiness, the number of systems you integrate, and how high your accuracy bar is.
| Project type | Typical example | Estimated cost | Timeline |
|---|---|---|---|
| Scripted / FAQ chatbot | Website widget answering a fixed set of support questions | $3k – $12k | 2 – 4 weeks |
| LLM assistant (no private data) | Conversational helper using a base model + prompt engineering | $8k – $25k | 3 – 6 weeks |
| RAG over your documents | Answers grounded in your knowledge base, with citations | $20k – $60k | 6 – 12 weeks |
| Workflow / process automation | AI triage, classification, or drafting wired into your tools | $25k – $80k | 8 – 16 weeks |
| Custom LLM feature in a product | An AI capability shipped inside your app with usage at scale | $40k – $150k+ | 3 – 6 months |
Ranges are 2026 estimates for fixed-scope, agency-delivered projects and exclude ongoing model usage. Your number depends on the cost drivers below.
What drives the cost up or down
Two projects with the same one-line brief can differ 5× in price. Almost all of that gap comes from four levers.
1. Data preparation and readiness
This is the line item buyers consistently underestimate. For any project that grounds answers in your own content, getting documents clean, chunked, de-duplicated, permission-aware, and indexed is frequently 30–50% of the total effort. If your knowledge lives in a tidy, well-tagged source, you save weeks. If it's scattered across PDFs, wikis, email threads, and someone's head, that becomes the project. Cutting this corner is the number-one cause of AI that answers confidently and wrongly.
2. Model choice and accuracy bar
The right model is rarely the biggest or newest one. A cheaper, faster model often handles classification, extraction, or routing perfectly — reserving a frontier model for the genuinely hard reasoning. Matching model to task is where a good build saves you money twice: lower per-call cost forever, and faster responses. The higher your required accuracy and the lower your tolerance for a wrong answer, the more evaluation and guardrails the build needs — which moves cost.
3. Infrastructure and integrations
A chatbot that lives in a widget is simple. One that reads your CRM, writes to your ticketing system, respects per-user permissions, and runs reliably under load is a real software project. Each system you connect adds integration, auth, and error-handling work. This is why “automate our support” can be $25k or $80k depending entirely on how many tools the AI must touch.
4. Evaluation and ongoing tuning
The difference between a demo and a dependable system is an evaluation harness — a way to measure answer quality against known-good cases so you catch regressions before users do. Serious projects budget for it up front and for ongoing tuning as your data and models change. It is the least glamorous part of the build and the one that most determines whether the AI is still trusted six months later.
How to size the ROI before you build
The fastest way to waste an AI budget is to build first and look for value after. Tie the project to one measurable outcome before scoping it. Three framings cover almost every case:
- Time saved. Hours of manual work the AI removes × loaded labour cost. If a support agent saves 8 hours a week drafting replies, that is concrete, recurring value you can annualise.
- Deflection. Tickets, calls, or emails the AI resolves without a human × your cost per contact. Even a 25% deflection rate on a high-volume queue compounds fast.
- Revenue. Faster response times, higher conversion, fewer drop-offs, or new self-serve capabilities that let you sell or serve more without adding headcount.
The test is simple: if the annualised value of that outcome clears the build cost plus first-year run cost, the project pays for itself. Well-scoped automation typically targets payback inside 6–12 months. If you can't name the outcome and a rough number, that's a signal to narrow the scope before spending — not a reason to build and hope.
Build vs. API: which one you actually need
Most teams assume serious AI means training a model. In 2026, that is rarely true. The honest default is to call a hosted model API and invest your budget in the parts that are uniquely yours — your data, your integrations, your evaluation.
- Use an API (almost always). Hosted models from OpenAI, Anthropic, and Google give you state-of-the-art quality with zero training cost and a ship time of weeks. Prompt engineering plus RAG now covers what teams previously assumed required custom training.
- Consider custom or fine-tuned (rarely). It earns its keep only with a narrow, very high-volume task where per-call cost dominates, strict data-residency rules that rule out third-party APIs, or a specialised domain where general models genuinely underperform after good prompting and retrieval.
The right partner will push you toward the cheaper path when it fits. If a proposal leads with "we'll train you a custom model" before understanding your data and volume, treat it as a flag. For features that live inside your product rather than a chat window, our custom software development work wires the AI directly into the application so it ships as a real, maintained capability.
Pitfalls that quietly inflate cost
The expensive mistakes in AI projects are rarely technical exotica — they are scoping and process gaps. The recurring ones:
- No evaluation harness. Without a way to measure quality, the system drifts and nobody notices until trust is gone and a rebuild is on the table.
- No human in the loop for high-stakes outputs. Anything touching money, legal text, or customer commitments needs review, not blind automation.
- Ignoring data permissions.A RAG system that surfaces content a user shouldn't see is a security incident, not a bug. Permission-aware retrieval has to be designed in, not bolted on.
- Over-spec'd models. Paying frontier-model prices for a task a cheaper model handles burns budget on every single call, forever.
- Treating it as one-and-done. Models update, your data changes, and prompts decay. An AI feature is a system to maintain, not a deliverable to forget.
Getting an accurate number for your project
The ranges here will get you to a defensible budget for a board or a plan. To get from a range to a real figure, the questions that move the number are: where does your data live and how clean is it, how many systems must the AI integrate with, how costly is a wrong answer, and what is the one outcome you’re buying. If you can answer those, scoping is fast and the estimate is tight.
If you'd like that scoping done for you, you can compare how the AI layer is built for a standalone assistant versus a feature inside an existing product on our AI integration and custom software pages — or skip ahead and tell us about your project for a tailored estimate.