AI integration for businesses: what owners and product leaders need to know before talking to a vendor

A vendor told you AI could “transform your operations.” A board member asked when you’ll have an AI strategy. A competitor announced an “AI-powered” feature and you’re not sure if it matters. You don’t need a deeper technical briefing. You need a calm framework for deciding what to do.

This is that framework. Written for owners and product leaders who are evaluating AI projects without an engineering background and don’t want to be sold to.

What AI actually does well today

Stripped of marketing, large language models — the technology behind ChatGPT, Claude, and everything else that uses the word “AI” right now — are very good at four classes of task:

Summarizing long text into short text. Meeting transcripts into action items. Support tickets into a one-line description. Contracts into key terms.

Drafting first versions a human will edit. Email replies, marketing copy, proposal sections, internal memos. Note “a human will edit.” AI as a first-drafter is reliable. AI as a final-author is not.

Extracting structured information from messy input. Pulling order details out of forwarded emails. Parsing resumes into fields. Reading invoices into your accounting system. Anything where a human is currently retyping data from one format into another.

Answering questions over a defined set of documents. Customer support chatbots over your help docs. Internal “ask the wiki” bots. Sales enablement assistants that draw from your product documentation.

These four work today. They’re well-understood, predictable in cost, and have clear failure modes. A vendor pitching one of these can give you a confident quote.

What AI is not good at

Equally important, and rarely discussed in sales calls:

Making decisions you can’t audit. AI is excellent at suggestions. It’s bad at decisions where you need to defend the outcome to a regulator, a court, or your board. The “show me the source” pattern matters.

Replacing structured systems. If a regex, a SQL query, or a spreadsheet formula can do the job, an LLM is the wrong tool. It’s slower, more expensive, and less reliable.

Open-ended creativity beyond drafting. Generating a campaign concept, a brand voice, a product strategy. AI produces serviceable mediocrity here. The signal of taste a customer responds to is still a human capability.

Anything that requires real reasoning over many sources. Multi-step research, complex analysis, audit work. Current models stumble on chains of reasoning longer than two or three steps. The marketing implies otherwise; production performance does not.

When a vendor pitches an AI project that falls into the second list, the project is going to disappoint you. Either you’ll lower the bar to declare success or you’ll cancel it.

The questions to ask before committing budget

Before you sign anything, run through this list with the project’s sponsor on your side and the vendor on the other.

1. What specific business process changes if this works? A working AI project should reduce a measurable cost, shorten a measurable timeline, or increase a measurable revenue line. If the sponsor can’t name the metric in advance, the project is exploratory, not operational, and should be budgeted accordingly.

2. What does success look like in numbers? “X% of tickets deflected.” “Y hours per week of operations time saved.” “Z% reduction in proposal turnaround.” If the answer is “users will love it,” you’re funding a vibes project.

3. How will we measure quality after launch? This needs to be set up before launch. After launch, it’s too late.

4. What’s the rough cost per use? Most LLM-powered features cost a fraction of a cent to a few cents per use. At scale, that adds up. Get a forecast for your expected volume.

5. What happens if the AI gives a bad answer? Every system has a failure mode. The right answer involves humans in the loop, escalation rules, or audit logs. The wrong answer is “it won’t happen.”

6. Whose data is going to the AI model? Where does it go? For most modern setups, the answer is one of the major API providers (OpenAI, Anthropic) with a data processing agreement. For regulated industries, the answer might need to be a self-hosted model. Get this nailed down before signing.

7. What’s the ongoing cost? Build cost is one line. Per-month operating cost (model usage, infrastructure, monitoring, maintenance) is the bigger one over time. A trustworthy vendor will quote both.

The three project sizes

In our experience, AI integrations fall into three rough categories.

Small (1-4 weeks, $3,000-$12,000). A single feature embedded in an existing product. A “summarize this thread” button. A “draft a reply” feature. An “extract data from this PDF” workflow. Small projects are excellent ways to learn what AI can and can’t do in your specific context. Most companies should start here.

Medium (2-4 months, $12,000-$45,000). A multi-feature integration. A custom AI chatbot. A document-understanding pipeline. An internal AI assistant for a specific team. This is where most production AI projects land. The cost reflects real engineering: prompts, evaluation, monitoring, integration with existing systems.

Large (6+ months, $50,000+). A platform. A new product line built around AI. A multi-tenant enterprise system. These have all the complexity of regular enterprise software plus the additional engineering load of making AI features reliable.

These ranges reflect SE Asia agency pricing for senior teams serving international clients. Western shops typically quote 2.5 to 4 times these numbers for the same scope. If someone is pitching a “small AI project” at a six-figure price, ask why. If someone is pitching an enterprise AI platform for $15,000, the budget will run out before the platform ships.

The pitfall: doing too much at once

The single most common pattern we see in failed AI projects is overscoping the first attempt. The team commits to a six-month build of an AI-everything platform before they’ve shipped a single AI feature.

The path that works: pick one specific workflow, ship an AI feature for it, measure for two months, then decide what to do next based on real data. The companies that learn AI fastest are the ones that ship narrow first. The companies that spend the most and learn the least are the ones that try to ship broad.

If a vendor or internal team is pushing you to a large initial build, push back. Ask what the smallest valuable version is and start there.

A budget reality check

For mid-market businesses adding AI features to existing software, a realistic first-year spend with a senior SE Asia agency is in the range of $15,000 to $60,000 of build cost plus $4,000 to $18,000 of operating cost. That gets you one or two production features running well, with monitoring in place, and a clear sense of what to build next. The same scope from a US shop will run two to four times higher; the LLM bills are identical regardless of where the team sits.

If you’re being quoted dramatically less, something important is being skipped. If you’re being quoted dramatically more, you’re paying for prestige overhead. Get a second opinion.

If you’re at this stage

We do scoping calls for AI integration projects, including for clients who eventually decide not to do the project. If you want an outside read on whether your AI initiative is well-scoped, sensibly priced, or solving the right problem, see our LLM integration services page or email [email protected]. One hour, no deck, honest answers.