AI-Powered Custom Applications: Where It Actually Helps (And Where It Doesn't)

In 2026, almost every custom application brief mentions “AI.” Some of them should. Most shouldn’t. Here’s how we help clients separate AI-enabled features that create real value from ones that add complexity, risk, and ongoing cost for marginal benefit.

What “AI” actually means in custom applications

When clients say “AI,” they usually mean one of four things:

LLM-powered generation. GPT-4o, Claude, Gemini producing text, summaries, drafts.
LLM-powered extraction. Pulling structured data from unstructured input (PDFs, emails, voice notes).
LLM-powered search/Q&A. Retrieval-augmented generation on your own data (“ask our documents”).
Classical ML. Recommendation engines, forecasting, classification. The “AI” of the 2010s, still useful.

These four are very different in cost, complexity, and reliability. Lumping them together causes most of the bad decision-making we see.

Where AI genuinely earns its place

1. Document and email processing

A finance company processes 200 supplier invoices a month, each with slightly different layouts. A junior team member spent 8 hours/week typing them into the accounting system.

We added an LLM extraction step: invoices uploaded → LLM extracts vendor, line items, amounts, dates → human reviews and confirms → posts to Xero. 8 hours/week became 90 minutes/week.

The LLM doesn’t replace the human (the review step matters - these are accounting records). It replaces the typing.

This pattern - LLM-assisted extraction with human review - is the workhorse of useful AI in business applications. It’s cheap, reliable enough, and the human-in-the-loop catches the cases where the model gets it wrong.

2. Search across internal documents

A 50-person law firm has 12 years of precedent letters, briefs and templates. Searching for “have we ever drafted X?” was unreliable. We built a retrieval-augmented Q&A tool: documents indexed, lawyers ask natural language questions, the system surfaces relevant passages with citations.

Use was instant. Lawyers found documents they didn’t remember existed. The cost was modest ($50/month in embeddings + LLM calls). The value was hours of saved time and better legal research.

This pattern - RAG over a corpus you own - is excellent when there’s enough internal content that traditional search fails.

3. Summarisation as a workflow step

Customer support tickets routinely run 30+ messages by resolution. When the ticket gets escalated, the new owner has to read the full thread. A simple summarisation step (“write a 3-bullet summary of this conversation so far”) halves the time-to-engaged for every handover.

The LLM is good enough at this, the cost is trivial, and the worst case (a summary that misses something) just means the agent reads the full thread anyway.

4. First-draft generation, never final

LLMs are excellent at producing 70% of a customer email, a proposal, a job description, a policy document. They’re terrible at producing the final 30% - the bits that require taste, brand voice, and specific factual accuracy.

Build that into your workflows: AI drafts, human refines. The productivity lift is real. The brand integrity stays with the human.

5. Classical ML where it always was useful

Inventory forecasting based on historical sales
Lead scoring based on conversion patterns
Recommendation engines for content or products
Anomaly detection for fraud and operational issues

These aren’t new. They were ML 10 years ago. They still work. They don’t require LLMs.

Where AI doesn’t help (or actively hurts)

“Let users ask the chatbot anything about our business”

This is the most-requested AI feature and almost always the wrong one. Problems:

Users don’t ask the questions you expect them to ask.
The chatbot answers confidently even when it’s wrong.
Confident wrong answers create real liability (incorrect pricing, wrong policy interpretations, made-up product features).
Most users would have preferred a search box and a clear navigation.

If you’re considering “chat with our knowledge base” for customer support, first ask: do we have a strong navigation and search experience? Usually not. Fix that first. The chatbot is the patch over a documentation problem.

Replacing core business logic with prompts

“The pricing engine will be an LLM prompt.” No. Pricing should be deterministic. Customers and accountants need it to be reproducible. An LLM that returns slightly different numbers for the same inputs is unfit for purpose.

Use LLMs where probabilistic-but-mostly-right output is acceptable, not where you need exactness.

Generating user-visible content unsupervised

A marketing site that auto-generates blog posts. A product page that auto-writes descriptions. A customer email that’s never reviewed.

These produce mediocre output, dilute the brand, and occasionally produce embarrassing or factually wrong content. Every saved hour costs you 10 hours of damage control when something slips through.

Replacing entire workflows with “AI does it”

Companies sometimes ask us to “use AI to do” the entire customer onboarding, or the entire claims process. They picture a magic box that takes input on one side and emits perfect output on the other.

That’s not how production AI works. Even the best LLM-powered workflows have:

Specific failure modes that need explicit handling
Confidence thresholds where the system defers to a human
Audit trails because compliance still cares
Fallback paths when the model gives nonsense

The “AI replaces the workflow” pitch is selling fiction. The right framing: “AI accelerates the existing workflow.”

The cost reality of AI in 2026

LLM costs have dropped dramatically - but they’re not zero, and they compound in ways most clients don’t expect.

For reference, in 2026:

GPT-4o-mini: ~$0.15 per million input tokens, $0.60 per million output tokens.
Claude Haiku: similar.
GPT-4o / Claude Sonnet (smarter): ~10x more expensive.
GPT-o3 / Claude Opus (smartest): 30–50x more expensive than the small models.

For a typical extraction or summarisation task with the mini-class models, costs are usually $20–200/month for a small business. For RAG search or chatbot workloads with thousands of queries, it’s $200–2000/month. For heavy generation workloads, it can run into low thousands.

The hidden costs are bigger than the raw API:

Engineering time for prompt engineering, evaluation, error handling.
Storage and compute for embeddings (vector databases).
Monitoring and quality assurance - you can’t just “ship and forget.”
Iteration time when models change behaviour with new releases.
Compliance review because AI features carry specific risk and regulatory attention.

Budget realistically. AI features add 15–30% to development cost and ongoing operations.

The data and privacy question

Every Sydney business adding AI needs a clear answer to:

What data is the LLM seeing? Customer data? Internal documents? Personally identifiable information?
Where does it go? OpenAI’s servers, Anthropic’s, Azure’s, Google’s, your own?
What happens to it after? Is it used to train models? Logged? Deleted?
Are there contracts in place? Most providers offer enterprise tiers with no-training and data-residency guarantees.
Are users informed? Some jurisdictions require disclosure when AI is processing their data.

These aren’t theoretical concerns. We’ve seen prospects who’d uploaded thousands of customer records to a third-party AI tool with no contractual protection - a regulatory landmine waiting.

A simple framework for “should we add AI to this?”

For any proposed AI feature, ask:

What does success look like? Specific. Measurable. “Reduce time to draft a proposal by 50%” is good. “Make us more AI-forward” is not.
What does failure look like? When the model gets it wrong, what’s the impact? Wasted time? Wrong information shown to a customer? Legal exposure?
Is there a human in the loop? If no, the failure mode determines whether this is acceptable.
Is there a simpler way? Could a database query, a templated form, or a search box do most of the job?
What’s the maintenance cost? Who watches the model output for quality drift? Who tunes the prompts when behaviour changes?

If you can answer those clearly, the AI feature is probably worth building. If you can’t, the brief isn’t ready yet.

The honest bottom line

AI is genuinely valuable in custom applications when used as a productivity multiplier for human work - extraction, summarisation, draft generation, search.

It’s expensive theatre when used as a magic wand for problems that have simpler solutions.

The teams getting the most from AI right now aren’t the ones with the most features. They’re the ones who picked one or two surgical use cases, evaluated them honestly, and shipped them with appropriate human oversight.

Considering AI in your application? Get in touch - we’ll do a 30-minute call separating the hype from the high-leverage use cases for your specific business.