AI Integration Service

AI Integration

We put AI inside the products you already run — your app, your website, your admin panel, your support inbox. Not a chatbot bolted onto a homepage. Features that read your own data, answer accurately enough to be trusted, and cost what you budgeted.

Built into your existing product Accuracy measured before launch Your data stays in your accounts
What we build

AI features, built into what you already run

We do not sell you a platform to migrate onto. We add capabilities to the app, website or management software your team uses today — through the same APIs any other feature would use.

Assistant Inside Your Product

An in-app assistant that knows your catalogue, your pricing rules and who the user is — so it answers about their order, not about the world in general.

In-app chat · User context · Your business rules

Answers From Your Own Documents

Retrieval over your contracts, manuals, price lists and past tickets, with a citation on every answer so staff can check the source instead of trusting a paragraph.

RAG · Vector search · Source citations

Support Automation

Incoming messages classified, routed and answered — either drafted for an agent to approve or sent automatically for the question types you decide to trust.

Triage · Draft replies · Human approval

Document & Data Extraction

Invoices, contracts, delivery notes, ID documents and handwritten forms turned into structured fields and pushed straight into your ERP, CRM or accounting system.

OCR · Structured output · ERP/CRM sync

Content & Translation At Scale

Product descriptions, listings, SEO copy and Vietnamese–English translation generated in your tone of voice, with a review step before anything is published.

Bulk generation · VI ⇄ EN · Brand voice

Semantic Search & Recommendations

Search that understands what the customer meant rather than which keywords they typed, plus recommendations built from behaviour instead of a static rule table.

Embeddings · Intent search · Personalisation

Voice, Image & Video

Call transcripts and summaries, voice notes turned into records, product photos tagged automatically, and visual checks that flag defects before shipping.

Speech to text · Vision · Auto-tagging

AI Agents & Workflow Automation

Multi-step tasks that actually call your systems: check stock, create the booking, issue the refund, update the record — each step logged and reversible.

Tool calling · API actions · Audit log
Why AI integration

Why a business would add AI to a product that already works

The gain is rarely the novelty. It is the work sitting in a queue because a person has to read something, decide something routine, or retype what a document already says. That is a narrow target on purpose — and it is where the return is easy to measure.

  • Answers from your data, not from the internetThe model reads your own documents and records, so the reply matches your prices, your policy and this customer's actual order.
  • Work that queued overnight finishes in secondsInvoice entry, ticket triage, listing copy and translation stop being a backlog and become a step in the workflow.
  • Your staff stop retyping what a document already saysExtraction moves the fields for them, and they spend their time on the exceptions instead of the other ninety percent.
  • Available at 2am, in the customer's languageRoutine questions answered outside office hours, in Vietnamese or English, with an escalation path to a person when it matters.
  • Every answer logged and scoredYou can see what it was asked, what it replied and how often it was right — so the feature improves instead of drifting.

And when it is the wrong tool

We will say so during discovery, before you spend the budget. AI is a poor fit when:

  • The task must be exactly right every single time — totals, tax, legal deadlines. That is code and rules, not a model.
  • There is no written knowledge for it to read; it cannot summarise what nobody wrote down.
  • The data may not leave your premises and there is no budget to run a model on your own hardware.
  • The real problem is a broken process. Automating it only makes it break faster.

In those cases we build the conventional feature instead — and you keep the budget for the use case that does pay.

Technology

The models and tools we work with

We are not tied to one vendor. The model is chosen per task — a small fast one for classification, a stronger one for reasoning — and swapped when a better option appears, without rebuilding your feature.

Model providers

OpenAIAnthropic ClaudeGoogle Gemini Azure OpenAIAWS BedrockOpen models (Llama, Qwen)

Retrieval & data

pgvectorQdrantPinecone ElasticsearchLlamaIndexLangChain

Integration

REST / GraphQLWebhooksNode.js PythonLaravelQueuesMCP

Quality & operations

Evaluation setsPrompt versioningGuardrails Rate limitsCachingCost dashboards
Where it pays off

Solutions by department

AI projects succeed where the work is frequent, written down and tolerant of a review step. These six areas meet all three conditions.

Sales & e-commerce

Product questions answered on the spot, descriptions generated for thousands of items, and search that finds what the customer meant.

→ pre-sales assistant
→ bulk product copy
→ intent-based search

Customer support

Tickets classified and routed, replies drafted from your help centre, and a summary of the whole thread waiting for whoever picks it up.

→ auto triage
→ drafted replies
→ thread summaries

Back office & finance

Invoices and delivery notes read into your accounting system, contracts checked against a clause list, exceptions escalated to a human.

→ invoice extraction
→ contract review
→ exception queue

HR & internal knowledge

Policies, procedures and onboarding material answerable in one place — so the same question stops arriving in someone's inbox every week.

→ policy Q&A
→ onboarding assistant
→ CV screening support

Marketing & content

Campaign copy, social variants, SEO briefs and Vietnamese–English translation produced in your tone and reviewed before publishing.

→ campaign variants
→ SEO content briefs
→ VI ⇄ EN translation

Field & operations

Voice notes turned into job records, photos checked for defects, and daily reports written from the data your team already captures.

→ voice to record
→ photo quality checks
→ automated reports
How we deliver

Six stages, and accuracy is one of them

The stage most AI projects skip is evaluation. We put a number on how often the feature is right before it reaches a customer — and you decide whether that number is good enough to launch.

Stage 01

Use-case selection

We look for work that is frequent, written down and tolerant of review, then estimate the hours or the revenue at stake. Weak use cases get dropped here, not after the build.

1 week→ shortlist + value estimate
Stage 02

Data & access review

What the feature is allowed to read, where that data lives, how it gets there, and which records must never leave your systems — agreed in writing with your side.

1 week→ data & privacy plan
Stage 03

Prototype & evaluation

A working prototype plus a test set of real questions with known answers. You get a scored accuracy figure instead of a demo that happened to go well.

2 weeks→ prototype + score
Stage 04

Integration build

The feature built into your app, website or admin panel, with the human-in-the-loop step wherever an error would reach a customer or a ledger.

3–8 weeks→ feature in staging
Stage 05

Guardrails & cost control

Spending caps, rate limits, caching, fallbacks when a provider is down, refusal rules for out-of-scope questions, and a deliberate attempt to break it before launch.

1–2 weeks→ limits + cost model
Stage 06

Launch & improvement

Live behind a staged rollout, with dashboards for quality, latency and spend. Prompts and model choice get retuned as usage teaches us what real users ask.

ongoing→ live feature
15+Years of experience
1,000+Projects delivered
30+Passionate team members
Engagement models

Three ways to start with us

Prove one use case cheaply, ship the whole feature, or keep a team improving it every month. Pricing is quoted after discovery, so you never pay for scope we have not yet understood.

AI Feasibility Sprint

2 weeks · fixed fee

For teams who want evidence before they commit a budget to AI at all.

  • Use-case shortlist and value estimate
  • Data and privacy assessment
  • Working prototype on your real data
  • Accuracy score on a test set
  • Running-cost projection and a go / no-go
Scope a sprint

End-to-End Feature Build

2–4 months · milestone-based

From prototype to a feature your customers or staff use every day.

  • Everything in the Feasibility Sprint
  • Retrieval pipeline over your content
  • Integration into your app, web or admin
  • Human review and approval flows
  • Guardrails, caching and spend caps
  • Dashboards, documentation and handover
Request a proposal

Dedicated AI Team

Monthly · 2–5 engineers

For companies rolling AI across several departments rather than one feature.

  • Named engineers on your sprint cadence
  • New use cases delivered continuously
  • Evaluation sets maintained as you grow
  • Model and provider migrations handled
  • Scale the team up or down each month
Discuss a team
Why CTG

Software engineers who use AI, not an AI shop learning software

The hard part of an AI feature is rarely the model call. It is the data, the integration, the permissions and the failure cases — which is ordinary software engineering, and it is what CTG has been doing for years.

01

We build into the product you already have

No migration, no second system for your staff to learn. The feature appears inside the app, website or admin panel they already open every morning.

02

Accuracy is measured, not claimed

Every project ships with a test set and a score. When we change a prompt or a model, we re-run it — so improvements are proven and regressions are caught before your users find them.

03

Your data stays yours

Provider accounts and API keys in your company's name, business terms that exclude training on your data, and a written record of exactly what each feature is allowed to read.

04

Cost is designed in from the start

Cheaper models for the easy nine tenths, caching for repeated questions, and hard spending caps — so the monthly bill is a number you approved, not a surprise.

Questions

What clients ask us first

Which model do you use — ChatGPT, Claude or Gemini?

Whichever fits the task, and often more than one in the same product. Classification and extraction run well on small, cheap, fast models; reasoning over a long contract needs a stronger one. We build the integration so the model is a configuration choice rather than something welded into your code, which matters because a better or cheaper option appears every few months. If you already have a provider contract, we work within it.

Will our data be used to train someone else's model?

Not under the business terms we set up. The major providers exclude API data from training by default on their business plans, and we configure retention settings and sign the data-processing terms as part of the build. Where the data cannot leave Vietnam or your own network at all, we use a self-hosted open model instead and tell you honestly what capability you trade away for that.

How accurate is it, and what happens when it is wrong?

We measure it rather than guess. Stage three produces a test set of real questions with known answers and a score you can see. Then we design around the remaining error: a human approves before anything reaches a customer or a ledger, the answer carries a citation the reader can check, and the feature refuses instead of inventing when it does not know. Any project where a wrong answer cannot be caught is one we would rather not build.

How much does it cost to run every month?

Two parts: the build, and the usage. Usage is billed per request by the provider and depends on volume, how much text each request carries and which model handles it — so we project it during the sprint using your real numbers, then reduce it with caching and by routing easy requests to cheaper models. Hard spending caps are part of stage five, so the bill cannot run away while you sleep.

Can it run on our own servers?

Yes, with open models such as Llama or Qwen on your hardware or a private cloud. It is the right answer when data genuinely cannot leave your network. Be aware of the trade: you take on GPU cost and operations, and open models still trail the leading commercial ones on hard reasoning — though for extraction, classification and translation the gap is often small enough not to matter.

Do we have to rebuild our app to add AI?

Almost never. In most projects the AI feature is a service your existing app calls, plus a screen or a panel in the interface you already have. If your current system has no API at all we build a thin one around it — which is usually worth having anyway, and is far cheaper than a rewrite.

Get started

Tell us which task is eating your team's hours

Describe the work you would hand to AI and the system it lives in today. Within two working days you will get a written first opinion — whether it is a good fit, roughly what it would cost to run, and what we would build first. No obligation, no sales call required.

We use your details only to answer this enquiry. Prefer email? Write to [email protected] or call 0912.80.50.86.