Solution
AI customer support
Most support questions are asked again and again, and most have an answer that already exists in your help pages, policies, or past replies. We build an assistant that answers those from your own content, and hands the rest to your team with the conversation attached.
- 1,200+ clients helped
- Building since 2010
- 15 services, one team
- 50+ technologies
At a glance
What we build
We set up AI customer support that answers routine questions from your own approved content using a language model, and passes anything it cannot handle to a person with the context attached.
Problems we solve
Support teams answer the same questions all day while customers wait, and urgent or sensitive cases get lost in the queue.
Who we serve
Support leaders and owners whose teams are swamped by repeat questions, and businesses that want faster first answers without hiring for every peak.
How to choose
Automate the repeated, well-documented questions. Keep people for anything sensitive, unusual, or costly if wrong, and measure accuracy before you widen the scope.
The problem
The same questions, every day
Where is my order. How do I reset my password. What is your returns policy. These questions fill a large share of most support inboxes. Each is easy for a person to answer, and together they take the hours that the team needs for the cases that need thought.
Customers feel the cost as waiting. A simple question sits behind a difficult one. After hours there is nobody. During a busy week the queue grows, and the first reply comes a day later.
An assistant fixes the routine part. It answers at once, at any hour, from content you control. It does not fix the hard part, which is why the design matters: what it may answer, how it hands over, and how you know it is right.
How we solve it
What the assistant answers and what it hands over
Should the assistant answer this? A common question with a clear answer in your own content: Let the assistant answer; The question needs an account lookup or a simple action: Let it help, with checks; The customer is upset, or the case involves money, health, or safety: Hand to a person; The answer is not in your content: Say so, and hand over.
If: A common question with a clear answer in your own content
Choose: Let the assistant answerFast, consistent, and available at any hour, with a link to the source.If: The question needs an account lookup or a simple action
Choose: Let it help, with checksIt can fetch an order status or start a return, within strict rules.If: The customer is upset, or the case involves money, health, or safety
Choose: Hand to a personA trained colleague takes over, with the history already in front of them.If: The answer is not in your content
Choose: Say so, and hand overThe assistant admits it does not know and routes the case, instead of guessing.
- Systems we build
- AI SolutionsAI AgentsKnowledge and Content Systems
- Often built with
- OpenAIAnthropic ClaudeGoogle GeminiLangChainHubSpot
What it looks like
What the assistant looks like to a customer
A customer sees a short answer, the source it came from, and a way to reach a person.
Scope
What should it handle?
The right scope starts small and widens as you see how it performs. Four starting points cover most businesses.
What is your biggest support burden?
Choose this when
Delivery times, returns, pricing rules, how-to steps, and opening hours.
- What you get
- An assistant that answers from your help pages and policies, and shows where each answer came from.
- Keep in mind
- Out-of-date content gives out-of-date answers, so someone must own it.
Choose this when
Customers ask for status, invoices, or changes to their own details.
- What you get
- An assistant connected to your systems, which looks up the answer after checking who is asking.
- Keep in mind
- Identity checks come first. It must never reveal one customer's data to another.
Choose this when
Messages arrive in one inbox and are sorted by hand.
- What you get
- Automatic tagging, priority, and routing to the right person, with a draft reply for review.
- Keep in mind
- Keep a person in the loop for urgent or sensitive messages.
Choose this when
Questions arrive at night and at weekends, and wait until morning.
- What you get
- An assistant that answers what it can and takes details for the rest, so the team starts the day informed.
- Keep in mind
- Tell customers clearly when they are talking to an assistant.
Control
How accuracy and safety are kept
| Risk | What can go wrong | How we control it |
|---|---|---|
| Wrong answers | The assistant states something that is not true. | It answers only from approved content, shows its source, and says when it does not know. |
| Exposing private data | One customer sees another's details. | Identity checks before any lookup, and access limited to the asking customer's own records. |
| Bad hand-offs | A customer is stuck in a loop with no person. | A visible route to a person at any point, and the conversation passed on. |
| Content drifting out of date | Policies change and the assistant keeps the old answer. | A named owner, a review schedule, and alerts when answers stop matching sources. |
| Hidden use of AI | Customers do not know they are talking to an assistant. | Clear labeling in the conversation. |
How we do it
From a pilot to a dependable assistant
Study the questions
We analyze a sample of real support messages to see what is asked, how often, and what answers exist.
What it produces
- Question analysis
- Content gaps
Prepare the content
We gather the approved answers, fix gaps, and set rules for what the assistant may and may not say.
What it produces
- Approved content set
- Answer rules
Build and test against real cases
We build the assistant, test it on hundreds of past questions, and review every wrong or doubtful answer.
What it produces
- Working assistant
- Test results
Pilot, measure, and widen
A pilot runs on a slice of traffic with staff watching. We measure accuracy and hand-over quality, and widen the scope only when the numbers support it.
What it produces
- Pilot report
- Scope decision
What you get
What the assistant comes with
An assistant
- Answers drawn from your approved content, with sources.
- Hand-over to your team with the conversation attached.
Connections
- Links to your help desk, order system, and customer records.
- Channels you choose: your website, messaging, or email.
Controls
- Rules for what it may say and do, tested before release.
- Review of a sample of conversations every week at first.
Measurement
- Reports on questions answered, hand-overs, and accuracy checks.
- A list of content gaps found, so your documentation improves.
Technical depth
How the assistant is built
For your developers and IT lead: our choices on answering from your content, models and orchestration, evaluation, and privacy and security, and the reasoning behind each.
The assistant uses a technique called retrieval-augmented generation, or RAG: it first finds the relevant passages in your approved content, then asks a language model to write an answer using only those passages. The answer carries its sources.
- We choose among OpenAI, Anthropic Claude, and Google Gemini on accuracy, cost, data handling terms, and where data may be processed.
- LangChain or plain code coordinates retrieval, the model call, and the hand-over.
- A fixed set of real questions with approved answers is run on every change, so quality cannot slip unnoticed.
- Conversations are sampled and reviewed, and wrong answers feed back into content and rules.
- Personal data is minimized before it is sent to a model, and retention is limited to what you need.
- Actions the assistant can take are limited to a short, audited list.
Common questions
Questions about AI in customer support
Will customers accept talking to an assistant?
Mostly, when it is quick, honest about being an assistant, and easy to leave for a person. Frustration comes from loops and wrong answers, which the design is meant to prevent.
How do we stop it making things up?
By restricting it to approved content, showing sources, testing it against real questions, and having it say it does not know when the answer is missing.
Will it replace our support team?
It takes the repetitive questions, so the team can spend its time on the cases that need judgment and care. We design it with your team, not around it.
What do we need to start?
A sample of real support messages, your existing help content, and someone who owns the answers. We handle the rest, including the connections to your systems.
Is our customers' data safe?
We design for it: identity checks, minimal data sent to models, limited retention, and provider terms reviewed. We explain the choices so you can approve them.
Have a different question?
Tell us what you are weighing up and we will answer it plainly.
Ask a questionAlready a client? Email support@dechytech.com
Tell us the questions you answer most
Describe your support volume and where the answers live. We will outline what to automate, what to keep human, and a first pilot.