A Custom AI Website Agent Built Around How Your Business Works

Custom goes seven layers deep, well past the color of the chat bubble

Nearly every website chat product calls itself customizable. Usually that means a logo, a color, a greeting and a box for instructions. A custom AI website agent is designed from the inside out, and the design work falls into seven layers.

1. Knowledge: which sources the agent may use, who owns each one, and which wins when they disagree.
2. Voice: the vocabulary, tone and reading level your company already uses, written down so it can be tested.
3. Actions: the specific things the agent is allowed to do in your systems, from checking an order to requesting a consultation.
4. Interface: how replies appear on the page, including buttons, date pickers, cards and forms rather than only paragraphs.
5. Guardrails: what it refuses to say or do, expressed as rules with tests attached.
6. Handoff: how a conversation reaches a person, shaped around the staff you actually have.
7. Data: where transcripts and analytics live, who can see them, and how you would take everything with you if you ever left.

A generic plugin lets you adjust a few of these at the surface. A custom build lets you decide all seven. Whether that difference is worth paying for depends on your site, and the rest of this page is meant to help you judge.

Layer by layer: what you can adjust in a plugin and what we design

Look at which rows matter for your business. If only the first two do, a plugin is probably fine. If your answers depend on live data, or a wrong reply would create a real obligation, the lower rows start to carry the weight.

Replies that do more than print paragraphs

Text works for explanations. It is clumsy for choices, dates and details, which is where a custom interface earns its keep. The components we build most often:

• Quick-reply buttons when the next step has a short list of options, such as “New project,” “Existing order” or “Something else.”
• Date and time pickers filled from your real availability, so a visitor can’t pick a slot that doesn’t exist.
• Cards for products, services or locations, each linking to the relevant page on your site.
• Inline forms that validate an email, phone number or order number as it is typed, instead of after a failed submission.
• Photo upload for businesses where a picture beats a description, with file type and size limits.
• A confirmation panel that repeats back exactly what was submitted and what happens next.

Placement is a design decision too. A floating launcher suits general browsing. An agent embedded directly in the body of a “Request an estimate” page, open by default, often works better than a bubble in the corner. We make the interface keyboard-operable, announce new messages to screen readers, keep contrast readable, respect reduced-motion settings and keep the launcher away from buttons it could cover on a phone. The script loads after your page content, so visitors who never open the chat pay little for its presence.

Writing your company’s voice down so the agent can keep it

“Friendly and professional” describes almost every business and helps no one. We ask for specifics and turn them into a voice guide the agent follows and the tests check:

• Words you use and words you don’t. An accounting firm may say “clients” and “engagement,” never “customers” and “deal.” A home-improvement brand may prefer “estimate” to “quote.”
• Length. Most agent replies should be two to four sentences. Visitors skim.
• How to say no. The exact pattern for declining a request, which always includes what the agent can do instead.
• How to apologize. Once, specifically, without grovelling or blaming the visitor.
• Emoji, exclamation points and humor. Allowed, limited or banned, your call.

Here is what a voice guide changes. Asked “can you do it cheaper?”, an untuned model tends to produce something like “I understand budget is important! While I can’t discuss pricing, our team offers great value.” With a voice guide in place, the same agent replies: “Pricing depends on the details of the job, so our estimator sets it. If you share what you have in mind, I’ll pass it along with your budget so the estimate reflects both.” The second answer is specific, avoids empty praise, and moves the conversation somewhere useful.

A source registry: every answer traces back to something you approved

We don’t point a custom agent at your whole website and hope for the best. We build a registry, a short list in which each source has four attributes: what it covers, who in your company owns it, when it was last reviewed, and whether the agent may paraphrase it or must quote it exactly.

Three rules follow from that registry:

• Precedence. If a product page and the returns policy disagree about a return window, the policy wins, and the owner of the product page gets a note to fix it.
• Freshness. Sources past their review date are flagged to their owner. A seasonal policy, a temporary closure or a holiday shipping cutoff is given an end date when it is added, so it can’t outlive its season.
• Exact wording where it counts. Warranty terms, cancellation rules and fee schedules are quoted with a link to the source, never summarized in the model’s own words.

When a question falls outside the registry, the agent says it doesn’t have that information and offers the right next step. That answer shows up in a weekly report, and a repeated gap usually means a page your website should have had anyway.

Three businesses, three different conversation designs

The same seven layers produce very different agents depending on how a business sells. These sketches are illustrative, not client stories.

An online store selling replacement parts

The most valuable question here is “will this fit my model?” The agent answers it from product compatibility data, never from general knowledge, and says so when the data doesn’t cover a model. Order-status requests require an order number plus the email on the order. Damaged-shipment reports get a photo upload and go straight to customer service. Lead capture barely matters on a store; a back-in-stock alert is the closest thing to it, sent only with the shopper’s consent.

An accounting firm

Visitors want to know whether the firm handles their situation, such as a small business with payroll or an estate return. The agent explains which services the firm offers and books consultations, asking about entity type and any filing deadline the visitor mentions. It gives no tax advice, whatever the phrasing of the question, and the engagement terms are discussed only by staff. The consultation request lands in the firm’s scheduling tool with the intake answers attached.

A manufacturer that sells through dealers

This agent’s main job is sending people to the right place. End customers get a dealer locator by ZIP code. Dealers and contractors get spec sheets and installation documents from the approved library. Large or custom orders become a structured request for quote, routed to the regional sales rep by state, with the product line, quantity and project timing captured along the way.

Guardrails written as tests you can read

A rule nobody tests is a hope. Every guardrail in a custom agent is paired with test messages and a pass condition, and the whole set reruns whenever anything changes. A sample of how those pairs look:

The checks that enforce these rules run in code before and after the model, so they hold even when a visitor finds a clever way to phrase a request.

A handoff plan that matches the staff you actually have

The right handoff depends less on technology than on who is available to receive it. We design around one of three staffing patterns, or a mix by time of day:

• Someone watches chat during business hours. A staff member can take over the live conversation from a shared inbox, with the full transcript and the agent’s summary on screen.
• Nobody watches chat. The agent offers a callback or email reply, collects the preferred time and channel, and sends your team a summary written for the person who will call back.
• After hours. The agent keeps helping with what it can, takes details for everything else, and states plainly when the office reopens.

Across all three, the lead flow follows the same order: answer first, offer to connect when the visitor shows intent, collect only what your team needs, confirm consent in plain words, and read the details back before submitting.

Analytics and transcripts you own, with an exit you could actually use

A custom agent reports into your analytics rather than a dashboard you rent. We send a small set of events, including conversation started, action used, lead submitted and handoff requested, tagged with the page and any campaign parameters, so you can compare visitors who chatted with those who didn’t using the same tools you use for everything else.

Transcripts are stored where you choose, with personal details redacted from the logs used for review, and kept only as long as your retention setting allows. The agent’s instructions, voice guide, source registry and tests live in plain files, not inside a vendor’s editor. If you ever move the agent to another developer or another platform, everything that defines its behavior goes with you.

When we’ll tell you to buy a plugin instead

A custom build is the wrong answer for a site with a handful of stable questions, no systems the agent needs to touch, and nobody available to review conversations. It’s also premature if your website content is out of date, because a custom agent will faithfully repeat whatever you feed it. In those cases we’ll suggest a reputable plugin or a site content refresh through our web design work, and point you back here when the business outgrows it.

From spec to launch: how EVOTECH builds a custom website agent

1. Free consultation. We talk by phone or video about your site, your visitors and the conversations you wish went better.
2. Agent spec. A short document covering all seven layers, which you review and approve.
3. Fixed-scope written quote. Priced against that spec, before any work starts.
4. Staging build. The agent is built and tested on a hidden copy or unlisted page, never on your live site first.
5. Test review. You see the results of the knowledge, voice and guardrail tests and can add your own cases.
6. Soft launch. It goes live on selected pages while we watch the first conversations closely.
7. Tuning rounds. Transcript reviews drive changes to sources, voice and flows, with the tests rerun each time.

We’re a Houston-area company with a US-based team, 20-plus years of operating history and a 5.0 rating on Google, and we serve businesses nationwide without anyone needing to travel. For agents that span email or internal tools as well as the website, see custom AI agent development; for how an agent fits an existing site’s pages, see AI agent for a business website.

Related services

Frequently asked questions