Somewhere in your inbox is an email promising that an AI chatbot will "10x your conversions". Somewhere in your memory is the last chatbot that trapped you in a loop of "I didn't quite catch that" until you closed the tab. Both of these are real, which is why "should we add one?" deserves a better answer than either the vendor or the memory provides.
We build these systems, so our bias runs toward yes. But we have also talked clients out of them, because a chatbot on the wrong website is a conversion leak with a monthly fee. Here is the honest picture for 2026.
What changed, and what did not
The old chatbots were decision trees wearing a chat window: press-1-for-opening-hours with typing. Modern AI chat is a language model grounded in your actual content - your pages, menus, policies, documentation - which means it can answer real questions in real sentences, cope with spelling and phrasing it has never seen, and hold context across a conversation.
What did not change: a chatbot still cannot know things you never wrote down, still cannot exercise judgement, and still cannot fix a broken enquiry process. It is a communication layer, not a business repair kit.
Where a chatbot genuinely earns its keep
The value shows up in a few specific situations. If yours is on this list, the case is usually strong:
- Out-of-hours enquiries. If enquiries arrive at 9pm and get answered at 9am, some of those customers bought elsewhere overnight. An assistant that answers instantly and captures the details - name, dates, requirements - turns dead hours into a pipeline. For hotels and restaurants this is the difference between a booking and a bounce.
- The same twenty questions. Every business has them: do you deliver, is there parking, can I bring a dog, what is included in the price. If your team answers the same questions daily by phone and email, deflecting most of that repetition pays for the system on staff time alone.
- Big catalogues and deep sites. When the answer exists but is buried - hundreds of products, extensive documentation, complex services - "just ask" beats making visitors learn your navigation.
- Lead capture with qualification. A good assistant does not just chat; it notices buying intent, asks the two questions that qualify the lead, and delivers a structured enquiry to your inbox or CRM with the conversation attached. That is a form that fills itself in.
Where it annoys people instead
- Simple brochure sites. Five pages that already answer everything do not need a concierge. A prominent phone number converts better.
- As a substitute for fixing the site. If visitors cannot find your prices, the fix is a pricing page, not a robot to apologise for the navigation. The leak-hunting in why your website isn't generating leads comes first.
- High-stakes, high-emotion conversations. Complaints, medical questions, anything where the customer is stressed: route to a human immediately. An AI handling a complaint badly is a screenshot on social media.
- When nobody reads the transcripts. The chat log is a free market-research feed of what customers actually want and where the site fails them. Ignoring it wastes half the value.
What it costs in 2026
- Off-the-shelf widgets (Intercom, Tidio, Chatbase and similar): free tiers exist; useful plans run £30 to £300 per month. Quick to launch, trained on your site with a few clicks. Limits: generic behaviour, shallow integration with your systems, per-seat or per-conversation pricing that climbs, and your customer conversations living on someone else's platform.
- Custom assistants: typically £5,000 to £25,000 to build, plus modest monthly model and hosting costs. Grounded in your content, matched to your brand voice, and - this is the real difference - connected to your systems: checking live availability, creating bookings, writing enquiries into your CRM, escalating with full context. A widget answers questions; an integrated assistant completes tasks.
The decision between them is volume and value. A modest site with occasional questions should start with a widget. A business where conversations convert to revenue - hospitality bookings, service enquiries, quote requests - usually outgrows widgets fast, and the custom build pays back the way any automation ROI calculation does: conversations handled x value per conversion.
The two traps: speed and nonsense
Two failure modes account for most chatbot regret, and both are preventable.
The page-speed trap. Many chat widgets load half a megabyte of JavaScript before your visitor sees anything, damaging exactly the Core Web Vitals that Google ranks on and slow connections feel. The fix is engineering, not abstinence - deferred loading, a lightweight launcher, loading the full widget only on interaction. We wrote the full technical approach in adding AI chatbots without hurting page speed. If a vendor cannot explain how their widget avoids this, that tells you something.
The nonsense trap. An ungrounded chatbot will eventually invent a price, promise an unavailable date, or offer a discount that does not exist - politely, confidently and in your brand's name. Deployed properly, this is a solved problem:
- Ground every answer in your actual content, retrieved at answer time, never the model's general knowledge of businesses like yours.
- Fence the scope: prices, availability and policies come only from live data or approved text, and outside its scope the assistant says so and offers a human.
- Build the handover: a clean escalation to email, phone or a person, with the conversation attached, the moment confidence drops or the customer asks.
- Log everything and review weekly, especially the conversations that ended without an answer - that list is your content roadmap.
This grounding-and-guardrails work is most of what separates a professional deployment from an embarrassing one, and it is where partner-level model experience matters - it is the core of what we build as an Anthropic Claude Partner and OpenAI Select Partner.
A decision checklist
Score yourself honestly:
- We get enquiries outside business hours that currently wait until morning
- Our team answers the same questions repeatedly by phone or email
- Our site is large or complex enough that visitors genuinely struggle to find answers
- A missed enquiry has a meaningful cost (a booking, a quote, a sale)
- Someone will own it: reading transcripts, updating content, handling escalations
Three or more ticks: a chatbot will probably pay for itself, and the last tick decides whether it keeps doing so. One or two: start with a widget trial or fix the underlying site first. Zero: spend the money on the site.
The bottom line
In 2026 an AI chatbot is neither a gimmick nor a magic conversion engine. On a site with real enquiry volume, out-of-hours traffic and answers worth finding, a well-grounded assistant captures leads you are currently losing and gives your team hours back. On the wrong site, or deployed carelessly, it is a slow-loading widget that makes up prices.
If you want a straight answer for your site - including "start with a £50 widget" when that is the truth - talk to us. We build grounded, integrated AI assistants for exactly this job, and the first conversation is a human one.