The same brief kept landing on our desk. We run GrandWorks, a design and engineering studio, and for the past two years client after client asked for a version of one thing: an AI support agent that answers from their real content, hands off to a person when it should, shows up where their customers already are, and costs the same in a busy month as in a quiet one. Hey Support is that brief, taken seriously and shipped as a product. Since this is the launch post, it covers who we are, why another AI support tool needs to exist, and what we believe.
The brief that kept coming back
Nobody asked us for an AI strategy. The requests were smaller and harder. An ecommerce founder wanted order status handled at 2am without hiring a night shift. A services firm wanted the same forty questions answered without their two-person team retyping the same forty answers. A SaaS client had good docs and had accepted that nobody reads docs; people read answers, and increasingly they expect a chat window to produce them.
Underneath the specifics sat the same four requirements, every time. Answer from what we have actually published, and nothing else. Hand off to a human without a fight. Work where our customers are, which mostly means the website today and messaging apps close behind. And send a bill we can predict before the month happens, not after it.
That reads like a minimum. It was surprisingly hard to buy. Every tool we tried did some of it well. None did all four, and the four only really work together.
We tried the platforms first
Recommending tools is a big part of studio work, so we evaluated the market properly before deciding to add to it. Trials, test content, real client questions pasted into demo widgets, notes compared across a dozen products. Three things kept breaking.
Pricing broke first. The large AI support products charge per resolution: Intercom's Fin costs $0.99 for every conversation the AI resolves, and Zendesk prices its AI the same way. That means the bill scales with your success. A launch goes well, a video takes off, and the invoice grows to meet the moment. We watched clients run the numbers on their own holiday season and quietly close the tab. I have laid out the full argument in the per-resolution pricing trap; the short version is that metering answers taxes the exact behavior you want more of.
Handoff broke second. On several platforms, reaching a human is a workaround. It lives on a higher tier, or behind an integration, or behind a bot that collects your email and promises that someone, somewhere, may write back. The handoff moment is where a customer decides whether your company is actually reachable. It kept being treated as an edge case.
Honesty broke third. We tested bots that confidently answered questions the client's site never answered. Invented return windows, links that lead nowhere, policies from thin air. A support bot speaks with your brand's authority, so a wrong answer reads as a commitment you never made. That failure mode alone disqualified more tools than any missing feature did.
To be fair, the underlying technology in most of these products is genuinely good. The models are capable, the interfaces are polished. The problems were choices: pricing choices, priority choices, and a tolerance for confident nonsense that we could not sign our clients up for.
Built for clients before it was a product
So we did what a studio does and built it. First as client work: retrieval wired over one client's site, a review step so they could see exactly what the bot knew, handoff into an inbox their team already watched. A second client wanted the same setup. Then a third. Somewhere around the third rebuild we admitted we were maintaining a product with extra steps, and we gave it a name.
Building for clients first shaped Hey Support more than any roadmap would have. Every feature had to survive a real deployment before it earned a settings page. The scan-review step exists because early test crawls kept swallowing cookie banners and menu text and serving them back as knowledge, which is funny exactly once. The improve-answer loop exists because every weak answer raised the same question from whoever spotted it: now what? "Edit your website and wait" was not an acceptable reply. What the product does today is on the features page; nearly all of it traces back to a specific person asking for a specific thing.
Five things we decided early
Some decisions predate the logo, and none of them have moved since.
We charge flat prices with conversation pools, never per resolution. When a vendor earns money on every resolved conversation, the roadmap quietly bends toward maximizing resolutions, and we did not want that incentive anywhere near our product decisions. Plans run from a free tier with 50 conversations a month up to Scale at 20,000, and the details are on the pricing page. The point is a support bill you can predict in January.
We set it up for you if you want. We are a studio; doing the work is not beneath us. Done for you means we index your content, wire your integrations, and hand over something that already works. If you would rather do it yourself, the same tools are self-serve, and most of setup is pasting a URL. Critics would call the hands-on part unscalable. They are right, and we are comfortable with that; a hundred well-set-up customers beat a thousand abandoned trials.
We chose honest over impressive. The bot answers from your indexed content, cites its sources on every answer, and says it does not know when your content does not cover the question. This demos worse than a bot that always has an answer. It operates better, and operating is the job.
Your content stays yours. We never train models on it, and you can delete it whenever you like. This should not count as a differentiator in 2026. Here we are.
And prices hold. The price you join at is written into our terms for as long as you stay. This industry has a short memory: Drift was acquired by Salesloft in 2024 and the standalone product was wound down. Tools get acquired, pricing pages get revised, and grandfathering lasts until the next rebrand. A held price, in writing, is our answer to that. The longer version of all five beliefs lives on our about page.
Who we are
GrandWorks is a design and engineering studio, fifteen years old. We have done client work for Spotify, OnePlus, OPPO, Hidesign, Borosil, and UNDP India, and we made Abstruct, a wallpaper app that has passed 1.5 million downloads. The team behind Hey Support is small and specific: I run product and engineering, Prairna runs customer and sales, Hampus runs design.
I mention the history because it answers the fair question every new tool faces: why trust this one? We are a studio that ships, not a startup that appeared last quarter. Fifteen years of maintaining software for clients teaches you what launch week cannot: the boring details decide whether software still works in month eighteen. Support software earns its keep in month eighteen, at 2am, on a page you forgot the widget was installed on.
What this blog will be
Two kinds of posts. First, teaching what we learn as we build: how retrieval actually works, how to train a chatbot on your website without indexing junk, which support metrics mislead small teams, how pricing models shape the products they fund. Second, documenting our decisions, including the wrong ones. We have already shipped things and then deleted them, and writing that down plainly is more useful to you than a highlight reel.
There is a real gap here. Most writing about AI support is either vendor marketing with a blog attached or research papers nobody running a store has time for. The middle, practitioners explaining what actually happens when this stuff meets customers, is thin. That is the middle we intend to occupy.
The masthead says notes from people building AI support. We mean it literally. Notes, from the people doing the building, published as we go.
If the brief at the top of this post sounds like the one on your desk, try it free: 50 conversations a month, no card required. And if you are just here to read, that is a fine way to use us too. More soon.
