AI-integration engagements, run as software. From what the client said, to a costed recommendation, to a build plan, to whether it actually worked and what it costs to run.
The real app, running in your browser on three fictional clients: one part-way through its build, one where the honest answer is “don’t build”, and one weeks after go-live. Change any fact and watch the answer move; your changes stay in your browser.
Most AI tools recommend building something, every time. Emerald splits the work so a recommendation can be checked — and so "don't build this" and "leave it human" are answers it can actually give.
Reads what people said and proposes facts. Every proposal must quote their words exactly, and none is used until a person accepts it.
Whether to build, which route, what it costs on both sides, how the options rank. Deterministic and tested: the same answer on Tuesday as on Monday.
Accept or reject what the AI proposes, settle the facts it isn't sure of, and make the call the evidence can't.
Interview comments are kept verbatim and tagged: pain points, workarounds, signals, goals.
Under each comment sit the facts drawn from it, so "$85 a no-show" is never far from the person who said it.
What to automate first, what not doing it costs each month, and what building it takes.
Below it, the questions worth asking next, ranked by the money riding on the answer.
Each option says what it automates, how, what it costs to build and run, and how sure it is.
Every cheaper route it passed over says what ruled it out. Leaving it human is always on the list.
When the client chooses, the engine turns the option into the work it takes: what to confirm first, what to build, the approval steps and alerts a mistake would call for, and who to walk through it.
Every step says why it is there. If facts change mid-build, the plan says so — and keeps building what was chosen until a person decides otherwise.
After go-live, each automation is judged against the figures recorded when it was chosen: is it running, does anyone use what it produces, and what did it save?
Built and unused gets the same answer as "don't build": switch it off.
The automations report their own runs and AI calls. Each call is priced from the model and its tokens, and results are shown net of what the AI costs.
Cost levers are rules too: cache, batch, send only what's new, switch off what nobody uses. Each is priced against the hours to set it up, and one that never pays back says "not worth it".
Rules fire on specific things said or recorded: a whiteboard of callbacks, a spreadsheet nobody else trusts, a paid tool nobody opens.
Each flag shows what was seen, what to do, and whether its rule comes from a real incident.
Each fact carries its value, how it's known and how sure anyone is. "Unknown" is an answer, and it becomes a question — never a guess.
Change a fact and the analysis re-runs as you type.
An AI model proposes facts from the comments. Code checks each one before anyone sees it: the quote must match word for word, the fact must still be open, the value must fit.
Nothing changes until a person accepts it. (Switched off in the online demo — it needs an API key, and nothing you type there leaves your browser.)
Written down before any code, and enforced by tests that were each shown failing first.