Doow is a shell until it can see your identity provider, bank, accounting system and HRIS. The order of the four steps was most of the design.

The obvious way to open a spend product is one giant number. I argued against it. A finance lead opening Doow on a Tuesday already knows roughly what their total is. What they cannot get anywhere else is what changed since last week, and what has a date on it.

So the dashboard opens on three task cards: what needs attention, what renews soon and what the system has noticed. The totals come after.

Design has Figma, product has Miro and marketing bought Canva: three bills for one capability. Doow groups tools by what they do, which surfaces the overlap. The harder question is whether to switch.

A price comparison leaves out the cost of moving, which usually decides it. So each recommendation shows year-one net savings, in red when negative, payback in months, the estimated switch cost, and a feature match that names what is missing.

Slack charges per seat. AWS charges for usage. OpenAI sells prepaid credits that burn down as tokens. Salesforce sells a negotiated contract with a floor and overage above it. Some vendors do two at once.

One screen has to hold all of it, because four separate flows would give four totals that do not reconcile, and a dashboard nobody trusts. The panel asks a few yes or no questions and opens only what the answers need, so a Slack plan takes twenty seconds and a hybrid enterprise contract still fits. A running total in the footer lets you check it against the invoice before saving.

Derek and Mina are Doow's AI finance agents. They audit filings, run analysis and produce board ready decks. Returning work someone will act on is a different design problem from returning text: it needs review, provenance and a way back.

Derek investigates, tracing a jump in spend to the vendors and contracts behind it. Mina packages, turning an analysis into something a CFO can read. Either can do either, as the replay shows, but the split gives each a default posture and a reason to ask which one you want before you start typing.

Responses stream as they are written. Forty seconds on a still screen looks broken. The same forty seconds with text arriving reads as work in progress.

A trial with a card on file can never expire: the charge either succeeds or fails into dunning. I found that by building a working model of the billing states and finding no route into expiry. It removed a branch of the product and made dunning the main failure path, so that flow had to be good.

Doow prices on tracked spend, not seats, because three people managing twelve million dollars of software get far more from it than thirty managing two hundred thousand. Nobody can pick a tier before Doow knows their spend, so plan selection moved to the end of onboarding, and tiers below your spend show as disabled with the reason inline.

AI usage above the allowance is prepaid, so the balance is the cap, with no usage invoices and no limit buried in settings. Overage is billed at cost, which loses a little after card fees. We chose that over a margin hidden in the metering.

Close to four hundred screens only happens by composing from a system. I checked all 96 colour tokens against black and white text, and the point where a label flips from black to white falls between steps 500 and 700 on every ramp. That became a rule engineers could apply without me: lower steps are fills and borders, upper steps carry text. Focus states are designed into every button variant rather than left to the browser.

Nearly the whole interface comes from the library, so a component change spreads everywhere instead of becoming a two week audit.