International real estate brokerage, 7 regions
Offer and story studio
The problem
Brokers put every client proposal together by hand: they searched the live inventory for units that were still available, pulled the good ones into a slide deck, and rewrote the surrounding text for the particular buyer. It ran into days of work per offer, and how the result looked depended on who was assembling it that week.
What changed
- Assembling an offer went from 70 hours of work to 20 minutes.
- 41 seats across 7 regions, 22 of them opening the tool in a given week.
- The broker still picks the units and reads the offer before it goes out.
How the work ran
The generator is available on top of the catalogue the brokers already used, and a background agent keeps unit availability synchronised against the developer feeds, so nobody leaves the interface to confirm that something is still in stock. Conversion was not measured in this rollout, so there is nothing here about deals won.
CRM relaunch + legacy displacement (complex monolith rework)
The problem
A CRM that had been growing for 15 years carried the whole sales operation, and the rules of the funnel were distributed across views, processors, creators, models and signals. Nobody could say why a given lead refused to move to the next stage without reading the code, and one funnel served residential, commercial and investment products at the same time.
What changed
- Funnel rules collected into one declarative layer, so a refused transition returns a stated reason.
- Deal, Person and Company added above the lead, with the legacy lead system left read-only and intact.
- A head of sales configures a new funnel and its data requirements without a release.
How the work ran
Delivered in two-week increments against the client architecture review. Around 95% of the change is localised in the new layer, which is what made it safe to run inside a monolith that a sales department uses all day.
Read the full case →
Krenels, internal
Background agent factory
The problem
Delivery speed was capped by how often an agent had to stop and wait for a person. We read a week of session logs and counted the stops: 16 times in three days a run was blocked on a password or an access token, 10 of 18 long background runs died when a laptop lid closed, and around 33 test runs a week were driven by hand.
What changed
- Agents run on a server around the clock and are addressed from Telegram, so a long run no longer depends on somebody’s laptop staying open.
- The repeating parts of delivery became skills: take a task, run the regression, restore context, turn a discussion into a tracked issue with a notification.
- Agents work in isolated git worktrees, so several run against one repository without competing for ports or shared config.
- Capacity moved from one project per developer to two, against a target of three to four.
How the work ran
We track this work against one measure: where the person stands in the delivery loop, with projects per developer as the proxy. Gathering and analysis go to the agent. The decision and the conversation with the client stay with us, and that line is deliberate — it is the part of the work a client is paying a person to do.
European developer, regional sales office
AI-native CRM for a regional office
The problem
The regional office worked in the head-office CRM, which had been built for a different market and a different way of selling. So the office kept its own copy of everything that mattered — project data, client matching, the numbers for the weekly report — and the two versions drifted apart between updates.
What changed
- A CRM shaped around the way this office sells, with head-office reporting fed out of it by agents.
- A project database with matching, so a rep can ask what fits a client and get the answer in the system.
- Agents keep the developer portals and the head-office CRM in step, so nothing is entered twice.
- Voice control for the stretch of the day a rep spends away from a desk.
How the work ran
Still in flight. There are no adoption or cycle-time figures worth publishing yet, and a projection in their place would be worse than an empty line.
See the products →
Start a conversation