Every team stands on the same edge: reports rebuilt by hand each month, leads retyped between systems, AI everywhere except in production. We build the steps that get you across - lead routing, reporting that builds itself, CRM automation, and AI only where it genuinely earns its place.
We build and run AI automation for the workflows around your marketing and revenue operations: qualifying and routing inbound leads, automating the reporting nobody should assemble by hand, and keeping CRM data clean between the systems you already pay for. Working software in production, not strategy decks. This is a supporting service for the clients whose campaigns we already run, not a transformation programme. For the wider argument about why businesses are becoming systems that machines act on, see The Agent-Ready Web.
On this side
On the other side
We start with one workflow, not a transformation programme. Each step ends with something your team can stand on.
Where the time actually goes, and which workflow has the clearest payback.
A working version your team can use within weeks, not a specification.
Access control, data residency, evaluation, and the legal review regulated buyers require.
In production, documented, with someone on your side owning the controls.
Only once the first one has earned it.
Assistants that qualify inbound enquiries against your own criteria, filter the traffic that was never going to convert, and hand your team a shortlist with the context attached. Built on retrieval over your approved content, so the answers stay inside what you actually claim.
The pack someone rebuilds every month, automated end to end: reporting that builds itself, with the dashboard that replaces the spreadsheet and shows how the numbers were filtered and why you can trust them.
Leads enriched and de-duplicated, classified and routed to the right owner, synced between the systems you already pay for, and flagged before they go stale.
The same discipline extends past marketing: intake pipelines that carry an enquiry from first touch to a clean CRM record, approval chains that stop living in somebody's inbox, and handoffs between tools that currently depend on copy and paste. The scope stays narrow on purpose - marketing, sales, and the operations around them, one measured workflow at a time.
A clinical-trial logistics company needed inbound enquiries qualified before they reached the team. We built a retrieval-based assistant over their approved content: zero-retention model tier, UK hosting, and built to decline anything outside its scope. CSI case study
A travel-technology company needed reporting it could trust. We built an analytics platform with a self-serve report builder over a dataset filtered to genuine human sessions. RoutePerfect case study
Our own AI product suite, built and run inside Sorbet. First-party work, mentioned for one reason: the patterns we implement for you are ones we operate ourselves, not techniques we read about.
We do not run company-wide AI transformation programmes, write speculative strategy decks, or deliver generic staff training. We are a small senior team that ships specific systems into production. If your problem genuinely needs a large consultancy, we will tell you early rather than bill you to find out.
Including the ones about what we deliberately do not do.
AI automation is software that can read, classify, and decide - language models, retrieval, and plain rules - running the repetitive work a team currently does by hand: qualifying an enquiry, assembling a report, updating a record. The practical difference from classic automation is judgement: the system handles inputs that arrive messy and in natural language, and still routes them correctly. In production it always ships with controls - evaluation, access boundaries, and a human owner.
Working software, not strategy decks. We build and deploy AI into the workflows your team already runs: qualifying and routing inbound leads, automating repetitive marketing and reporting work, enriching and maintaining CRM data, and turning internal knowledge into something your team can query. Each engagement ends with a system in production that someone on your team owns.
Yes. For CSI, a clinical-trial logistics company serving pharma and biotech sponsors, we built and deployed a retrieval-based AI assistant that qualifies inbound business enquiries, filters misrouted traffic, and hands off qualified enquiries with structured metadata. For RoutePerfect, a travel-technology company, we built a reporting and analytics platform with a self-serve report builder over a bot-filtered dataset.
The consulting happens inside the build rather than instead of it. Every engagement starts with mapping where your team's time actually goes and which workflow has the clearest payback, and you get that judgement either way. What we do not sell is the deck without the system.
We write and operate the systems: retrieval pipelines, embeddings and vector search, evaluation, access control, data residency, and monitoring. We also run our own AI product suite, Sorbet Brains, so the patterns we implement for clients are ones we maintain in production ourselves rather than techniques we read about.
It is usually the first question in regulated industries, and it shapes the architecture. In practice that has meant zero-retention model tiers, hosting in a specific jurisdiction, keeping the data store under the client's control, and building the system so it declines to answer outside its approved scope. We design for your legal review rather than around it.
We do not sell company-wide "AI transformation" programmes, speculative strategy decks, or generic staff training. We are a small senior team that ships specific systems. If your problem needs a large consultancy, we will say so.
With one workflow, not a roadmap. We map where time actually goes, pick the process with the clearest payback, and build a working prototype your team can use. If it earns its place, we deploy it, hand over the controls, and move to the next workflow.
Bring the process your team repeats most. We will tell you honestly whether AI is the right answer for it, and what it would take to put it into production.
Discuss an AI workflow