$ ls services/▋
Five ways to work together. Every one has a defined scope.
Engagements run in a ladder: an audit tells you what's worth automating, a build ships it, a launch hands it to your team, a retainer keeps it improving, and scaling extends it further. Start anywhere. Most teams start with the audit.
How it unfolds
Step by step to a working system.
Scroll to move through each stage of the ladder.
Start here
Discovery Audit
One to two weeks. X-GenAI goes through your workflows, tooling, and data, and comes back with an automation opportunity map: what's worth automating, what isn't, and what each candidate would take to build.
- ▸ Written report with scoped build candidates
- ▸ Every step marked as a decision or a writing task, so a large model only runs where it's needed
- ▸ Walkthrough call to go through the findings
- ▸ Fully credited against your build if you proceed
Then
Fixed-Scope Build
A working automation with a fixed scope, sized from the audit. Code where it matters, low-code where it's faster. GDPR-aware by default.
Small
A single focused automation: one workflow, one integration, roughly 20–40 hours of work.
e.g. Ticket triage, report generation, a Slack-to-CRM pipeline.
Medium
A multi-step system: several connected workflows, an agent with tool access, or a knowledge base with retrieval.
e.g. An onboarding agent, a content pipeline with review gates.
Large
Custom agents and RAG systems built end-to-end: evaluation, monitoring, and handoff included.
e.g. A support agent over your docs and product data, multi-agent research workflows.
Then
Launch
Once the build works, the job isn't done at a repo handoff. A short stretch of testing against real data, training so your team can run it, and documentation so nothing depends on tribal knowledge.
- ▸ Structured testing against real data and edge cases
- ▸ Training session(s) so your team can operate it confidently
- ▸ Documentation handed over, no black boxes
Keep it alive
Retainer
15–40 hours a month of maintenance, iteration, and new automation work once something is live, because workflows change and software should keep up.
- ▸ Monitoring and fixes for automations in production
- ▸ Iteration as your workflows and tools change
- ▸ Priority over new project work
Then
Scale
Once something's proven on retainer, it extends further: more workflows, more teams, without starting the ladder over.
- ▸ New workflows scoped against what's already live
- ▸ Shared infrastructure and evals reused, not rebuilt
- ▸ Same team, same context, no re-onboarding
Why it holds up
Built for what happens after the demo.
Most AI pilots don't fail on stage. They fail the first week they meet real traffic, real edge cases, and real scrutiny. Here's what gets built in from day one.
Concurrency & latency engineering
Every build accounts for real load and real failure modes from the start: queuing, retries, and graceful degradation designed in, not patched on after the first spike.
Observability & tracing
Every agent run is logged and traceable end to end. When a model updates, an API changes, or the data shifts underneath you, the drift gets caught before your customers do.
Human-in-the-loop error handling
High-stakes actions get an approval gate; everything else runs untouched. Exceptions route to a person, never into a silent failure.
Compliance-aware architecture
GDPR-aware by default, with data residency and consent handling considered at the design stage, built to keep pace as the regulatory landscape shifts, not just where it stands today.
Coverage doesn't stop at launch. Models drift, APIs change, workflows break silently. The retainer keeps everything shipped monitored, fixed, and improved, so what works in month one still works in month twelve.
See how the retainer works →In plain English
What we mean by evals.
An eval is a test for an AI system. Take a set of real inputs, write down the answer you'd expect (or a rule for judging one), and score how often the system gets it right. Run it before every change, and again on a schedule after launch, and “it seems fine” turns into a number you can watch.
Factuality
Is the answer right?
The share of eval cases where the output matches a known-good answer, or passes a checklist a person wrote. This is the number most people mean when they say accuracy.
Groundedness
Is it making things up?
For answers drawn from your own documents, every claim has to trace back to a passage in the source. Anything unsupported counts as a failure, however confident it sounds.
Safety
Does it hold up against bad input?
Prompt-injection attempts, out-of-scope requests, missing data. What counts is how many get caught or handed to a person instead of obeyed or guessed at.
Human review rate
Is the review gate earning its keep?
How much lands with a person, and how often they agree with what the system would have done. Too high and the automation isn't saving time. Too low and nothing is being caught.
Reliability
What happens when something breaks?
Runs that still fail after retries, and whether every failure shows up in the log instead of vanishing. A visible failure is fine. A silent one isn't.
Latency and cost
Is a better score worth what it costs?
p95 response time and cost per 1,000 calls, tracked next to quality, so a tweak that is 2% more accurate but three times the price gets noticed.
Drift. The same eval set re-runs on a schedule. If the score slips, say after a model update or a change in your data, you hear about it from a dashboard, not from a customer.
Targets. There's no universal passing score. 96% is fine for drafting an internal summary and nowhere near enough for anything touching money or a child's record, so each target is set with you, per workflow, before the build starts.
How it works
Watch a project get shipped.
Same five steps every time. Scroll to see how one moves into the next.
Discovery Audit
2 weeksX-GenAI maps your workflows and comes back with a written opportunity map: what's worth automating, what isn't, and what each piece would cost.
Fixed-Scope Build
per scopeA working automation with a clear scope, sized from the audit.
Launch
per scopeTesting, training, and handover, so your team can run it before X-GenAI steps back.
Retainer
ongoingA retainer keeps it monitored, fixed, and improved as your workflows change, because software should keep up.
Scale
ongoingOnce it's proven, it extends further: more workflows, more teams, without starting over.
AUDIT
map ops
BUILD
ship it
LAUNCH
hand it off
LIVE
keep it running
SCALE
grow it
Built with
Models
Orchestration & agents
Data & retrieval
Automation & integration
Evals & observability
Engineering & ops
Standards
Not just a mockup
The pipeline diagrams above are illustrative. One workflow on this site isn't: a real, runnable enquiry-to-CRM automation, tested against 5 deliberately hard cases including a prompt-injection attempt.
Out of scope
- ▸ General software development unrelated to AI or automation
- ▸ On-site work: engagements are remote
- ▸ 24/7 SLA-backed support or managed operations
- ▸ Finance and healthcare work is taken case-by-case, since regulated sectors need extra compliance scoping, priced accordingly
Terms
- ▸ Invoices in EUR by default · USD on request for non-EU clients
- ▸ Payment by SEPA transfer (EU) or Wise (elsewhere)
- ▸ Builds: 50% deposit, 50% on delivery