FIXED-SCOPE ENGAGEMENT
One workflow. Six weeks. Running in production.
The Production AI Sprint takes one workflow from an idea or a stalled prototype to a production AI system. It's integrated with your systems, tested on your real data, controlled by your people and handed over to your team.
PROBLEM
Most companies don't lack AI ideas. They lack AI in production.
- A prototype impressed people in a demo, then stalled on integration, security review or reliability.
- Nobody can say how accurate it is, because there's no evaluation set.
- Open-ended consultancy feels risky: unclear scope, unclear end date, unclear result.
- In-house teams are busy, and production GenAI (evaluation, guardrails, monitoring, integration) is a specialist skill.
CURRENT MANUAL WORKFLOW
- 1.Someone builds a prototype in a notebook or a no-code tool.
- 2.It works on hand-picked examples.
- 3.Security, IT and data protection questions arrive, and there's no architecture to answer them.
- 4.Connecting to the CRM, ATS or ERP turns out to be the real project.
- 5.Nobody defined what 'good enough' means, so nobody signs off go-live.
- 6.Months later, it's still a demo.
AI-OPERATED WORKFLOW
Agree one workflow
We fix the scope, integrations, success measures and fee in writing.
Baseline it
We measure how the workflow performs today.
Build the evaluation set
We turn 50–200 of your real examples into a test the system must pass.
Design the controls
We define approval points, review queue, escalation and audit logging before any build.
Build and integrate
We connect to your systems with least-privilege access, in your cloud or a UK/EU-region environment.
Run in shadow mode
The system works live alongside your team without acting, and we compare against the baseline.
Go live
We switch on, with monitoring for quality, cost and latency.
Hand over
You get code, documentation, runbook and a written recommendation: operate, extend or stop.
HUMAN CONTROL
Production means controlled. Every sprint ships with:
Approvals
Rules per action type, agreed with the workflow owner during design.
Review queue
Low-confidence and high-risk outputs go to people, not straight into your systems.
Audit trail
Every action is logged with its inputs, outputs, approver and timestamp.
Escalation
Named owners and SLAs for every exception type.
Evaluation gate
Changes must pass the regression set before they reach production.
Go-live sign-off
Your workflow owner decides, based on shadow-mode evidence.
IMPLEMENTATION APPROACH
| Week | Phase → You get |
|---|---|
| Week 0 | Workflow review: one call plus a written scope → Agreed workflow, success measures, integrations and fee |
| Week 1 | Discover → Baseline metrics, system access, evaluation set started |
| Week 2 | Design → Architecture, data flow, approval points and risk controls signed off |
| Week 3–4 | Build → Integrated worker in sandbox, review queue, logging, evaluations passing |
| Week 5 | Shadow mode → Live parallel run and results against the baseline |
| Week 6 | Go live & handover → Production switch-on, monitoring, runbook, recommendation |
What we need from you: a workflow owner for about 2 hours a week, a technical contact, system access, and real historical examples.
Common starting points: AI Sales Worker · AI Operations Worker · AI Recruitment Worker, or your own workflow or existing prototype.
FAQ
- What counts as 'one workflow'?
- A process with a clear trigger, steps and outcome. Examples are 'new inbound lead to qualified and routed in HubSpot', 'PO email to validated order in the ERP' and 'new application to summarised and ranked in Bullhorn'. We agree the boundary in writing before starting.
- We already have a prototype. Do we start again?
- Not necessarily. We review what exists in week 1 and keep what holds up. The usual gaps are evaluation, integration, controls and monitoring, and those are what the sprint adds.
- How is the fee set?
- It's fixed for the agreed scope, and agreed in writing after the workflow review, before any work starts. There's no open-ended time and materials. New ideas go into a backlog for after the sprint.
- What if the shadow-mode results aren't good enough?
- Then we don't go live, and you get a written explanation of why and what it would take. A clear 'not yet' after six weeks costs far less than a stalled project that runs for six months.
- Who owns the system afterwards?
- You do: code, configuration, prompts, evaluation set and documentation. It runs in your environment. We can operate and extend it with you, but you're not dependent on us.
Bring one workflow. Leave with a production system.
Show us the workflow, or the prototype that stalled. We'll tell you plainly whether it fits a six-week sprint and what production would look like.