We take one idea at a time from a business case to a system your people use every day. The route has six stages. Each one has a fixed duration range, a written deliverable and a decision at its end that you take with the evidence in front of you. After go-live we keep running the system with you, and what it teaches us becomes the next idea.
Every stage ends in a decision
A stage closes with a short document and a meeting. The document states what we found, the evidence behind it and what we recommend. You decide whether to continue, adjust or stop, and nothing in the next stage starts before you have.
Most decisions are to continue. When the evidence says stop, we say so in writing, for example: the pilot did not move the metric, so we are not scaling it. The decision gates are where a stop costs least.
Why the pilot comes after the production build
The common order is to pilot a prototype and engineer later. In a Swiss regulated setting that order does not work. Real users and real personal data cannot meet a system before its security, data protection and human oversight exist, so the pilot runs on a governed system.
That is why stage 03 is scoped to the pilot’s footprint only: one site, one team, one desk. It is a production system, but a small one. Stage 05 does the engineering that wider use needs, and by then the pilot has shown whether wider use is worth paying for.
Stage 01 · Ground the case
3-5 weeks
Stage 01 answers whether the idea is worth proving. We name the one number the project has to move, the person accountable for it and its baseline today. We check the idea against your regulation and your data at the same time, so neither surprises you in month four.
What you receive
- A concept and business case: the use case, the metric, its owner, the baseline, the target and the reasoning behind the target.
- A first architecture sketch.
- A regulatory applicability note: whether revFADP, the EU AI Act or a sector rule such as GxP touches this use case, and how much.
- The data verdict, in one of three forms: the data supports the case, it does not, or it will once a named problem is fixed.
- The design artefacts listed below, which stage 02 builds on.
Design work in this stage
Design thinking starts here, before any model is chosen. We interview and shadow the people whose work will change, and we map the process as it runs today, including the workarounds nobody wrote down.
- Human personas. The people in the process: who does the work, who approves it, who carries the risk, who is on the receiving end.
- Agent personas. A draft persona card for every agent the case might need: its role, what it reads, what it may and may not do, when it hands over and its rung on the autonomy ladder.
- Journey map. The current process step by step, with the moments where time, money or trust is lost.
- Service blueprint with human and agent lanes. The future process, with a lane for people and a lane for agents, and every handoff between them drawn as a state with a trigger and an owner.
These artefacts often change the case. A step that looked like a job for an agent can turn out to be a rule a workflow handles better and more cheaply.
Who you need
- The person who will own the metric.
- Someone who can speak for data protection or compliance.
- The owner of each system of record the case depends on, who knows what the data contains.
The decision
You decide whether the idea is worth proving. You decide with a priced case, a named owner, a regulatory read and a verdict on the data. The usual outcomes are to move to stage 02, to commission a Data Readiness Review first when the data is the open question, or to park the idea with the reasons on file.
Stage 02 · Prove the value
3-6 weeks
Stage 02 answers whether a prototype moves the metric from stage 01. We write the evaluation first: cases drawn from your process, including the ones that went wrong, and an agreed definition of a correct outcome. That set is frozen before the prototype exists, and then we build the prototype that has to pass it.
What you receive
- The evaluation set, frozen and versioned, with the grading rules your experts agreed.
- A working prototype, scoped to the metric and run on your data or on realistic data.
- An evaluation report: the measured movement against the baseline, case by case, across repeated runs.
- A recommendation: go to production engineering, adjust one named variable and re-test, or stop.
Design work in this stage
The prototype goes in front of the people from the personas. We watch them use it on cases from their own week and note where they hesitate, override or stop trusting it.
- The handoff screen is prototyped alongside the agent, because the moment a person takes over decides whether the system is usable.
- Persona cards are revised against what the prototype did, and each agent’s proposed autonomy rung is tested against the evaluation results.
- The service blueprint is updated with the handoff states the prototype revealed.
Who you need
- The metric owner, for checkpoint reviews.
- Someone technical enough to check the evaluation method as well as the results.
- Access to the data confirmed in stage 01.
The decision
You decide whether to fund production engineering. You decide with an evaluation report that shows what the prototype did on cases you recognise, how consistent it was, and what it would take to engineer.
Stage 03 · Engineer the system
8-16 weeks
Stage 03 answers whether the system can run safely for real users in one place. We engineer it for production: integrations, access controls, monitoring, failure handling and the documentation an auditor will ask for. The scope is the pilot’s footprint, so the build stays small enough to finish and to check.
What you receive
- A production system scoped to the pilot, running in the hosting option agreed for your data.
- A governance pack: model documentation, access and data-handling controls, monitoring thresholds, the human oversight design and the audit trail.
- The evaluation set from stage 02, wired into the release pipeline as a regression suite.
- An operating runbook for the pilot: who is called, what is rolled back, what is logged.
Design work in this stage
The human oversight design is interface design. We build the approval queue, the handoff package and the screens where a person reviews, corrects or stops the agent, and we test them with the people who will use them in stage 04. Persona cards are signed by your accountable owner before go-live.
Who you need
- IT or platform engineering, engaged from week one.
- Your data protection contact, for the DPIA and the access model.
- The accountable owner who signs the persona cards.
The decision
You decide whether to put the system in front of real users at one site. You decide with the governance pack, a passing regression run and a signed persona card for every agent.
Stage 04 · Pilot in one place
6-12 weeks
Stage 04 answers whether people use the system in their daily work and whether the number moves. We run it live in one bounded scope: one team, one site, one product line or one desk. The scope is small enough that a correction is cheap and large enough that the result means something.
What you receive
- The system live with its first users, supported by us throughout.
- An adoption and value report: who used it, how often, what moved against the stage 01 baseline, what did not, and what has to change before anyone else gets it.
- A record of every handoff and override during the pilot, with the reasons people gave.
Design work in this stage
Pilot research is fieldwork. We observe the team at work, interview users at the start and the end, and read the handoff and override logs alongside them. The most useful signal is where people route around the system, because it shows where the service blueprint and the working day disagree. Findings go back into the persona cards, the handoff package and the interface before stage 05.
Who you need
- The operational manager whose team’s workflow is changing.
- A direct channel to the frontline users, as well as to their reporting line.
The decision
You decide whether to extend the system beyond the pilot. You decide with measured use at one site against the baseline you set in stage 01.
Stage 05 · Scale what works
3-6 months
Stage 05 answers whether the value holds across the full scope. A system built for one footprint now meets more load, more edge cases, more integrations and people who were not part of the pilot. We do the engineering and the change work together, and we measure the metric again at the new scale.
What you receive
- The system extended across the agreed scope.
- The engineering wider use needs: capacity, integrations, monitoring and access for every new site or team.
- Training and change work, so each team knows when to rely on the system and when to override it.
- A value report at the new scale that confirms where the pilot result held and names any place it did not.
Design work in this stage
The service blueprint is adapted site by site, because processes that look the same on an org chart rarely run the same way. Onboarding is designed from what the pilot taught us: the first ten minutes with the system, the first handoff, the first override.
Who you need
- A sponsor with authority across the full scope, beyond the pilot team.
- A named contact at each new site or team.
The decision
You decide whether the system moves into steady operation. You decide with the metric measured across the whole scope.
Stage 06 · Run and evolve
Ongoing
Stage 06 answers, every quarter, whether the system still pays for itself. AI systems drift as the world they model changes, as models are updated and as regulation moves. We run the system with you under an operating agreement with defined service levels, and we review its value against the stage 01 metric every quarter.
What you receive
- The running system under an operating agreement.
- A monitoring view you can read: cost per completed task, drift against the evaluation set and how often agents hand over.
- A quarterly value review with a recommendation.
- A roadmap of improvements and new cases, each one entering the method at stage 01.
Design work in this stage
Users keep being heard after launch. Each quarterly review includes a short round of user interviews and a read of the handoff and override logs. Persona cards are updated when an agent moves up or down a rung, and the card is signed again.
Who you need
- A standing owner on your side, part time, who acts on the quarterly review.
The decision
Every quarter you decide whether to keep the system as it is, improve it, retrain or rebuild it, or retire it, and which idea enters stage 01 next.
How engagements are contracted
Three ways to start
Most clients start with one of three fixed-scope offers. Each offer’s page sets out the scope, the length and what you leave with. Price is on request.
| Offer | Where it sits | Duration | You leave with |
|---|---|---|---|
| AI Opportunity Assessment | Stage 01, sold on its own | 1-2 weeks | A business case with a named metric and owner, a regulatory applicability note, a first read on the data and a recommendation |
| Data Readiness Review | A diagnostic inside stage 01, used when the data is the main risk | 1-3 weeks | A scorecard for every data source on completeness, accuracy, accessibility and lineage, and a remediation plan |
| Proof of Value Sprint | Stage 02, sold on its own | 2-4 weeks | A working prototype, an evaluation report against the stage 01 baseline and a go, no-go or adjust recommendation |
The assessment is the shorter read of stage 01. When a use case is complex, regulated or spread across several systems of record, stage 01 runs closer to its full three to five weeks and usually includes the Data Readiness Review. The sprint is stage 02 for a case whose metric and data are already confirmed; a stage 02 that still has to collect its evaluation cases runs closer to six weeks.
Engineering, pilot and scale
Stages 03 to 05 are contracted one at a time, each at the decision that opens it. Before you sign, you see the scope, the duration, the deliverables, the exit criterion and who we need from your side. Signing stage 03 does not commit you to stage 04 or 05.
| Stage | Contracted as | Duration | Exit criterion |
|---|---|---|---|
| 03 · Engineer the system | A fixed scope for the pilot footprint | 8-16 weeks | Governance pack complete, regression suite passing, persona cards signed |
| 04 · Pilot in one place | A fixed scope for one team or site, with our support on site and remote | 6-12 weeks | Adoption and value report delivered |
| 05 · Scale what works | A fixed scope for the agreed sites and teams, planned in phases | 3-6 months | Value report at scale delivered |
| 06 · Run and evolve | An operating agreement with defined service levels | Ongoing, reviewed quarterly | Renewed, changed or ended at any quarterly review |
Price for each stage is on request and quoted once its scope is fixed.
Support after go-live · Run and evolve
Run and evolve is a named service with a written operating agreement. It covers everything a system in daily use needs after the launch.
| What we do | What it means for you |
|---|---|
| Operation | We run the system in the hosting option agreed in stage 03: Swiss-hosted cloud, your own infrastructure or on premises. |
| Monitoring | You see the dashboard we see: traces of every run, cost per completed task, drift against the evaluation set, how often agents hand over, errors and latency. Each signal has a threshold that triggers an alert to us. |
| Regression evaluations | Every change to a model, prompt, retrieval source or tool reruns the evaluation suite before release. When a model provider updates or retires a model, the replacement passes the same suite before it goes live. |
| Incident response | The operating agreement sets severity levels, response times and who is called on both sides. Every incident ends with a written report: what happened, the cause and the change that prevents it. |
| Autonomy reviews | An agent moves up a rung on the autonomy ladder only when its evaluation results support the move, and moves down after an incident. Your accountable owner signs every change. |
| Quarterly value review | The stage 01 metric, measured again, with a recommendation: keep, improve, retrain, rebuild or retire. |
| Regulatory watch | When a rule that touches the system changes, we assess the impact and tell you what has to change and by when. |
| Roadmap | Improvements and new ideas from the review enter the method at stage 01, with their own business case. |
If you want your own team to run the system, we hand it over with the runbooks, the evaluation suite, the persona cards and the dashboards, and we support the handover period.
Frequently asked questions
Do we have to start at stage 01?
No. If the business case is already written and owned, start with the Data Readiness Review. If the case and the data are both confirmed, start with a Proof of Value Sprint. We read what you already have and tell you which stage it lets you skip.
How long until real users have the system?
Add up the stages you need. At their full published ranges, stages 01 to 03 take 14 to 27 weeks, and the pilot starts after that. Starting with the two fixed-scope offers shortens the first two stages.
What does it cost?
Each stage is priced on request once its scope is fixed, before you commit to it. The scope depends on the number of systems, data sources and sites involved. The three entry offers have the smallest and most predictable scope.
Can we stop after any stage?
Yes. Every stage ends with a document or a system you can use without us: a business case, an evaluation report, a governed system, a pilot report. Contracts run one stage at a time, so stopping costs you nothing further.
Who do you need from our side?
At every stage, an owner for the metric. Beyond that it changes with the work. Stage 01 needs a data protection contact and stage 02 someone technical. Stage 03 needs IT or platform engineering from the first week. Stage 04 needs the operational manager and the frontline users, and stage 05 a sponsor across the full scope. The offer pages list hours per role.
Where does our data go?
That is settled in stage 01, use case by use case. The options are Swiss-hosted cloud, your own infrastructure or on premises. We run a data protection impact assessment by default on every engagement that touches personal data, and personal data is replaced with aliases before it reaches a model wherever the design allows it.
Who runs the system after go-live?
Either we run it with you under the Run and evolve agreement, or your team takes it over with our runbooks, evaluation suite and dashboards. You can change that choice at any quarterly review.
What if the system stops paying for itself?
The quarterly review will show it. You then decide whether to improve it, retrain it, rebuild it or retire it. Retiring a system that no longer moves its metric is a normal outcome of stage 06.