Technical Lead & BPMN Educator·9 min read
Start with the process, not the technology
The organisations that get value from AI do not start by asking "where can we use AI?" They start by understanding how their work actually happens and what it costs. AI is then a decision made inside a known process - not a tool bolted onto a process nobody has mapped.
The framework below runs on one idea: a trusted process model with the numbers attached. Everything - what to automate, how to design it, whether it worked - builds on that foundation. Five gated stages take a small number of funded initiatives from baseline to realised impact, and each gate must be met before the next stage begins.
The five stages
Understand
You work from one accurate model of how the process actually runs today - and you capture its numbers (cost per run, volume, cycle time, error rate) as a baseline. No baseline, no business case.
- -Capture the current process from interviews, documents, or a workshop - not the idealised version.
- -Have the people who do the work validate it, so the model reflects reality.
- -Record the baseline metrics after observing real runs, and freeze them as a version.
Gate: One trusted process model with a frozen baseline
Identify
You assess opportunities broadly but fund only the best few. Each funded initiative has an expected value, a named owner, and a target outcome. Everything else stays in the backlog.
- -Surface friction continuously - the people doing the work know where it hurts.
- -Assess each candidate on value and AI-suitability (volume, rule-clarity, data, judgement).
- -Shortlist and fund 3-4 initiatives. Detailed work starts only once something is funded.
Gate: Three to four funded opportunities with quantified value
Design for AI
You design how people and AI will collaborate on every task, and refine the business case before anything is built. The decision for each task lives in the process model where everyone can see it.
- -For each task, choose: rule-based, AI agent, AI assist, or keep human.
- -Define validation points and the metrics that will prove impact - decided now, never after go-live.
- -Compare as-is against to-be and approve the refined business case. Approval authorises the build.
Gate: Approved future-state process with a refined business case
Deploy
AI runs in production, connected to the approved model rather than undocumented tribal knowledge, and rolled out so its impact stays measurable from day one.
- -Build agents and automations against the approved process model.
- -Test behaviour against the designed process before production rollout.
- -Go live per the rollout plan, so results can be attributed with confidence.
Gate: AI live with measurable impact
Measure & Scale
You compare actual outcomes against the Stage 1 baseline, then scale, improve, or stop each initiative. The learnings make the next round of prioritisation more accurate.
- -Measure realised impact against the frozen baseline. Agent metrics explain the result; they are not the result.
- -Run a gate review: scale, refine, or stop each initiative.
- -Roll successful patterns out to similar processes; new friction feeds back into Stage 2.
Gate: Realised impact captured, next cycle ready
"The single biggest mistake I see: skipping Stage 1. Teams jump to "let's add an AI agent here" without a baseline, then six months later nobody can say whether it helped. Freeze the numbers first - it is the cheapest insurance you will ever buy."
Deciding what AI does on each task
Stage 3 is where most of the value - and most of the risk - lives. For every task in the process, pick one of four options. The decision belongs in the process model, so the whole team can see why each task was designed the way it was.
| Task profile | Choose | Why |
|---|---|---|
| High volume, clear rules, low judgement | Rule-based automation | Deterministic, cheap, auditable. No AI needed. |
| High volume, fuzzy inputs, low error cost | AI agent | AI handles variability; a wrong call is cheap to correct. |
| Judgement-heavy, high error cost | AI assist + human | AI drafts, a person decides. Keep the human accountable. |
| Rare, high-stakes, context-dependent | Keep human | Automation maintenance costs more than it saves. |
For a deeper scoring method, see our guide on how to decide what to automate with AI.
Why the process model is the backbone
Every stage leans on one shared model of the process. It is where you baseline the numbers, record the human-vs-AI decision for each task, give AI agents a structured context to work from, and anchor the before/after measurement. Capturing that model used to take weeks of workshops; AI-native tools such as Crismo can now draft it from interviews or a transcript in a day, which is what makes Stage 1 practical rather than a project in itself.
Related guides
What to Automate with AI
Score each task and choose rule, assist, agent, or human.
What is Process Automation?
Types, benefits, and how BPMN connects to automation.
AI in Continuous Improvement
Where AI helps and where it does not in Lean and Six Sigma.
Process Mapping for Automation
Map first, then automate. The first step in any AI project.
Frequently asked questions
What is the first step to implementing AI in a business?▼
Understand the process you want to improve and baseline its numbers - cost per run, volume, cycle time, and error rate. Without a baseline you cannot build a credible business case or prove impact later. The rule is simple: no baseline, no business case.
Why do most AI implementation efforts fail?▼
They start with the technology instead of the process. Teams deploy AI onto a process nobody has mapped or measured, so they cannot tell what to automate, cannot quantify value up front, and cannot prove impact afterwards. Automating a process you do not understand just produces the wrong result faster.
How do you decide which processes to automate with AI?▼
Assess broadly, fund selectively. Score each candidate on volume, rule-clarity, data availability, error cost, and how much human judgement it needs. Fund only the three or four with the strongest quantified value, a named owner, and a clear target - and keep the rest in the backlog.
How do you measure the ROI of an AI implementation?▼
Compare the process metrics after go-live against the frozen baseline you captured before you started. Outcome metrics (cost, cycle time, quality) are the ROI; agent-level metrics are only diagnostics that explain the outcome.
Do you need a process model to implement AI?▼
For anything beyond a one-off experiment, yes. A shared process model tells you where AI fits, records the human-vs-AI decision for each task, gives agents a structured context to work from, and anchors the before/after measurement. It is the backbone the whole implementation hangs on.