Framework
AI Solutioning Framework
Choose and design the route for an AI idea that has passed assessment, then prove it phase by phase before you scale.
Stage 3 of four: Discovery to Assessment to Solutioning to Adoption and Improvement. A Go from Assessment arrives with its value case; a decision to scale moves next to Adoption and Improvement.
Idea lifecycle (Solutioning is Stage 3 of 4)
This is my method for an idea that has already passed the assessment: carry the value case forward, choose and design the route, then prove it phase by phase before scaling. I work with Quality and AI leaders. GxP guidance says how much assurance a solution needs; this adds a practical way to compare delivery routes, lifecycle cost and phased investment.
1 When to use it
- After a Go from the assessment, before committing to a vendor, a build or a partner.
- At every phase gate of a pilot or slice already underway.
2 The seven steps at a glance
Solutioning an AI idea, in seven steps
Seven steps in order, top to bottom. Step 1: Bring the value case forward. Step 2: Check what you already own. Step 3: Choose the route, with four options: Buy a vendor feature, Configure an owned suite, Build on cloud-hosted models, and Co-build with a partner. Step 4: Design it with the business. Step 5: Plan the controls few publish. Step 6: Prove it in phases, as a loop. Slice or pilot, then Phase gate checks, then three outcomes. Expand: next slice or wider pilot. Expand loops back to the phase. Adjust: change and rerun the phase. Adjust loops back to the phase. Stop: close it and keep the evidence. When the phases are done, Expand continues to the final gate. Step 7: Final gate, two outcomes. Scale: validate and roll out. Not this one: the quality need still gets a funded solution. Rule: ROI decides whether to do the AI project, never whether required quality work gets done.
3 The steps
- Bring the value case forward
Baseline, targets (including adoption), owner and check date, plus a lifecycle cost ceiling that includes validation.
- Check what you already own
An existing PV safety database, MLR system, CRM or EDC may offer the AI as a feature. Check the intended use and whether its validation evidence still applies. Extending a validated platform is often the cheapest route, but price integration, supplier releases and retesting. "Already in the suite" is not the same as "already validated": reuse, not exemption.
- Choose the route
Compare first on how validation work is divided (the cost that lasts). Buying shifts work, never accountability.
Route What you validate What to watch Buy a vendor feature or product Intended use, configuration, integration, performance on your data, and supplier oversight Release cadence against your change control Configure an owned suite The change and its intended use Scope creep beyond the original validation Build on cloud-hosted models Application, prompts, retrieval, integration, and controls as custom components Model versions, deprecations, and maintenance load Co-build with a partner The whole system you will own, plus the partner's practices Knowledge transfer and ownership of the validated state at handover - Design it with the business
Deterministic backbone, GenAI only on the narrow language step, a person at the decision with authority and time. The business owner co-designs the workflow. SOP, training and change-control updates are in the plan from the start (adoption is designed in, not hoped for).
- Plan the less visible controls
- Governance layer between staff and models (identity, approved models, logging, cost attribution, policy).
- Suppliers you may not be able to audit: map available assurance across hosting provider, model provider and contract; name gaps and residual risk; restrict or reject if inadequate.
- Model version change control: pin versions, keep a regression set, and treat model, prompt, retrieval, data, configuration and intended-use changes as change records with impact assessment.
- Performance monitoring: sampled accuracy against acceptance criteria, reviewer edit rates, and usage, where falling usage is often the earliest signal.
- Prove it in phases
Each phase is co-designed with the business owner. The phase is a loop.
Slice or pilot
Build: one narrow slice, for example one section type, tested on expert-adjudicated historical cases. Keep development cases as the regression set and hold back independent cases. Buy: a contracted paid pilot with pre-agreed success criteria and an exit clause, on your own documents.
GxP
Segregated from production, on authorized, protected, representative data. No GxP record or decision relies on pilot output until validated.
Gate
Quality first (criteria met on held-out cases; if not, adjust or stop), then value metrics against the baseline, real usage against the adoption target, reviewer burden (review time, edits) and cost to date against the ceiling.
Call
Expand (next slice or wider pilot), adjust and rerun the phase, or stop.
- Final gate
Recheck total lifecycle cost (integration, licenses, human review, maintenance, change assessments on model, prompt or data changes, retirement) and the exit plan (regulated records and audit trails portable). If the numbers work: Scale (complete validation, roll out with SOPs and training, keep monitoring, confirm adoption after rollout). Otherwise: "Not this one."
4 The rule
Key rule
ROI decides whether to do the AI project; it never decides whether required quality work gets done.
5 After solutioning
The final gate, step 7, ends this stage with one of two calls.
- Scale. The total lifecycle cost and the exit plan still hold up. Complete the validation, roll out with standard operating procedures (SOPs) and training, and keep monitoring. After rollout, confirm that people are actually using it. The idea then moves to the next stage, Adoption.
- Not this one. The numbers do not work.
Monitoring does not end with this stage. Sampled accuracy against the acceptance criteria, how much reviewers edit, and usage carry on into the Adoption and Improvement stage. Falling usage is often the earliest warning.
"Not this one" closes the AI idea, not the quality need, which still gets a funded solution.
Want to work through solutioning on your project? That is the work I do.
6 Where the detail lives
The companion article, Solutioning AI in GxP: how to choose the route and prove it in phases, carries the public examples and their citations that this page leaves out.
7 A note on agentic designs
Full agentic, where a system plans and acts on its own across steps, is justified only in rare cases. Many processes that look agentic can be broken into deterministic steps, which are easier to specify, test and validate, and the route table above applies to each step. A separate article on this is coming soon.
8 Caveats
This is my own method and my recommendation. It is advisory, not a regulatory requirement and not legal advice. The examples on this page are illustrative.
Followed through, the reasoning arrives at one line:
ROI decides whether to do the AI project; it never decides whether required quality work gets done.