Implementation Roadmap

​

Overview

A typical implementation runs 8–10 weeks from kickoff to go-live. It has one preparation phase and six working phases. Each phase ends with an exit review: the FFE, the client PM and the super user confirm the exit criteria are met before moving on. Adjust the length to the number of use cases, the number of data sources and how available the client’s experts are.

​

Phase summary

​

Phase 0: Prepare

Goal. Nobody waits on access in week 1. Exit criteria. The environment is healthy. Admin accounts work. Access requests are filed with named owners. Use cases and experts are named. Workshops are booked.

​

Phase 1: Discover (weeks 1–2)

Goal. Agree what success looks like, in the client’s own words, and where the data to support it lives.

• Kickoff (90 minutes). Goals, scope, roles, the roadmap, how Experio works (How Experio Works), and how decisions are made and logged.
• WS1 Questions & Agents, once per use case. Each session produces golden questions with expected answers and agent candidates. The FFE and super user merge the sessions into one de-duplicated questions.md.
• WS2 Data Inventory. For every must-have question, identify the system or library where its answer lives. The output is a data inventory, a key crosswalk (how the same client, person or project is identified in each source) and the pilot sample.

Exit criteria.

• questions.md has at least 15–30 must-have questions across the use cases, each with an expected answer and a source of truth.
• Success criteria are signed off by the sponsor. For example: “80% of must-have golden questions pass, with correct citations, by the end of week 7.”
• Every must-have question maps to at least one source in the inventory. Questions with no source are dropped or marked “needs new data”.
• Pilot sample chosen: full structured exports, plus the document folders for 2–3 representative clients or projects.

​

Phase 2: Model (weeks 2–4)

Goal. Capture the business rules and enter them in Experio. Between sessions the FFE enters the configuration. The super user reviews it in Experio before the next session. Test Ingestion on the artifact types, run on real samples, should happen in this phase, not after. Exit criteria.

• Ontology v1 is saved, and every entity type and relationship traces to at least one golden question.
• Taxonomies are imported, with owners and a refresh process agreed.
• Each artifact type has passed Test Ingestion on 3–5 real samples: correct classification, and the expected entities extracted.
• Mappings are defined for each structured source, with the load order agreed.
• Matching strategies are set for the key entity types, and someone is named to staff the review queue.
• Model & Define > Compatibility shows no invalid configuration.

​

Phase 3: Pilot ingest (weeks 4–6)

Goal. Real data in the graph, and the configuration tuned on a sample.

1. Load structured master data in dependency order, and check the counts against the source.
2. Ingest the document sample (a few hundred files). Watch Process > Jobs.
3. Work the Conflict Resolution queues daily, and feed repeated patterns back into classification instructions and matching strategies.
4. Run Graph Evaluation per artifact type and fix the extraction instructions it points to.
5. Hold WS8 Enrichment Rules once there’s real data to look at. Run the rules and spot-check the results.
6. Expect a second modeling pass. Pilot data always reveals a missing attribute or a relationship nobody mentioned.

Exit criteria. Structured counts reconcile with the sources. The document sample is processed, with under 5% failed and the review queues cleared. Graph Evaluation is “good” or “acceptable” for each artifact type. Enrichment rules are run and spot-checked.

​

Phase 4: Validate (weeks 6–7)

Goal. Prove answer accuracy against the golden questions.

• Scoring round 1. Run every must-have question in chat. Record the answer, check the citations, score pass, partial or fail, and assign a root cause. See Accuracy Validation.
• Fix by root cause: classification, extraction, matching, mapping, missing data, or retrieval and assistant instructions.
• Scoring round 2. Re-run the full list, not just the failures, to catch regressions.
• WS9 Access Control (optional). Design the policy, run it in shadow mode, and preview it as real users.

Exit criteria. The success criteria from Phase 1 are met, or the gap is understood and accepted by the sponsor. The fix log is up to date. If access control is used, it is validated in shadow mode.

​

Phase 5: Consume (weeks 7–8)

Goal. Put the graph to work in the tools users will actually use.

• WS10 Assistants & Agent Flows. Configure the assistant(s), AI instructions and document templates. Build the agent candidates from WS1 as agent flows.
• User acceptance testing. 5–10 pilot users per use case work through real tasks for a week and log issues.

Exit criteria. Pilot users sign off each use case. Every agent flow produces usable output for its sample inputs. Blocking issues are fixed.

​

Phase 6: Launch & hand over (weeks 9–10)

Goal. Full corpus, all users, and client ownership.

• Ingest the full corpus in batches, library by library, and watch jobs and cost.
• Turn on access control enforcement, if used.
• Roll out in waves, with training by persona.
• Two weeks of hypercare, with daily triage.
• Handover: the decision log, the ontology rationale, questions.md with the final scores, the runbook for routine tasks, and the operating cadence.

Exit criteria. See the go-live readiness checklist in Go-Live & Handover.

​

Status reporting

The PM sends a weekly status report with the phase, completed and upcoming workshops, decisions made, open decisions, risks, and the latest golden-question pass rate once scoring has started. A template is in Templates.

​

Common risks