FFDE (Fortified Forward Deployed Engineering) first fortifies the conditions SMBs need for AI implementation. The public framework is 5·6·7: 5 AI foundations → 6 delivery stages (1–6) → 7-day fast validation, with a 2–4 week standard pilot when deeper evidence is needed. It is not feature stacking, but closing the gap between how you work today and how you should work in the AI era.
One visual map: five foundations, six stages and a 7-day path into evidence-led delivery.
5 · Cross-cutting foundations
What are the 5 AI pillars?
Tools alone do not create lasting value. Usable AI foundation = ② LLM Platform + ① Data & Knowledge +
③ Tools & Applications + ④ Governance & Security + ⑤ People & Processes; missing any pillar, pilots stall or get abandoned.
Every FFDE project documents three things: current process, target state, and the
verifiable gap this phase closes. At kickoff we freeze Spec (scope),
Eval (acceptance criteria), and Handoff (transfer conditions). Requirements
that cannot map to a business gap do not enter current scope.
#
Pillar
Definition
Typical deliverables
Acceptance focus
①
Data & Knowledge
Document/ledger standards + private RAG
Knowledge base, spreadsheet templates, golden Q&A
Accurate answers, traceable sources
②
LLM Platform
Cloud APIs, unified gateway, governable selection
Gateway config, tiered accounts, cost baseline
Single entry point, visible cost
③
Tools & Applications
Scenario Agents + workflow embedding
Scenario Agents, automation flows, prompt library
Embedded in daily work
④
Governance & Security
Access, red lines, audit
Permission matrix, security policy, compliance red lines
Rules for sensitive data
⑤
People & Processes
Training, SOPs, ROI, handoff
Runbook, training notes, UAT, Phase 2 plan
Team can run and iterate
6 · Delivery stages
6 stages from first use case to operable capability
6 stages (1–6) apply to all FFDE services — from qualification to operable capability. The 7-day path compresses depth; the 2–4 week standard pilot adds fuller evaluation and evidence in stages 5–6. See also pilot timeline FAQ.
1 · Qualification
Confirm Sponsor, business owner, acceptance signatory, and IT contact; assess Agent/RAG fit, data sensitivity, deployment mode; draft graduation criteria (timeline, value metrics, activation observation window).
2 · Use case & value hypothesis
Top 3 → Top 1 matrix with scene-fit three checks; freeze Spec, Eval, Handoff; intent alignment gate signed by acceptance owner.
3 · Business & data readiness
Site survey (process, documents, data red lines); knowledge base and golden Q&A; unified model gateway, account tiers, security baseline; scope confirmation gate with Out of Scope in writing.
4 · MVP & workflow embedding
Single-scenario main path build and integration; connect existing entry points; high-risk steps as copilot (human review); exception handling and refusal rules.
5 · Validation, governance & decision
UAT, golden evaluation, access/red-line spot checks; five-dimension ROI review and GO/NO-GO. Acceptance is signed here; activation starts trial and baseline — long-term observation can sit in graduation criteria (e.g. 2–4 weeks post-pilot).
6 · Handoff, ops & expansion
Runbook, owner training, knowledge updates and eval baseline; acceptance report and evidence archive; activation rate review and Phase 2 roadmap.
Real samples in KB, RAG, golden Q&A, data red lines
Days 3–5
4
Same Top 1 main path: build → integrate → business trial (not parallel office/finance/sales layers)
Day 6
5
Access matrix, audit, compliance red lines
Day 7
5 + 6
UAT, GO/NO-GO, training handoff, ROI review, Phase 2 recommendations
Note: by Day 7 you complete acceptance sign-off (stage 5). Activation targets in graduation criteria may include a post-pilot observation window — stable daily use by everyone on Day 7 is not required.
7 · Fast validation
What can you see in 7 days of AI deployment?
Within 7 days, run a minimum viable version of all 5 pillars around the frozen Top 1 scenario: Agent demonstrable and
UAT-ready, plus ROI narrative and Phase 2 recommendations. Daily playbook below.
Day 1Discovery, Top 3 scenarios, unified gateway & accounts② LLM Platform
Day 2Data collection, cleanup, private RAG knowledge base① Data & Knowledge
Day 3Map workflow and first runnable steps for Top 1③ Tools & Applications (main path)
Day 4Build, integrate I/O, exceptions, and human review nodes③ Tools & Applications (integration)
Day 5Business user trial, refine main path, prepare UAT③ Tools & Applications (trial)
Day 6Permissions, red lines, audit, compliance④ Governance & Security
Day 7Training, acceptance, ROI review, iteration rhythm⑤ People & Processes
Note: Days 3–5 always focus on the same Top 1 business scenario; office, finance, and sales are common candidate directions only. Out-of-scope needs go to Phase 2 — not parallel stacks during the pilot.
Before you build
How do you know a use case is worth doing?
Most AI projects fail on problem selection, not model quality. FFDE filters candidates with a three-part scene fit check — used with the Top 3 workshop matrix.
① Pain check
It must be a specific person’s specific pain — e.g. “the support lead spends 3 hours every Monday consolidating tickets,” not “improve support efficiency.”
② Economics check
Estimate weekly hours saved, cost of errors, or a KPI leadership cares about. Pain without a number rarely survives the next budget cycle.
③ Feasibility check
Where is the data, who owns it, what sensitivity applies, and what accuracy bar is needed (90% + human review vs 99% automated).
AI accelerates execution; three decision gates require human sign-off: intent alignment
(kickoff) → scope confirmation (before build) → GO/NO-GO (acceptance). Final bar:
one Agent MVP + ROI review + release decision. See How to accept a pilot.
Graduation criteria
Agree in writing at kickoff: timeline cap, value metrics, activation goals. Meet them → expand; miss → exit cleanly and archive — no endless PoC.
Launch ≠ activation
Provisioning accounts is administrative launch. Activation means target users use the system in daily work — we track activation rate (target users as denominator, not account count).
Copilot first
For high-risk flows, start with “AI drafts + human confirms,” then raise automation as trust builds — lower resistance, higher adoption.
Define scenario and acceptance first, build the Agent MVP, then iterate via training, governance, and a reusable standard pack.
Governance first
Data classification, Agent least privilege, human review on critical outputs — security built into the flow, not patched later.
Measurable ROI
Five-dimension review: efficiency, quality, cycle time, cost, and knowledge assets — then Phase 2, not endless PoC.
Prerequisites
Make the business describable before the Agent is runnable
AI does not organize tacit knowledge or resolve process disputes by itself. Before build, FFDE makes inputs, owners, rules, and exceptions locatable and confirmable — only then can AI enter daily workflows reliably.
Describable
Map data sources, key fields, and real workflows — separate official rules from habits and stale practice.
Executable
Confirm standard actions, exceptions, and decision boundaries with the business owner; capture in Spec and acceptance samples.
Runnable
Start one high-frequency, lower-risk scenario; connect data, access, and workflow; expand after UAT and handoff.
Most use cases succeed with prompt engineering first, then RAG. When you need
stable JSON, multi-label classification, evidence spans, or sensitive-field detection
and prompt tuning still wobbles, we move to a controlled path:
corpus → teacher pre-label → human Gold → holdout → LoRA → same Gold-set evaluation.
This is not knowledge-base Q&A — you are training output shape and business rules, not indexing documents.
Good fit signals
Output schema is frozen; production needs parseable structure; human Gold samples exist; errors are reviewable and reversible.
Three data rules
Teacher pre-labels speed up work — they are not ground truth. Human Gold is holdout-only — never training labels. Reference drafts are read-only — annotators must not copy teacher fields.
How we accept
After Gold freeze: base-model baseline → LoRA fine-tune → same Gold set for JSON validity, field accuracy, and related metrics — written into the acceptance report, not “it feels better.”
Steps
What happens
Gate (summary)
1–2
Corpus intake & de-identification; freeze output schema and annotation rubric
Provenance traceable; rubric signed by business owner
3–4
Teacher model batch pre-label + automated validation / audit
Structural validity meets bar; failures routed to priority queue
5–6
Human Gold review; lock holdout (by sample id and content hash)
Validation pass; holdout version frozen before training
7–8
Train LoRA on non-holdout teacher labels; baseline vs fine-tuned evaluation
No holdout leakage; comparison report in evidence pack
Who does the labeling? High-volume field work is often done by interns, part-timers, or outsourced teams — that is normal in the industry.
Quality depends on process, not on who is cheapest: calibration sessions before scale-up, validate-on-save, honeypot items in the queue, dual annotation with adjudication, expert spot checks, and batch freeze before expanding volume. Schema design, high-risk semantics, and acceptance sign-off stay with domain experts and the business owner.
Note: synthetic pilot corpora validate the pipeline only — not production claims. Formal projects use customer de-identified real data with agreed sample size and acceptance thresholds.
FFDE applies
12-Factor Agents
(Dex Horthy / HumanLayer): most agents stall at demo because prompts, context, control flow, and state are hidden in framework black boxes — not because the model isn’t smart enough.
Production-ready delivery = mostly deterministic software with LLM steps at the right points, aligned with Spec/Eval/Handoff and human gates.
Three foundations
LLM as pure function → context engineering; Agent as software → Sidecar + structured next step; own control flow → pause and human review.
Small & focused
One Agent MVP per use case; main path ~3–20 steps — context, tools, state, errors, control flow.
Own control flow
Query steps may continue; writes and external publish must pause — human approval as a first-class tool, matching the three gates.
How do you choose between 7-day fast validation and a 2–4 week standard pilot?
Short decision chain, want a single-scene win first → 7-day fast validation.
Need formal acceptance and full pillars → 2–4 week standard pilot. Same methodology;
difference is depth and evidence requirements.
