AI Implementation Methodology

The method Akontec applies to every AI engagement. If a step is skipped, the deployment usually fails in month two — which is the same failure pattern as unprepared outsourcing.

Step 1 — Baseline (non-negotiable)

Capture, before anything is built:

Everything reported afterwards is measured against these. Without a baseline, "it's working well" is an opinion.

Step 2 — Automatable-share estimate

Rank intents by volume. For each, ask: is the answer in a system? Is the decision rule stable? Does a wrong answer cause harm? Only intents that clear all three enter phase one.

Present the honest number. If only 30% of volume is automatable, say 30% — a client who is told 70% and gets 30% will not renew.

Step 3 — Pilot design before build

Define, in writing, before a line is configured:

• Which traffic slice, at what percentage
• Success criteria with thresholds
• The rollback trigger
• Who reviews transcripts in week one, and how many
• The tuning cadence

A pilot without pre-agreed success criteria becomes an argument about whether it worked.

Step 4 — Build with guardrails as a first-class artefact

Guardrails are configured alongside flows, not bolted on. Allowed promises, restricted statements, regulated topics, escalation triggers and the confidence floor — all defined and tested. See 27_Guardrails.

Step 5 — Evaluate before exposing

An evaluation set of at least 100 real scenarios per intent family, including adversarial and edge cases: angry customers, ambiguous phrasing, out-of-scope requests, attempts to extract information the agent should not give, and multi-intent messages. The agent must pass the agreed thresholds before any real traffic reaches it.

Step 6 — Pilot with heavy human review

Week 1: 100% transcript review. Week 2: sampled review plus daily tuning. Weeks 3–4: weekly tuning, expanding traffic only if the metrics hold.

Step 7 — Measure and report honestly

Weekly during pilot, monthly thereafter, against the baseline. Report the failures as prominently as the wins — the failure categories are what drives the next improvement.

Step 8 — Expand or roll back

Expansion is a decision made on data against the pre-agreed criteria, not a default. If the criteria are not met, we roll back and say why. A client whose pilot we honestly rolled back will trust the next recommendation.

Step 9 — Maintain

Knowledge refresh cycle, drift monitoring, monthly review, and re-evaluation whenever the client changes products, pricing or policy. An unmaintained agent degrades silently, which is the worst way to fail.