SME / PME track — 90 days

Target

An organization with several departments, business systems, suppliers, and levels of accountability.

Intended outcome

A prioritized portfolio, a shared foundation, and a first governed pilot — not a collection of isolated tools.

The synthetic A2 business agent for B2B quotes applies this track to an industrial SME: low, central, and high ranges; a global denominator; total cost; a frozen test set; a live pilot; and an explicit decision not to confuse A2 with A3.

Days 1–30 — Govern and discover

Lightweight governance

• executive sponsor;
• AI portfolio owner;
• business, IT/security, data, and data-protection owners;
• legal, HR, procurement, or compliance roles as cases require;
• monthly gate with written decisions.

Work

• inventory official use and shadow AI;
• establish data classification and usage rules;
• map high-volume processes;
• build baselines and a use-case portfolio;
• immediately filter high-impact cases;
• define a target architecture and common supplier criteria.

Days 31–60 — Build the evidence

• select one to three pilots that differ mechanically;
• assign an owner and budget to each;
• create test sets, critical segments, and thresholds;
• conduct legal, data-protection, and threat analyses;
• establish identities, secrets, logs, limits, and environments;
• run shadow mode and adversarial tests.

Avoid a general platform before common needs are proven. Share only what materially reduces risk or cost.

Days 61–90 — Pilot and industrialize

• launch bounded copilots for cases that passed G2;
• measure business outcomes, human corrections, and incidents;
• test rollback and continuity;
• decide separately on value, reliability, and risk;
• version models, instructions, tools, corpora, and evaluations;
• establish supplier reviews and a reassessment calendar.

Portfolio architecture

Leadership indicators

• realized value versus announced value;
• share of systems recorded in the register;
• share of cases with a baseline and evaluations;
• severe errors and time to detection;
• supplier dependency and concentration;
• human-correction and workaround rates;
• total cost per useful outcome.