Developing your AI Roadmap

Learning Objective

Draft a robust 90-day AI roadmap that sequences capabilities—from isolated assistants to multi-agent workflows—while establishing strict data boundaries, operating ownership, and commercial stop conditions.


1. The 2026 AI Operating Reality: Assistants, Workflows, and Agents

The era of isolated, undirected generative AI experimentation is over. In 2026, enterprise architecture must support dynamic, goal-oriented systems. To build a viable roadmap, you must stop treating “AI” as a monolith and instead sequence your investments across three distinct operational tiers:

  • Assistants: Single-user, reactive tools bounded by immediate human prompts (e.g., code copilots, document summarisers). These offer baseline productivity gains with minimal integration overhead and low commercial risk.
  • Workflows: Deterministic business processes augmented by AI (e.g., classifying incoming vendor invoices and routing them to the correct ledger). The workflow controls the AI, passing data through rigid, predefined pipelines.
  • Agents: Autonomous or semi-autonomous systems that chain reasoning and tool use to achieve a stated goal (e.g., dynamically adjusting supply chain orders based on live API telemetry and supplier email negotiation). Agents dictate the workflow. They carry high commercial impact but require stringent governance and dynamic controls.

Your roadmap must reflect this progression. Do not attempt to deploy autonomous agents into a department that has not yet mastered deterministic AI workflows or secured its foundational data architecture.

Strategic sequencing of AI capabilities across enterprise maturity stages

2. The “Authority-to-Act” Framework

Every AI initiative on your roadmap must pass through a decision gate governing its authority to act. Assess each capability against this framework to determine required controls and operating ownership:

  • Level 0: Advisory (Human-in-the-loop). The system analyses data and suggests an action. A named human operator must explicitly approve the action before execution. Example: Drafted client responses.
  • Level 1: Bounded Execution (Human-on-the-loop). The system executes tasks autonomously within strict, predefined numerical or categorical boundaries. A human operator reviews logs asynchronously and handles exceptions. Example: Automatic approval of expense claims under £50 that match receipt OCR data.
  • Level 2: Autonomous (System-to-System). The system negotiates and acts without human intervention, constrained only by system-level fail-safes and commercial stop conditions. Example: Programmatic ad bidding or dynamic server provisioning.

Security, risk and governance constraints should be treated as operating design inputs, not delegated after launch. You cannot abdicate risk to the IT department. The commercial owner of the business process must own the AI’s outcomes and the definition of its boundaries.

3. Execution: Build, Buy, or Configure?

Your roadmap must explicitly state the acquisition strategy for each capability. The default enterprise posture in 2026 should be to configure or buy, reserving expensive build efforts exclusively for proprietary data advantages.

  • Configure: Leverage existing SaaS platforms (e.g., Microsoft 365 Copilot, Salesforce Einstein) by activating their native AI features. Requires robust data hygiene and permissions mapping, but zero engineering.
  • Buy: Procure vendor-specific point solutions designed for vertical use cases (e.g., Harvey for legal analysis). Validated by faster time-to-value, provided integration constraints are met.
  • Build: Construct bespoke agentic workflows using orchestration frameworks (e.g., LangChain) and foundational models over proprietary enterprise data. This mandates dedicated MLOps, telemetry, and continuous evaluation pipelines.
Decision matrix for AI integration scenarios and resource allocation

4. Structuring the Sequenced Portfolio

A static 12-month plan is a liability. Your AI roadmap must operate as a sequenced portfolio of 90-day sprints. Each sprint must culminate in a hard decision gate tied to commercial KPIs, not just technical deployment. If a system hallucinates beyond acceptable thresholds or fails to reduce processing costs by the target metric, the explicit stop condition is triggered: funding halts, and the workflow rolls back to human operation.


Worked Example: The Cold-Chain Logistics Hub

Role: Director of Operations
Industry: Third-Party Logistics (UK)

The Situation: Freight routing during temperature fluctuations requires immediate carrier renegotiation. Currently, dispatchers spend 40% of their shift manually calling carriers when initial plans fail.

The 90-Day Roadmap:

  • Sprint 1 (Days 1-30) – Configure: Implement an email assistant (Level 0) that monitors the dispatch inbox, extracts delay data, and drafts suggested re-routing requests. Stop Condition: If extraction accuracy is below 95%, pause rollout and refine the prompt parameters. Owner: Shift Manager.
  • Sprint 2 (Days 31-60) – Buy & Integrate: Deploy a deterministic workflow using an off-the-shelf logistics middleware. If a temperature sensor triggers a delay, the system automatically emails the top three secondary carriers for quotes (Level 1). Humans review replies.
  • Sprint 3 (Days 61-90) – Build: Launch an internal negotiation agent. The agent reads incoming carrier quotes and is authorised to accept quotes up to 12% above standard rate to save the load (Level 2). Data Boundary: Sandboxed strictly to the spot-freight API; no access to payroll or core CRM. Stop Condition: If the agent accepts a rate above the 12% threshold, immediately revoke API tokens and revert to manual approval.

Reference point: Use NIST’s AI Risk Management Framework resources and the official EU AI Act text (2024) as inputs to governance planning, not as substitutes for legal, security or sector-specific review.

Apply it to your work

This is the capstone requirement for the course. Produce a one-page, 90-day AI roadmap for one real operating situation in your business. Your roadmap must explicitly detail the following components:

  • Commercial Outcomes: The specific financial or operational metric you intend to move.
  • Workflow Boundary: The exact start and end points of the AI’s involvement in the process.
  • Data Sources: Which systems the AI will read from and write to.
  • Authority & Controls: The required level of autonomy (Level 0, 1, or 2) and the human fail-safes.
  • Operating Owner: The named commercial leader responsible for the AI’s output.
  • Decision Gates & Stop Conditions: The explicit performance failures that will trigger an immediate rollback to manual processes at the end of each 30-day sprint.

Do not proceed until you have documented these parameters clearly. Ambiguity here is what causes enterprise AI deployments to fail.


Next Step: Get Certified

You have now completed the core roadmap frameworks. To validate your capability to architect, govern, and deploy enterprise AI solutions, transition to the structured, project-based, and credentialed AI Fluency Certification.


Accredited Executive Program • The Tesseract Academy

Level 33: Executive AI Fluency Certification

Master AI strategy, LLM deployment economics, and vendor evaluation without needing a coding background.