AI for Operations, Supply Chain & Project Management
Operations is process, and process is what agents are built for: take a request, check the rules, act across systems, report. AI moves from summarising work to doing steps of it, so the design question becomes which steps, and who handles exceptions.
How AI is changing this work
- Status and coordination work, such as updates, chasing and routing, is done by AI inside the work tool.
- Process automation grows from fixed scripts to agents that handle varied inputs and decide the next step.
- Forecasting gains ready-made models that predict demand without months of custom modelling.
Your layers: the 4 that matter most
1. Layer 8 · Agents
Why it matters
An operations process is a loop of check, decide, act and hand over. Agents now run those loops across systems, alongside existing automation and people.
What to do here
- Agent plus fixed automation. Use fixed rules where steps never vary and an agent where inputs are messy; let one process hold both. The agent loop
- Humans on exceptions. Route anything unusual, costly or irreversible to a named person, with the context attached. Agent permissions & sandboxing
- Many agents, one conductor. Coordinate specialist agents through one layer that tracks state and can pause or retry. Multi-agent systems
- Screens as a last resort. Where an old system has no interface for software, an agent can use its screens like a person. Computer & browser use
Examples as of 2026-11-11
2. Layer 9 · Applications & Interfaces
Why it matters
Project and work tools are where plans, tasks and updates live. AI inside them removes the reporting work that surrounds real work.
What to do here
- Automatic status. Generate the weekly update from task activity, and flag projects drifting off plan. AI features inside products
- Intake and routing. Sort incoming requests, fill missing details and assign them by rule. Copilots
- Ask across tools. Search tasks, documents and chat in one question, limited to what you may see. Chat assistants
Examples as of 2026-11-11
3. Layer 7 · Orchestration & Tooling
Why it matters
Agents are only useful if they can reach your systems. This layer connects them to orders, stock, tickets and calendars, with permissions and tests.
What to do here
- Systems as tools. Expose each system's actions, such as create order or update ticket, as named tools with limits. Tool use & protocols (MCP)
- Procedures as context. Give the agent your written operating procedure; unclear procedures produce unclear agents. Context engineering
- Dry runs. Replay last month's real cases through a new agent and compare with what staff did. Guardrails & evals
Examples as of 2026-11-11
4. Layer 5 · Foundation Models
Why it matters
Not every model is a chat model. Models trained on time series forecast demand and load out of the box, and belong in the planner's toolkit.
What to do here
- Forecast without custom training. Try a ready-made forecasting model on your history before commissioning a bespoke one. Specialised and fine-tuned models
- Use the right model for the task. Language models explain and plan; numeric forecasts come from numeric models and solvers. Specialised and fine-tuned models
- Reasoning for planning. Use a model that thinks in steps to draft plans and spot conflicts, then verify with real constraints. Reasoning models
Examples as of 2026-11-11
High-leverage use cases
- Request intake and routing. Classify, complete and assign incoming requests across teams.
- Status reporting. Project and portfolio updates written from task data, with risks flagged early.
- Exception handling. Agents resolve routine order, delivery and invoice exceptions and escalate the rest.
- Demand and capacity forecasts. Baseline forecasts from a ready-made model, reviewed by a planner.
- Procedure upkeep. Draft and update operating procedures from how work is actually done.
Start here: the next 30 to 60 days
- Week 1 to 2. Pick one high-volume process and write its steps, rules and exceptions on one page.
- Week 3 to 4. Turn on automatic status updates for one project and compare with the manual report.
- Day 30 to 60. Pilot an agent on that process in dry-run mode: it proposes, a person acts. Measure agreement.
- Day 30 to 60. Run a ready-made forecasting model on two years of history and compare with your current forecast.
Risks and pitfalls
- Automating a broken process. An agent runs a bad process faster. Fix and write down the process first.
- Errors that compound. A small mistake early in a long chain grows with each step. Add checkpoints and keep chains short.
- Irreversible actions. Orders, payments and deletions need a confirmation step or a limit. (OWASP Top 10 for Agentic Applications)
- Nobody owns the agent. Each agent needs a named owner, a log and a way to switch it off.
- Language models doing arithmetic. Quantities, dates and totals should come from systems and solvers, not from a model's recall.
Sources
A product named on this page is an example, not a recommendation.
- ServiceNow AI Agents (servicenow.com) (2026-11-11)
- UiPath Maestro (uipath.com) (2026-11-11)
- Asana AI Teammates (asana.com) (2026-11-11)
- Atlassian Rovo (atlassian.com) (2026-11-11)
- Zapier MCP (zapier.com) (2026-11-11)
- n8n (n8n.io) (2026-11-11)
- Microsoft Copilot Studio (microsoft.com) (2026-11-11)
- Chronos (Amazon) (github.com) (2026-11-11)
- TimesFM (Google Research) (github.com) (2026-11-11)
- OWASP Top 10 for Agentic Applications (2025-12-09)
- NIST AI Risk Management Framework (AI RMF 1.0) (2023-01-26)
Other roles
HR & People · Finance · Sales · Marketing · Product & Design · Engineering · Cybersecurity · Legal & Compliance · Customer Support