How should AI work with a planning system?
The safest pattern is: the AI proposes, a person approves, and the evidence is kept. The AI reads the plan with the same permissions as the person using it, drafts changes with a reason and a summary of its evidence, and nothing changes until a person approves it in the planning system.
Once planners see what an AI assistant can explain, the next question is obvious: can it make the changes too? It can draft them. But in a plan that drives buying, production, and financial commitments, the AI should never be the one who decides. This article, part of our series on using AI in the planning process, covers how an AI assistant should work with a planning system: how it connects, how changes flow from proposal to approval, how a sales rep uses it during the week, and how permissions keep each person's AI inside their own lane.
Your own AI, connected to your planning system
…the technology’s larger impact will be in complementing and augmenting human capabilities, not replacing them.
The design we've found works best is not a chat window built into the planning tool. It's each user's own AI assistant, connected to the planning system the way you'd connect it to email or a calendar, and signed in as that person. The assistant reads the plan through a defined set of tools and sends back only what the user approves.
That matters most for data that should never leave its home. A rep's email, a meeting transcript, or a buyer's spreadsheet is read in the user's own AI. The planning system receives only the approved result: a number, a decision, and a short summary of the evidence. The source document itself never has to be copied into the planning system.
The AI proposes; a person decides
Every change the AI suggests becomes a proposal, not an edit. Each proposal carries the change, the reason, a short summary of the evidence (never the source text), and a batch name, so decisions made together, such as everything agreed in one monthly planning meeting, can be reviewed together. Nothing changes until a person approves it in the planning system.
| Item · account · month | Current | Proposed |
|---|---|---|
| Women's boots · Bellmont · Mar | 5,100 | 5,900 |
- Reason: buyer confirmed 12 more doors from March customer input
- Evidence: summary of the buyer's email, Oct 3, shared by the rep email summaryThe email itself stays in the rep's own mailbox
- Proposed by the assistant for Maria L. Oct 14
Illustrative. Names and numbers are made up.
When a person approves, the change is applied as their edit, and it keeps the reason, the evidence, and who proposed it, so the rationale stays attached to the number. The AI has no way to approve its own proposals. And because mistakes happen, an undo should be something the AI can propose too: one action that restores every affected value to what it was, rather than a manual fix someone has to work out.
The sales rep: from a buyer's email to a forecast with its basis
Sales reps know things the planning system doesn't: what a buyer said last week, which orders are in hand, which store is closing. Today that knowledge reaches the plan as a single number with no explanation. With an AI assistant, it arrives as a number with its basis:
- The rep asks their assistant to draft the forecast for their accounts, using their own email and open orders.
- The assistant drafts a projection per account and month from last year, orders in hand, what the emails say, and the current statistical forecast. Each number names its basis.
- The rep adjusts it and chooses what to share. Emails stay private; only the excerpts or summaries the rep attaches go with the submission.
- The planner sees the submission with its basis and can accept it, adjust it, or send it back.
Use my emails and open orders to fill in my spring forecast for Bellmont and Carroway's.
| Account · month | Statistical | Proposed | Basis |
|---|---|---|---|
| Bellmont · Feb | 4,200 | 4,800 | Orders in hand 3,900 (2 POs) plus the usual reorder pattern |
| Bellmont · Mar | 5,100 | 5,900 | Buyer email, Oct 3: "12 more doors from March" |
| Carroway's · Feb | 2,600 | 2,600 | No new information; statistical forecast kept |
Illustrative. Names and numbers are made up.
The rep's week, on the phone
Reps don't sit at a planning grid. They know the most right after a meeting, often in the car. So the capture happens by voice during the week, and the review happens once, at the end of the week. The forecast changes only when the rep approves, and even then it goes to the planner first.
Dictate after a meeting. The assistant asks one or two short questions and files the note.
Notes add up to a pending forecast for the rep's accounts.
"4 notes captured, 1 question open, review due tomorrow."
One summary per account. Approve, edit, or hold.
The submission goes to the planner with its evidence.
Just left Bellmont. Buyer says they're adding twelve doors in March and want boots up about ten percent. Sandals flat.
Got it. Is the 10% on all women's boots, or only the styles in the spring order?
Just the spring order.
Noted for Bellmont: spring-order boots up about 10% from March, 12 new doors, sandals unchanged. It's in your draft for Friday.
Your pending forecast · Territory Northeast
Illustrative. Names and numbers are made up.
A few rules keep this trustworthy:
- What the rep says is a note, not an edit. The assistant pulls out the account, products, timing, size, and source, and asks at most two questions, only about what it can't work out.
- Nothing needs the screen while driving. In the car, the assistant reads back a one-line summary. Anything that needs a look waits for the weekly review.
- Vague notes don't move numbers. "Sounded positive" is kept as context with no number attached.
- The rep's own words become the evidence the planner sees later.
The AI sees only what the person may see
An AI assistant must never become a way around the planning system's permissions. The rule is simple: the assistant has the same access as the person asking, enforced by the planning system itself, not by the AI.
Illustrative.
Questions to ask before connecting AI to your planning system
Whether you build it yourself or evaluate a vendor, these questions separate a useful assistant from a risky one:
- Does it write to the plan, or propose? Look for proposals that a person approves, applied as that person's edit.
- Where does my data go? Source documents such as emails and transcripts should stay in the user's own AI; only approved results and evidence summaries should reach the plan.
- Does it respect permissions? They should be enforced by the planning system's server, with the assistant signed in as the user.
- Does it show its evidence? Every answer and proposal should separate evidence, assumptions, and what wasn't checked.
- Is the reasoning kept? Approved changes should carry their reason and evidence, so the plan's history becomes corporate memory.
- Can mistakes be undone cleanly? An undo should restore every affected value in one step.
For the use cases these principles support, see how to use AI in demand planning. The last article in this series puts it all together in a real setting: the monthly planning meeting, from S&OP to MFP.
Frequently asked questions
Should AI be allowed to change the forecast directly?
No. The AI should draft proposals with a reason and evidence, and a person should approve them in the planning system. The approved change is that person's edit, with the reasoning attached, and the AI cannot approve its own proposals.
How does an AI assistant connect to a planning system?
The most controlled approach is to connect each user's own AI assistant to the planning system, signed in as that user, through a defined set of tools. The assistant reads what the user is allowed to see and sends back only what the user approves.
Does my email or meeting data end up in the planning system?
It shouldn't. Source material is read in the user's own AI. The planning system receives only the approved number, the decision, and a short summary of the evidence.
How do sales reps use AI for forecasting?
Reps capture what they hear from buyers by voice after meetings. The assistant turns those notes into a pending forecast with the basis for each number, and the rep reviews and approves it once a week before it goes to the planner.
How are permissions handled when AI accesses planning data?
The assistant has the same access as the person using it, enforced by the planning system's server. A rep limited to their own accounts and to demand planning gets only those answers and tools.