Your agent sent the contract question to the person who replies fastest, not the person who signs. That isn't a model failure, it's a context failure. The fix is three pieces of stored context: a client stakeholder map your agent can read, an approval chain with a hard stop in it, and communication preferences written as attributes instead of vibes. Get those into your AI agent operating system and the wrong-person routing mostly stops within a week.
Why does my AI agent keep emailing the wrong stakeholder?
Agents default to what's closest. The last message in the thread, the name in the from field, whoever answered in four minutes. It treats responsiveness as authority because nothing in its context says otherwise.
I watched this happen with a marketing coordinator who replied to everything inside ten minutes and signed off with "sounds good, let's move." Our agent read that as approval and sent her the scope change. The person who actually approved scope changes was a CFO who read maybe one email a week and wanted the dollar figure in the subject line. Two completely different kinds of people, and the agent had no way to tell them apart because I never told it.
Here's the part that took me a minute to accept: the model wasn't being dumb. It was being honest about how thin my context was. Most of the bad routing I've seen inside real companies was a human doing the exact same thing.
So the question isn't how to make the model smarter. It's what the model needs to know about this client that it currently doesn't.
What goes on a client stakeholder map your agent can actually use?
A stakeholder map is not an org chart. An org chart tells you who reports to whom. A stakeholder map tells your agent who can say yes, who can say no, and who just needs to be told afterward.
I keep one per client, one row per person, in a Notion database with a relation back to the client record. Airtable works the same way. If you're running light with three or four clients, a folder of markdown files is fine. The agent doesn't care about the tool, it cares that the facts are retrievable.
The fields that matter: name and title, plus how they actually sign emails. Authority level, meaning decides, influences, blocks, or informed. Pick one, because that field drives routing. What they're measured on. What they've said no to before, with a date. Preferred channel and format. Last verified date.
That "measured on" field does more work than every other field combined. If I write that the VP of Ops is measured on launch dates and the CFO is measured on cash timing, the agent can infer that moving a launch two weeks should route to the VP first, cc the CFO, and lead with the cash impact. That's an agentic workflow doing something useful instead of just drafting prettier emails.
Without that field, the agent sees two names and picks the one that appeared most recently.
How do I encode the approval chain so the agent stops inferring?
Write it as a numbered chain with thresholds, not as prose. Prose invites interpretation.
Here's the shape I use:
- Under $5,000: account lead approves.
- $5,000 to $25,000: account lead, then client's director of ops.
- Above $25,000: director of ops, then CFO.
- Any scope change: always the CFO, regardless of dollar amount.
- Anything touching a launch date: director of ops, then me.
Then the rule that saves you: if the next approver in the chain is not in the map, the agent stops and asks me. No inferring. No "probably the same person." I would much rather get a message at 9pm asking who signs than find out three weeks later that a proposal went to someone who can't approve a lunch order.
Two more pieces. Keep a separate veto list, because the person who can kill a deal is often not in the approval chain at all. And note which approvals run in parallel versus sequence, so the agent doesn't serialize something that can happen at once.
For enforcement, split the work. Stable chain rules go in the system prompt or project instructions. The per-client map lives in a retrievable knowledge file or database. If a step absolutely cannot be skipped, gate it in n8n or Make with a real approval node before the agent sends anything. Models are good at judgment. They are not a compliance department.
Then test it. Pull three real decisions from the last quarter and ask your agent who should approve each one. If it gets two out of three wrong, the chain isn't written tightly enough yet.
How do I capture communication preferences without writing a personality essay?
Preferences have to be attributes, not adjectives.
"Prefers concise communication" is useless to an agent. "Three bullets maximum, no attachments over 2MB, number in the subject line, nothing before 10am Monday" is usable. One is a mood. The other is a spec.
So for each person I store: channel, format, hard length cap, best send window, cc rules, and one line on what genuinely annoys them. That last field sounds soft, but it's the one that keeps a long-term relationship intact. If the agent knows this person hates being looped in on scheduling threads, it stops doing it.
Don't build these from memory, because your memory is worse than you think. Mine is. Open the last twenty emails from that person, look at what they answered and what they ignored, and write down the pattern. Repeat quarterly and you'll catch drift before it costs you.
Teaching AI your business is mostly this: writing down what your best account manager already knows and never wrote down. This is the unglamorous half of context engineering for founders, and it's the part of AI for founders and operators that nobody posts about. Not prompt tricks. Accurate records of real humans, kept somewhere the agent can reach.
How do I keep the stakeholder map from going stale?
Every client map gets a last verified field and a 30 day review. If a deal goes sideways, that client's map gets reviewed the same week.
Two rules keep it clean. One source of truth, because if the approval chain lives in three places the agent will pick the wrong one and you won't know why. And when two facts conflict, the agent surfaces the conflict instead of choosing. "The map says Sarah approves above $25k, but the thread says she moved teams in March. Which is right?" That message is worth more than any amount of clever prompting.
Version the files if you can. A markdown folder in git gives you a history of who you thought approved what, which is genuinely useful the one time a client asks.
FAQ
Can I just put all of this in the system prompt? For one or two clients, yes. Past that it gets long, expensive, and hard to update. Put the stable routing rules in the prompt and the per-client facts in a retrievable database or knowledge file.
What if the client reorganizes mid-project? The agent should never invent the new structure. It hits the approver-not-in-map stop rule, asks you, and you update the row with a new verified date. One minute of work, one avoided embarrassment.
Does this work if I only have five clients? It works best at five. That's the point where you still know everyone and you're starting to make mistakes anyway. Build the map now and it grows with you instead of getting rebuilt later.
I wrote the full playbook for this. How to Build Your Own AI Agent Operating System walks you through the exact architecture I use, step by step. You can get it at a.mastermindshq.business/ai-os-book.
