Every company has a set of decisions that recur frequently enough that they should be consistent.
How to respond to a discount request. When to escalate a client complaint. What to do when a project runs over scope. How to handle a payment that is 45 days overdue.
In most companies, these decisions are made freshly each time, by whoever encounters the situation, with whatever information they have available. The quality of the decision depends entirely on who made it and what mood they were in.
AI does not improve this situation without decision rules. It accelerates the inconsistency.
Decision rules are the documented logic for the judgment calls that recur often enough to standardise. They are working instructions that tell the AI (and the team member using it) exactly what the company does in specific situations.
Loaded into the AI workspace, they convert every AI recommendation from “what a reasonable professional would do” to “what this company specifically does.” That distinction is the difference between an AI that assists and one that represents.
Why the AI without decision rules produces the wrong answer
The scenario: a client emails the account manager: “I was hoping we could discuss the project fee. The project has taken longer than expected and we’ve had to absorb some internal costs. Is there flexibility on the final invoice?”
What the AI produces without decision rules
The account manager asks the AI to draft a response. The AI produces a professionally worded acknowledgment that expresses empathy, offers to review the project timeline and cost breakdown, and suggests a call to discuss options.
It is a perfectly reasonable response; the kind a professional consultant would send.
Why this is wrong for the company
The company has a specific policy: for long-term clients (18+ months), the account lead can authorise a goodwill credit of up to 5% of the invoice without approval. For shorter-term clients, discounts require the founder’s sign-off.
The specific client in this scenario has been with the company for two years. The account manager has full authority to resolve this.
The AI’s response; “let’s have a call to discuss options”; was not wrong. It was the safe, generic response.
It was also unnecessary. The account manager could have resolved this in the first reply with a 5% credit offer, closing the situation in 30 seconds.
Instead: a call has been scheduled, the client feels the firm is reluctant to help, and the resolution takes three days.
With a decision rule loaded
SCENARIO: Client requests fee reduction citing unexpected project duration
STANDARD POSITION: We do not reduce fees based on project duration unless
the duration was caused by our actions
EXCEPTION (long-term clients, 18+ months): Account lead may offer a goodwill
credit up to 5% of the invoice at their discretion; no approval required
EXCEPTION (shorter-term clients, under 18 months): Requires founder approval
before any credit is offered
WHAT TO SAY: Acknowledge the challenge; reference the goodwill credit if
applicable; make the specific offer in the same email rather than
scheduling a callThe AI response: “I understand the project timeline extended beyond what either of us expected, and I appreciate you flagging it directly. As a long-term client, I’d like to offer a 5% credit on the final invoice; I hope that acknowledges the situation fairly. I’ll send a revised invoice today.”
The response is specific, confident, and appropriate. The client feels respected. The situation is closed in one email. The account manager exercised the authority they had but did not know they could use without checking.
The three categories of decision rules: what to document
Category 1: Commercial rules (build first)
Commercial decision rules cover pricing, discounts, payment terms, contract modifications, and the commercial situations that arise in sales and account management.
Why build first: commercial decisions are the ones most likely to cost money when made inconsistently.
An account manager who gives a 15% discount because they did not know the limit was 10% costs real margin.
An account manager who refuses a 5% goodwill credit to a long-term client; because they did not know they had the authority; costs a relationship.
Rules to include in Category 1:
- Discount request response: what the company does at each approval level and what the ceiling is before escalation
- Payment terms standards: the standard terms and what conditions allow deviation
- Scope change pricing: how out-of-scope work is priced and who approves the quote
- Retainer renewal pricing: the standard approach and loyalty considerations
- Non-standard contract terms: which terms the company accepts routinely and which require escalation
- Early termination: what happens when a client terminates early; the standard response and the commercial handling
Documentation format:
SCENARIO: [The specific commercial situation]
STANDARD POSITION: [What the company does by default]
EXCEPTION CONDITIONS: [What circumstances change the standard position; and how]
APPROVAL THRESHOLDS: [Who can authorise what; up to what amount]
WHAT TO SAY: [The specific language to use; not just what to do; but how to
communicate it]Category 2: Client communication rules (build second)
Client communication decision rules cover how the company handles specific relational situations; complaints, delays, scope disputes, difficult conversations.
Why build second: commercial rules have clearer financial stakes; communication rules have clearer relationship stakes.
Rules to include in Category 2:
- How to handle a client complaint: the response sequence, the tone, the authority to offer remediation
- How to communicate a project delay: what to say, when to say it, what the company offers when the delay is its own fault versus the client’s
- How to handle a client who is consistently late on payments: the escalation sequence; the tone at each stage
- How to communicate bad news (a missed deliverable, a team member change, a quality issue): the structure and the accountability statement
- How to handle a client asking about AI use: the disclosure approach and the language
- How to communicate scope boundaries: what to say when a client asks for work outside the agreed scope
Category 3: Escalation rules (build third)
Escalation decision rules define when a team member should stop handling a situation independently and bring in a senior person.
Why build third: escalation rules prevent two distinct failures.
- Too much escalation: a team member who escalates too readily creates a bottleneck at the founder
- Too little escalation: a team member who does not escalate when they should handles situations beyond their authority without sufficient support
Rules to include in Category 3:
- What client situations require the founder’s direct involvement: which complaints, which commercial requests, which relationship inflection points
- What communications require review before sending: which emails, to which clients, on which topics
- What financial decisions require founder sign-off: what thresholds trigger founder approval in both directions
- When to involve the AI system owner versus handle independently: when an AI output failure requires system-level intervention versus a one-off correction
The sourcing method: how to identify the rules worth documenting
The “ask the founder” method
The fastest way to identify the most valuable decision rules is to track the questions the team asks the founder.
Every “should I…?” and “what do we do when…?” question the founder receives is a decision that should be documented.
The two-week tracking exercise:
For two weeks, the founder keeps a running list of every decision question the team brings to them.
At the end of two weeks, they have the decision rule backlog; the complete set of situations the team does not have guidance for.
The categories that emerge most often:
- Discount and credit requests from clients
- Client complaints and escalations
- Scope interpretation disputes
- Payment chase situations
- Resource allocation conflicts between clients
- Non-standard client requests
- Team member questions about what they can and cannot commit to without approval
These categories become the Category 1, 2, and 3 rules above.
The “reverse-engineer from mistakes” method
An equally useful source is the decisions that have gone wrong; the situations where a team member handled something in a way the founder would not have, and the outcome was suboptimal.
For each past mistake or near-miss: “What specific rule, if it had existed and been loaded, would have produced the right decision?”
That rule is the highest-priority addition to the document.
The decision rules document format: how to write rules the AI can use
The core format (used for all three categories)
SCENARIO: [The specific situation; named precisely enough that the AI recognises
when it applies]
STANDARD POSITION: [What the company does by default in this situation]
EXCEPTION CONDITIONS: [The specific circumstances that change the standard position]
ESCALATION THRESHOLD: [The point at which this decision requires senior involvement]
WHAT TO SAY / HOW TO COMMUNICATE: [The specific language or framing to use]
WHAT NOT TO SAY: [The specific language or framing to avoid]The difference between human-readable and AI-usable
Human-readable rule: “Use good judgment on discounts. Be fair to long-term clients but protect margin.”
This rule is useful for a human who can fill in the specifics from experience. It is not useful for an AI; which will apply “good judgment” in a generic way that does not reflect the company’s specific thresholds.
AI-usable rule:
SCENARIO: Retainer client requests discount on renewal engagement
STANDARD POSITION: Standard pricing applies to all renewal engagements
EXCEPTION CONDITIONS:
18+ months: account lead can offer 5% reduction at their discretion;
no approval required
24+ months with annual contract value above $60K: account lead can offer 10%;
requires account director sign-off
Any client; any tenure: amounts above 10% require founder approval before offer
ESCALATION THRESHOLD: Any discount above 10%; or any situation where the client
is threatening to leave rather than simply asking
WHAT TO SAY: Acknowledge the relationship; make the specific offer at the
appropriate tier; frame as a gesture to the partnership rather than a
negotiation
WHAT NOT TO SAY: Do not use "standard policy" or "our hands are tied"; these
are relationship damage phrasesThe second version gives the AI everything it needs to produce an accurate, specific recommendation. The first version does not.
Length and completeness
A complete decision rules document for a $10M–$25M company covers 15–25 rules across the three categories.
Total length: 1,500–3,000 words. It is not a legal policy document; it is a working instructions file that is updated as new situations arise and as the company’s positions evolve.
Common questions on AI decision rules
”How is a decision rule different from a policy?”
A policy is a formal compliance document; it defines what is permitted or required, often for legal or regulatory reasons.
A decision rule is a working instruction; it tells the AI (and the team) what to do in common situations and is updated as the company’s positions evolve.
Policies are written to be permanent. Decision rules are written to be useful. The format is different; the purpose is different; both are needed.
”What if the right answer genuinely varies by situation: how do I write a rule for that?”
Write the rule for the most common case and document the exception conditions. A rule that says “the answer always varies” gives the AI nothing to work with.
The goal is to document what determines the variation; not to pretend the variation does not exist. A rule that says “gentle reminder for 0–15 days overdue; firm for 15–30; formal demand for 30+ with a client-tier exception for long-term relationships” captures the variation and makes it usable.
”How many decision rules do I actually need?”
15–25 rules covers the most common decision scenarios for most $10M–$25M companies. Below ten rules: the most important commercial and escalation scenarios are probably uncovered. Above thirty rules: the document becomes too long to maintain and too complex for consistent AI use.
Start with ten to twelve of the highest-frequency rules. Add rules as new situations arise from the two-week tracking exercise.
”Should decision rules be visible to clients?”
No. Decision rules are internal operational documents. Clients benefit from the consistency they produce; they do not need to see the underlying logic.
The one exception: the disclosure rules in Category 2 (how to handle a client asking about AI use) should reflect a consistent, honest communication approach that the company is comfortable with clients knowing.
”Who updates the decision rules when the company’s position changes?”
The AI system owner owns the decision rules document.
Any team member who encounters a situation not covered; or covered incorrectly; flags it to the AI system owner.
The AI system owner drafts the new or revised rule and reviews it with the relevant stakeholder (usually the founder or COO) before updating the document.
The update cadence: whenever a “should I…?” question reaches the founder that is not already in the rules document; that question becomes a new rule within one week.
”Can AI help me write the decision rules themselves?”
Yes; with the right structure. Tell the AI: “I’m going to describe a commercial situation we encounter regularly. Help me write a decision rule in this format: [paste the core format above].” Then describe the situation in plain language.
The AI structures the rule accurately. The founder reviews and corrects any specific thresholds or language that the AI inferred incorrectly. The AI is good at structuring rules; it cannot know the company’s specific thresholds without being told.
Want the decision rules written, formatted, and loaded: before the team starts using AI on commercial and client situations?
Decision rules are the component of AI Foundations that most directly determines whether the AI represents the company or merely assists it.
Without them, every AI output is calibrated to generic professional practice.
With them, every output reflects the specific logic the company has developed over years of serving its clients; the thresholds, the exceptions, the language, and the judgment calls that are distinctively how this company operates.
The rules take 3–4 hours to document. They prevent the inconsistency that costs margin, relationships, and the founder’s time answering questions that should not need to come to them.
Path one: start the two-week tracking exercise today. Keep a running list of every “should I…?” question the team brings to you this week. By Friday you will have six to eight rules to write. Write them in the format above and load them into your Claude Project before next Monday.
Path two: bring in a partner. If you want the decision rules built through a structured founder interview; converting the accumulated judgment in the founder’s head into documented rules the AI can use and the team can follow; that is the Phos AI Labs Phase 1 Foundations work. We have run 400+ AI engagements. Clients include Zapier, Coca-Cola, Medtronic, Dataiku, and American Express. Thirty minutes, no deck. Start here.