While building DealMind, one design decision became surprisingly important:
Which parts of the system should actually use AI?
Negotiation decisions are not only about what a customer asks for.
They are also about what a concession actually costs.
That is why I wanted the economics layer in DealMind to remain deterministic.
A language model can explain a number.
It should not be responsible for inventing the number.
Starting with the Current Deal
DealMind receives information such as:
- Deal value
- Initial offer
- Customer counteroffer
- Requested discount
- Contract length
- Competitor pressure
- Customer objection
The system then calculates the economics of the current negotiation.
For example, consider a $100,000 deal.
At a 20% discount:
$100,000 × 20% = $20,000 concession
At an 8% discount:
$100,000 × 8% = $8,000 concession
Difference:
$12,000
The correct interpretation is:
$12,000 less in discount concession compared with a 20% discount.
We deliberately avoid calling this guaranteed profit or margin savings because that would require actual cost and margin data.
Why Deterministic Calculations Matter
This is one of the areas where using an LLM for everything would be unnecessary.
The calculation is simple.
The application can perform it directly and consistently.
The LLM can then explain the result in natural language.
This gives us a simple pipeline:
Deal Data
↓
Deterministic Economics
↓
Historical Evidence
↓
Strategy Generation
↓
LLM Synthesis
The important part is that the model receives reliable numbers rather than being asked to calculate and interpret everything at the same time.
Combining Economics with Memory
The economics becomes more useful when combined with historical experience.
Suppose Hindsight retrieves previous negotiations where large discounts were unsuccessful.
At the same time, the current deal has a large discount request.
That historical context can influence which strategies are worth considering.
This is where DealMind's Strategy Lab becomes useful.
Instead of showing a single unexplained answer, the system can present possible approaches.
Hold Price and Increase Value
Respond to the price objection by emphasizing additional value instead of immediately reducing the price.
Trade Concession for Commitment
Offer a smaller concession in exchange for a longer contract or another meaningful commitment.
Respond to Competitor Pressure Carefully
Avoid automatically matching the competitor's requested price. Instead, use the available evidence to determine an appropriate response.
The actual options depend on the current negotiation and the historical evidence available.
What-If Analysis
The What-if Simulator allows the salesperson to explore different negotiation assumptions.
For example:
| Discount | Concession |
|---|---|
| 20% | $20,000 |
| 15% | $15,000 |
| 10% | $10,000 |
| 8% | $8,000 |
This makes the economic trade-off visible.
Instead of asking an AI to tell the salesperson which number is "best," DealMind lets the salesperson explore the consequences of different assumptions.
Counteroffer Guidance
The Counteroffer Advisor extends the same idea into a live negotiation.
The current customer counteroffer becomes the starting point.
The application can combine:
- Current offer
- Requested discount
- Historical evidence
- Customer context
- Competitor pressure
- Economic impact
The user can then consider possible responses based on the available information.
Human Judgment Still Matters
The system is not designed to make the negotiation decision automatically.
The salesperson still needs to consider factors that may not be represented in the application.
DealMind's role is to make the available evidence and economics easier to reason about.
That is why the product is better thought of as negotiation decision support rather than autonomous negotiation.
The Architecture
The economics and strategy workflow can be summarized as:
Current Negotiation
↓
Hindsight Evidence
↓
Economics Calculation
↓
Confidence
↓
Possible Strategies
↓
What-If Analysis
↓
Counteroffer Guidance
↓
Salesperson Decision
After the negotiation, the outcome returns to Hindsight.
That closes the loop.
The system can therefore learn from the outcome without making the original decision autonomous.
What This Taught Us
The economics layer taught us an important engineering lesson:
Not every part of an AI product needs to be powered by AI.
Calculations, thresholds, evidence counts, and state transitions are often better handled deterministically.
The language model becomes more useful when it is given reliable inputs to explain rather than being asked to calculate and invent everything itself.
That balance is what makes DealMind's negotiation guidance practical.
And, for me, that is one of the more interesting parts of building an AI product:
Deciding where AI should be used is just as important as deciding where it shouldn't be.
Top comments (0)
Some comments have been hidden by the post's author - find out more