LinkedIn
DecisionGuard stops risky deals slipping through. It was built by a two-person team for the TDX 2026 hackathon in San Francisco, and I did the build. We didn't win, but we walked out with a working build that solves a real problem: discount policies live in documents, while decisions happen in tools.
What it does:
- The moment a seller submits a risky quote, Agentforce flags the rule about to break and quantifies the hit to margin
- The agent writes the executive summary, proposes a counter-offer, or raises a clean exception request in Slack
- Approvers get one Slack message with full context and approve, reject or ask a question in the thread
- Status flows back to Salesforce automatically
- Every intervention is logged as a Decision Event, feeding a Tableau dashboard of margin preserved, approval times and the rules that keep firing
Rules and counter-offers are held as configuration, so the deal desk can change them without a release.
How I built it
The problem is not that sellers want to break policy. It is that the policy usually sits somewhere else, in a document, a process note or someone's head, rather than in the moment the decision is made. DecisionGuard brings it onto the Opportunity.
Behind the scenes, I used the right tool for each job. Flow does the deterministic work: cost of goods, gross margin, discount percentage, a risk score and band, and the guard status, so the seller and the approver work from the same controlled numbers. The rules and the commercial counter-plays are held in custom metadata, so the deal desk can change thresholds, weights and plays without a release.
Agentforce comes in where it adds real value. The agent receives the structured figures from Salesforce and grounded content from Data 360, explains the risk in plain language, and proposes ranked options, each with its rationale and projected impact. Put simply: the agent does not invent the numbers, it helps the seller act on them.
From recommendation to action
A Pitch button gives the seller guidance on positioning the recommended option with the customer, including likely objections. Apply updates the Opportunity with the chosen option. And because approvals do not always happen on the record, the same context goes into Slack, where a Canvas carries the deal summary and recommended option and the approver decides in the thread.
What I would tell another team
- Calculate the numbers deterministically and let the agent explain them
- Hold rules and plays in configuration the business can change
- Take the decision to where approvers already work
I design, build and review agents like this as an Agentforce consultant in the UK.