“Ask Ana before you send the offer.”
“Check ERP too.”
“Service might know more.”
“Marketing uses a different customer ID.”
“Export it to Excel first.”
None of this shows up in the official process. But often, it is still how the process works.
After a while, nobody calls these workarounds. You just learn them. And that can work surprisingly well until you ask AI to follow the same process.
That is usually when the gaps become visible. Some context is missing. Important checks still happen outside the workflow. Experienced employees know exceptions the system has never captured.
That is where our team comes in. The AI may already be able to produce an answer. Our job is to work through the gaps around it: which ones need a business decision, and which ones we can solve with better data, integration or process automation.
We have seen the same pattern since our first Agentforce projects. Production is where AI meets old integrations, incomplete history, local rules and manual checks.
Most of those problems were already there. Trying to bring AI into the process makes them much harder to ignore.
Pressure-test one AI use case in seven places
When we start talking about AI, we re-check seven parts of the process. Some may already be defined in the existing CRM, data and integration setup. Others become specific to the AI use case.
Our consultants start by reconstructing the real decision: what information, systems, rules and exceptions does a good employee rely on before acting? We compare that with what AI can actually access, how current the information is and what the systems allow it to do.
For a use case like “recommend the next best offer”, we follow the decision from start to finish: which system gets checked, what happens when data conflicts, and where does someone “call Ana”?
1. How the customer is identified
If CRM, Service and ERP identify the same customer differently, we have to define a matching rule and a source of truth, then make that identity available consistently across the systems AI uses.
2. Which context AI actually needs
List what an employee needs before acting. For an offer, that could mean product ownership, open service issues, active quotes, payment status or consent.
We define the minimum required context for this use case.
3. Which system AI should trust
Enterprise systems disagree. Somebody usually knows which one to trust for a particular piece of information. AI needs that rule too.
We define the authoritative source and business owner for each critical input, then make that rule work across Salesforce and connected systems.
4. Is the data fresh enough?
A daily update can be fine for one task and useless for another. If an order placed ten minutes ago changes the recommendation, freshness is part of the rule.
We define the maximum acceptable age of each critical input.
5. Steps nobody wrote down
The offer is ready, but before it goes out, someone still emails Finance to confirm the price.
That check may never appear in the process map. But if people rely on it, AI has to account for it too.
We decide whether the check still makes sense. If it does, we make it part of the workflow. If it does not, we remove it instead of teaching AI to copy the workaround.
We usually learn more from the last real offer than from the process map. Someone will say, ‘Yes, but before I send it, I always check this.’ That’s usually where the missing context is.
Ivana Holjevac Brdar
Solution Consultant, Agilcon
6. Where AI must stop
AI can recommend the next offer. But if changing the price needs a manager’s sign-off, it stops there. It can prepare the recommendation, but nothing goes to the customer until someone approves it.
7. What happens when something doesn’t add up
Say AI is about to recommend an offer. CRM says the contract is active, but ERP says otherwise.
That recommendation stops there. Someone checks which status is right before anything goes to the customer.
If the same conflict keeps coming back, don’t add another exception around it. Fix what’s causing the systems to disagree.
What the seven context checks actually uncover
These seven checks do not cover everything needed for production. Security, access control, model evaluation, monitoring, auditability, performance and cost still matter.
They cover the business-context layer: whether AI has enough of the real process to produce a reliable answer, recommendation or action.
What comes out of the review is a practical production backlog:
use case → required context → source → rule → gap → owner → go-live decision
When “ask Ana” becomes part of the AI use case
The useful moment is when someone says: “Yes, but before I do that, I always check…”
That tells you the process depends on something the official workflow may not capture.
Find out why. Maybe context is missing, a business rule was never made explicit or the systems do not agree.
Whatever the reason, sort it out before AI starts inheriting the workaround.
Moving an AI use case towards production? Our team can help uncover the context gaps, decide what needs to change and turn those decisions into working processes across Salesforce and your connected systems.
