Your SOPs are fiction. Your event log is not.

Every operations leader has lived this moment. A new starter follows the standard operating procedure to the letter, and the case gets stuck. An experienced colleague leans over, says "oh, for those ones you have to...", and the case moves. The knowledge that moved it was never in the SOP. It lived in the colleague.

That gap has always been a cost. With AI agents it becomes the whole game. An agent grounded in the SOP is that new starter, permanently, at scale, with no colleague to lean over.

What the documented process leaves out

Standard operating procedures are written to describe intent. They tell you how work should flow when everything arrives correctly and every system behaves. They are usually written by someone who no longer does the work, reviewed by someone who never did, and updated less often than the products, systems and policies they describe.

What they leave out is precisely the part an agent needs most:

• The variants. A seven-step process on paper is routinely fifty or sixty distinct paths in practice, most of them rework loops.
• The exceptions. The cases resolved by an email, a phone call, a message to a lead. In one lender we looked at, a quarter of cases carried a decision that never touched the system of record.
• The real rules. The spreadsheet on a shared drive that everyone actually uses, in four slightly different versions.
• The judgement. When a senior handler accepts a non-standard document, they are applying a rule. It is just not written anywhere.

Where the truth lives

Your systems have been recording the real process for years. Every status change, every document upload, every reassignment leaves a timestamp against a case ID. Reconstruct those events and you get the process as it actually runs: every variant, its frequency, its duration, its cost.

The event log is not the whole truth, because the exceptions escape it. That is why we pair the data with shadowing: sitting with the people who handle the hard cases and capturing what they do when the system runs out. Data tells you where the process leaks; people tell you what happens in the leak.

What to do instead

Ground the agent in the empirical process, then decide deliberately what to keep. Not every variant deserves to survive. Some rework loops exist because a system was slow in 2019 and nobody removed the workaround. Discovery is the moment to see them; redesign is the moment to remove them.

Then treat the undocumented rules as first-class citizens. Capture them, reconcile the contradictions with the people who own the decisions, and give the agent the reconciled version with a trace back to the source. That is what a context pack is, and it is the difference between an agent that works in the demo and one that works on a Tuesday afternoon in November.

The SOP still has a role. It becomes the output of this work rather than the input: a document that finally describes what happens, because what happens was designed.